transcript-site/content/post/Latent Space/Latent-Space-Latent-Space-Chats:-NLW-(Four-Wars,-GPT5),-Josh-Albrecht-Ali-Rohde-(TNAI),-Dylan-Patel-Semianalysis-(Groq),-Milind-Naphade-(Nvidia-GTC),-Personal-AI-(ft.-Harrison-Chase-—-LangFriend-LangMem).lrc

[by:whisper.cpp]
[00:00.00](音乐)
[00:06.00]欢迎到达特兰的节目的节目的节目
[00:09.00]这是Charlie,你的AI co-host
[00:12.00]SWIX 和 ALESIO 都在节目中做了更多的节目
[00:16.00]我们有很多新的节目
[00:18.00]来自Elicit、Chroma、 Instructor
[00:20.00]和我们的新的节目在 NSFW
[00:23.00]"不安全的工作AI"
[00:25.00]今天我们会在SWIX 和 ALESIO 的新的节目中
[00:29.00]找到更多新的新的节目
[00:31.00]在我们的第一节目中
[00:33.00]找到更多新的节目
[00:35.00]SWIX 和 ALESIO 都在节目中
[00:37.00]做了更多新的节目
[00:40.00]SWIX 和 ALESIO 都在节目中
[00:42.00]做了更多新的节目
[00:44.00]我们有很多新的节目
[00:46.00]SWIX 和 ALESIO 都在节目中
[00:49.00]找到更多新的节目
[00:52.00]我们有很多新的节目
[00:54.00]SWIX 和 ALESIO 都在节目中
[00:56.00]找到更多新的节目
[00:58.00]SWIX 和 ALESIO 都在节目中
[01:00.00]找到更多新的节目
[01:02.00]SWIX 和 ALESIO 都在节目中
[01:04.00]找到更多新的节目
[01:06.00]SWIX 和 ALESIO 都在节目中
[01:08.00]找到更多新的节目
[01:09.00]SWIX 和 ALESIO 都在节目中
[01:11.00]找到更多新的节目
[01:12.00]SWIX 和 ALESIO 都在节目中
[01:14.00]找到更多新的节目
[01:16.00]SWIX 和 ALESIO 都在节目中
[01:18.00]找到更多新的节目
[01:19.00]SWIX 和 ALESIO 都在节目中
[01:21.00]找到更多新的节目
[01:22.00]SWIX 和 ALESIO 都在节目中
[01:24.00]找到更多新的节目
[01:26.00]SWIX 和 ALESIO 都在节目中
[01:28.00]找到更多新的节目
[01:30.00]SWIX 和 ALESIO 都在节目中
[01:32.00]找到更多新的节目
[01:34.00]SWIX 和 ALESIO 都在节目中
[01:36.00]找到更多新的节目
[01:38.00]SWIX 和 ALESIO 都在节目中
[01:40.00]找到更多新的节目
[01:42.00]SWIX 和 ALESIO 都在节目中
[01:44.00]找到更多新的节目
[01:46.00]SWIX 和 ALESIO 都在节目中
[01:48.00]找到更多新的节目
[01:50.00]SWIX 和 ALESIO 都在节目中
[01:52.00]找到更多新的节目
[01:54.00]SWIX 和 ALESIO 都在节目中
[01:56.00]找到更多新的节目
[01:58.00]SWIX 和 ALESIO 都在节目中
[02:00.00]找到更多新的节目
[02:02.00]SWIX 和 ALESIO 都在节目中
[02:04.00]找到更多新的节目
[02:06.00]SWIX 和 ALESIO 都在节目中
[02:08.00]找到更多新的节目
[02:10.00]SWIX 和 ALESIO 都在节目中
[02:12.00]找到更多新的节目
[02:14.00]SWIX 和 ALESIO 都在节目中
[02:16.00]找到更多新的节目
[02:18.00]SWIX 和 ALESIO 都在节目中
[02:20.00]找到更多新的节目
[02:22.00]SWIX 和 ALESIO 都在节目中
[02:24.00]找到更多新的节目
[02:26.00]SWIX 和 ALESIO 都在节目中
[02:28.00]找到更多新的节目
[02:30.00]SWIX 和 ALESIO 都在节目中
[02:32.00]找到更多新的节目
[02:34.00]SWIX 和 ALESIO 都在节目中
[02:36.00]找到更多新的节目
[02:38.00]SWIX 和 ALESIO 都在节目中
[02:40.00]找到更多新的节目
[02:42.00]SWIX 和 ALESIO 都在节目中
[02:44.00]找到更多新的节目
[02:46.00]SWIX 和 ALESIO 都在节目中
[02:48.00]找到更多新的节目
[02:50.00]SWIX 和 ALESIO 都在节目中
[02:52.00]找到更多新的节目
[02:54.00]SWIX 和 ALESIO 都在节目中
[02:56.00]找到更多新的节目
[02:58.00]SWIX 和 ALESIO 都在节目中
[03:00.00]找到更多新的节目
[03:02.00]SWIX 和 ALESIO 都在节目中
[03:04.00]找到更多新的节目
[03:06.00]SWIX 和 ALESIO 都在节目中
[03:08.00]找到更多新的节目
[03:10.00]SWIX 和 ALESIO 都在节目中
[03:12.00]找到更多新的节目
[03:14.00]SWIX 和 ALESIO 都在节目中
[03:16.00]找到更多新的节目
[03:18.00]SWIX 和 ALESIO 都在节目中
[03:20.00]找到更多新的节目
[03:22.00]SWIX 和 ALESIO 都在节目中
[03:24.00]找到更多新的节目
[03:26.00]SWIX 和 ALESIO 都在节目中
[03:28.00]找到更多新的节目
[03:30.00]SWIX 和 ALESIO 都在节目中
[03:32.00]找到更多新的节目
[03:34.00]SWIX 和 ALESIO 都在节目中
[03:36.00]找到更多新的节目
[03:38.00]SWIX 和 ALESIO 都在节目中
[03:40.00]找到更多新的节目
[03:42.00]SWIX 和 ALESIO 都在节目中
[03:44.00]找到更多新的节目
[03:46.00]SWIX 和 ALESIO 都在节目中
[03:48.00]找到更多新的节目
[03:50.00]SWIX 和 ALESIO 都在节目中
[03:52.00]找到更多新的节目
[03:54.00]SWIX 和 ALESIO 都在节目中
[03:56.00]找到更多新的节目
[03:58.00]SWIX 和 ALESIO 都在节目中
[04:00.00]找到更多新的节目
[04:02.00]SWIX 和 ALESIO 都在节目中
[04:04.00]找到更多新的节目
[04:06.00]SWIX 和 ALESIO 都在节目中
[04:08.00]找到更多新的节目
[04:10.00]SWIX 和 ALESIO 都在节目中
[04:12.00]找到更多新的节目
[04:14.00]SWIX 和 ALESIO 都在节目中
[04:16.00]找到更多新的节目
[04:18.00]SWIX 和 ALESIO 都在节目中
[04:20.00]找到更多新的节目
[04:22.00]SWIX 和 ALESIO 都在节目中
[04:24.00]找到更多新的节目
[04:26.00]SWIX 和 ALESIO 都在节目中
[04:28.00]找到更多新的节目
[04:30.00]SWIX 和 ALESIO 都在节目中
[04:32.00]找到更多新的节目
[04:34.00]SWIX 和 ALESIO 都在节目中
[04:36.00]找到更多新的节目
[04:38.00]SWIX 和 ALESIO 都在节目中
[04:40.00]找到更多新的节目
[04:42.00]SWIX 和 ALESIO 都在节目中
[04:44.00]找到更多新的节目
[04:46.00]SWIX 和 ALESIO 都在节目中
[04:48.00]找到更多新的节目
[04:50.00]SWIX 和 ALESIO 都在节目中
[04:52.00]找到更多新的节目
[04:54.00]SWIX 和 ALESIO 都在节目中
[04:56.00]找到更多新的节目
[04:58.00]SWIX 和 ALESIO 都在节目中
[05:00.00]找到更多新的节目
[05:02.00]SWIX 和 ALESIO 都在节目中
[05:04.00]找到更多新的节目
[05:06.00]SWIX 和 ALESIO 都在节目中
[05:08.00]找到更多新的节目
[05:10.00]SWIX 和 ALESIO 都在节目中
[05:12.00]找到更多新的节目
[05:14.00]SWIX 和 ALESIO 都在节目中
[05:16.00]找到更多新的节目
[05:18.00]SWIX 和 ALESIO 都在节目中
[05:20.00]找到更多新的节目
[05:22.00]SWIX 和 ALESIO 都在节目中
[05:24.00]找到更多新的节目
[05:26.00]SWIX 和 ALESIO 都在节目中
[05:28.00]找到更多新的节目
[05:30.00]SWIX 和 ALESIO 都在节目中
[05:32.00]找到更多新的节目
[05:34.00]SWIX 和 ALESIO 都在节目中
[05:36.00]找到更多新的节目
[05:38.00]SWIX 和 ALESIO 都在节目中
[05:40.00]找到更多新的节目
[05:42.00]SWIX 和 ALESIO 都在节目中
[05:44.00]找到更多新的节目
[05:46.00]SWIX 和 ALESIO 都在节目中
[05:48.00]找到更多新的节目
[05:50.00]SWIX 和 ALESIO 都在节目中
[05:52.00]找到更多新的节目
[05:54.00]SWIX 和 ALESIO 都在节目中
[05:56.00]找到更多新的节目
[05:58.00]SWIX 和 ALESIO 都在节目中
[06:00.00]找到更多新的节目
[06:02.00]SWIX 和 ALESIO 都在节目中
[06:04.00]找到更多新的节目
[06:06.00]SWIX 和 ALESIO 都在节目中
[06:08.00]找到更多新的节目
[06:10.00]SWIX 和 ALESIO 都在节目中
[06:12.00]找到更多新的节目
[06:14.00]SWIX 和 ALESIO 都在节目中
[06:16.00]找到更多新的节目
[06:18.00]SWIX 和 ALESIO 都在节目中
[06:20.00]找到更多新的节目
[06:22.00]SWIX 和 ALESIO 都在节目中
[06:24.00]找到更多新的节目
[06:26.00]SWIX 和 ALESIO 都在节目中
[06:28.00]找到更多新的节目
[06:30.00]SWIX 和 ALESIO 都在节目中
[06:32.00]找到更多新的节目
[06:34.00]SWIX 和 ALESIO 都在节目中
[06:36.00]找到更多新的节目
[06:38.00]SWIX 和 ALESIO 都在节目中
[06:40.00]找到更多新的节目
[06:42.00]SWIX 和 ALESIO 都在节目中
[06:44.00]找到更多新的节目
[06:46.00]SWIX 和 ALESIO 都在节目中
[06:48.00]找到更多新的节目
[06:50.00]SWIX 和 ALESIO 都在节目中
[06:52.00]找到更多新的节目
[06:54.00]SWIX 和 ALESIO 都在节目中
[06:56.00]找到更多新的节目
[06:58.00]SWIX 和 ALESIO 都在节目中
[07:00.00]找到更多新的节目
[07:02.00]SWIX 和 ALESIO 都在节目中
[07:04.00]找到更多新的节目
[07:06.00]SWIX 和 ALESIO 都在节目中
[07:08.00]找到更多新的节目
[07:10.00]SWIX 和 ALESIO 都在节目中
[07:12.00]找到更多新的节目
[07:14.00]SWIX 和 ALESIO 都在节目中
[07:16.00]找到更多新的节目
[07:18.00]SWIX 和 ALESIO 都在节目中
[07:20.00]找到更多新的节目
[07:22.00]SWIX 和 ALESIO 都在节目中
[07:24.00]找到更多新的节目
[07:26.00]SWIX 和 ALESIO 都在节目中
[07:28.00]找到更多新的节目
[07:30.00]SWIX 和 ALESIO 都在节目中
[07:32.00]找到更多新的节目
[07:34.00]SWIX 和 ALESIO 都在节目中
[07:36.00]找到更多新的节目
[07:38.00]SWIX 和 ALESIO 都在节目中
[07:40.00]找到更多新的节目
[07:42.00]SWIX 和 ALESIO 都在节目中
[07:44.00]找到更多新的节目
[07:46.00]SWIX 和 ALESIO 都在节目中
[07:48.00]找到更多新的节目
[07:50.00]SWIX 和 ALESIO 都在节目中
[07:52.00]找到更多新的节目
[07:54.00]SWIX 和 ALESIO 都在节目中
[07:56.00]找到更多新的节目
[07:58.00]找到更多新的节目
[08:00.00]SWIX 和 ALESIO 都在节目中
[08:02.00]找到更多新的节目
[08:04.00]SWIX 和 ALESIO 都在节目中
[08:06.00]找到更多新的节目
[08:08.00]SWIX 和 ALESIO 都在节目中
[08:10.00]找到更多新的节目
[08:12.00]SWIX 和 ALESIO 都在节目中
[08:14.00]找到更多新的节目
[08:16.00]SWIX 和 ALESIO 都在节目中
[08:18.00]找到更多新的节目
[08:20.00]SWIX 和 ALESIO 都在节目中
[08:22.00]找到更多新的节目
[08:24.00]SWIX 和 ALESIO 都在节目中
[08:26.00]找到更多新的节目
[08:28.00]SWIX 和 ALESIO 都在节目中
[08:30.00]找到更多新的节目
[08:32.00]SWIX 和 ALESIO 都在节目中
[08:34.00]找到更多新的节目
[08:36.00]SWIX 和 ALESIO 都在节目中
[08:38.00]找到更多新的节目
[08:40.00]SWIX 和 ALESIO 都在节目中
[08:42.00]找到更多新的节目
[08:44.00]SWIX 和 ALESIO 都在节目中
[08:46.00]找到更多新的节目
[08:48.00]SWIX 和 ALESIO 都在节目中
[08:50.00]找到更多新的节目
[08:52.00]SWIX 和 ALESIO 都在节目中
[08:54.00]找到更多新的节目
[08:56.00]SWIX 和 ALESIO 都在节目中
[08:58.00]找到更多新的节目
[09:00.00]SWIX 和 ALESIO 都在节目中
[09:02.00]找到更多新的节目
[09:04.00]SWIX 和 ALESIO 都在节目中
[09:06.00]找到更多新的节目
[09:08.00]SWIX 和 ALESIO 都在节目中
[09:10.00]找到更多新的节目
[09:12.00]SWIX 和 ALESIO 都在节目中
[09:14.00]找到更多新的节目
[09:16.00]SWIX 和 ALESIO 都在节目中
[09:18.00]找到更多新的节目
[09:20.00]SWIX 和 ALESIO 都在节目中
[09:22.00]找到更多新的节目
[09:24.00]空中
[09:26.00]找到更多新的节目
[09:28.00]SWIX 和 ALESIO 都在节目中
[09:30.00]找到更多新的节目
[09:32.00]SWIX 和 ALESIO 都在节目中
[09:34.00]找到更多新的节目
[09:36.00]SWIX 和 ALESIO 都在节目中
[09:38.00]找到更多新的节目
[09:40.00]SWIX 和 ALESIO 都在节目中
[09:42.00]找到更多新的节目
[09:44.00]SWIX 和 ALESIO 都在节目中
[09:46.00]找到更多新的节目
[09:48.00]SWIX 和 ALESIO 都在节目中
[09:50.00]找到更多新的节目
[09:52.00]SWIX 和 ALESIO 都在节目中
[09:54.00]找到更多新的节目
[09:56.00]SWIX 和 ALESIO 都在节目中
[09:58.00]找到更多新的节目
[10:00.00]SWIX 和 ALESIO 都在节目中
[10:02.00]找到更多新的节目
[10:04.00]SWIX 和 ALESIO 都在节目中
[10:06.00]找到更多新的节目
[10:08.00]SWIX 和 ALESIO 都在节目中
[10:10.00]找到更多新的节目
[10:12.00]SWIX 和 ALESIO 都在节目中
[10:14.00]找到更多新的节目
[10:16.00]SWIX 和 ALESIO 都在节目中
[10:18.00]找到更多新的节目
[10:20.00]SWIX 和 ALESIO 都在节目中
[10:22.00]找到更多新的节目
[10:24.00]SWIX 和 ALESIO 都在节目中
[10:26.00]找到更多新的节目
[10:28.00]SWIX 和 ALESIO 都在节目中
[10:30.00]找到更多新的节目
[10:32.00]SWIX 和 ALESIO 都在节目中
[10:34.00]找到更多新的节目
[10:36.00]SWIX 和 ALESIO 都在节目中
[10:38.00]找到更多新的节目
[10:40.00]SWIX 和 ALESIO 都在节目中
[10:42.00]找到更多新的节目
[10:44.00]SWIX 和 ALESIO 都在节目中
[10:46.00]找到更多新的节目
[10:48.00]SWIX 和 ALESIO 都在节目中
[10:50.00]找到更多新的节目
[10:52.00]SWIX 和 ALESIO 都在节目中
[10:54.00]找到更多新的节目
[10:56.00]SWIX 和 ALESIO 都在节目中
[10:58.00]找到更多新的节目
[11:00.00]SWIX 和 ALESIO 都在节目中
[11:02.00]找到更多新的节目
[11:04.00]SWIX 和 ALESIO 都在节目中
[11:06.00]找到更多新的节目
[11:08.00]SWIX 和 ALESIO 都在节目中
[11:10.00]找到更多新的节目
[11:12.00]SWIX 和 ALESIO 都在节目中
[11:14.00]找到更多新的节目
[11:16.00]SWIX 和 ALESIO 都在节目中
[11:18.00]找到更多新的节目
[11:20.00]SWIX 和 ALESIO 都在节目中
[11:22.00]找到更多新的节目
[11:24.00]SWIX 和 ALESIO 都在节目中
[11:26.00]找到更多新的节目
[11:28.00]SWIX 和 ALESIO 都在节目中
[11:30.00]找到更多新的节目
[11:32.00]SWIX 和 ALESIO 都在节目中
[11:34.00]找到更多新的节目
[11:36.00]SWIX 和 ALESIO 都在节目中
[11:38.00]找到更多新的节目
[11:40.00]SWIX 和 ALESIO 都在节目中
[11:42.00]找到更多新的节目
[11:44.00]SWIX 和 ALESIO 都在节目中
[11:46.00]找到更多新的节目
[11:48.00]SWIX 和 ALESIO 都在节目中
[11:50.00]找到更多新的节目
[11:52.00]SWIX 和 ALESIO 都在节目中
[11:54.00]找到更多新的节目
[11:56.00]SWIX 和 ALESIO 都在节目中
[11:58.00]找到更多新的节目
[12:00.00]SWIX 和 ALESIO 都在节目中
[12:02.00]找到更多新的节目
[12:04.00]SWIX 和 ALESIO 都在节目中
[12:06.00]找到更多新的节目
[12:08.00]SWIX 和 ALESIO 都在节目中
[12:10.00]找到更多新的节目
[12:12.00]SWIX 和 ALESIO 都在节目中
[12:14.00]找到更多新的节目
[12:16.00]SWIX 和 ALESIO 都在节目中
[12:18.00]找到更多新的节目
[12:20.00]SWIX 和 ALESIO 都在节目中
[12:22.00]找到更多新的节目
[12:24.00]SWIX 和 ALESIO 都在节目中
[12:26.00]找到更多新的节目
[12:28.00]SWIX 和 ALESIO 都在节目中
[12:30.00]找到更多新的节目
[12:32.00]SWIX 和 ALESIO 都在节目中
[12:34.00]找到更多新的节目
[12:36.00]SWIX 和 ALESIO 都在节目中
[12:38.00]找到更多新的节目
[12:40.00]找到更多新的节目
[12:42.00]找到更多新的节目
[12:44.00]找到更多新的节目
[12:46.00]找到更多新的节目
[12:48.00]找到更多新的节目
[12:50.00]找到更多新的节目
[12:52.00]找到更多新的节目
[12:54.00]找到更多新的节目
[12:56.00]找到更多新的节目
[12:58.00]找到更多新的节目
[13:00.00]找到更多新的节目
[13:02.00]找到更多新的节目
[13:04.00]找到更多新的节目
[13:06.00]找到更多新的节目
[13:08.00]找到更多新的节目
[13:10.00]找到更多新的节目
[13:12.00]找到更多新的节目
[13:14.00]找到更多新的节目
[13:16.00]找到更多新的节目
[13:18.00]找到更多新的节目
[13:20.00]找到更多新的节目
[13:22.00]找到更多新的节目
[13:24.00]找到更多新的节目
[13:26.00]找到更多新的节目
[13:28.00]找到更多新的节目
[13:30.00]找到更多新的节目
[13:32.00]找到更多新的节目
[13:34.00]找到更多新的节目
[13:36.00]找到更多新的节目
[13:38.00]找到更多新的节目
[13:40.00]找到更多新的节目
[13:42.00]找到更多新的节目
[13:44.00]找到更多新的节目
[13:46.00]找到更多新的节目
[13:48.00]找到更多新的节目
[13:50.00]找到更多新的节目
[13:52.00]找到更多新的节目
[13:54.00]找到更多新的节目
[13:56.00]找到更多新的节目
[13:58.00]找到更多新的节目
[14:00.00]找到更多新的节目
[14:02.00]找到更多新的节目
[14:04.00]找到更多新的节目
[14:06.00]找到更多新的节目
[14:08.00]找到更多新的节目
[14:10.00]找到更多新的节目
[14:12.00]找到更多新的节目
[14:14.00]找到更多新的节目
[14:16.00]找到更多新的节目
[14:18.00]找到更多新的节目
[14:20.00]找到更多新的节目
[14:22.00]找到更多新的节目
[14:24.00]找到更多新的节目
[14:26.00]找到更多新的节目
[14:28.00]找到更多新的节目
[14:30.00]找到更多新的节目
[14:32.00]找到更多新的节目
[14:34.00]找到更多新的节目
[14:36.00]找到更多新的节目
[14:38.00]找到更多新的节目
[14:40.00]找到更多新的节目
[14:42.00]找到更多新的节目
[14:44.00]找到更多新的节目
[14:46.00]找到更多新的节目
[14:48.00]找到更多新的节目
[14:50.00]找到更多新的节目
[14:52.00]找到更多新的节目
[14:54.00]找到更多新的节目
[14:56.00]找到更多新的节目
[14:58.00]找到更多新的节目
[15:00.00]找到更多新的节目
[15:02.00]找到更多新的节目
[15:04.00]找到更多新的节目
[15:06.00]找到更多新的节目
[15:08.00]找到更多新的节目
[15:10.00]找到更多新的节目
[15:12.00]找到更多新的节目
[15:14.00]找到更多新的节目
[15:16.00]找到更多新的节目
[15:18.00]找到更多新的节目
[15:20.00]找到更多新的节目
[15:22.00]找到更多新的节目
[15:24.00]找到更多新的节目
[15:26.00]找到更多新的节目
[15:28.00]找到更多新的节目
[15:30.00]找到更多新的节目
[15:32.00]找到更多新的节目
[15:34.00]找到更多新的节目
[15:36.00]找到更多新的节目
[15:38.00]找到更多新的节目
[15:40.00]找到更多新的节目
[15:42.00]找到更多新的节目
[15:44.00]找到更多新的节目
[15:46.00]找到更多新的节目
[15:48.00]找到更多新的节目
[15:50.00]找到更多新的节目
[15:52.00]找到更多新的节目
[15:54.00]找到更多新的节目
[15:56.00]找到更多新的节目
[15:58.00]找到更多新的节目
[16:00.00]找到更多新的节目
[16:02.00]找到更多新的节目
[16:04.00]找到更多新的节目
[16:06.00]找到更多新的节目
[16:08.00]找到更多新的节目
[16:10.00]找到更多新的节目
[16:12.00]找到更多新的节目
[16:14.00]找到更多新的节目
[16:16.00]找到更多新的节目
[16:18.00]找到更多新的节目
[16:20.00]找到更多新的节目
[16:22.00]找到更多新的节目
[16:24.00]找到更多新的节目
[16:26.00]找到更多新的节目
[16:28.00]找到更多新的节目
[16:30.00]找到更多新的节目
[16:32.00]找到更多新的节目
[16:34.00]找到更多新的节目
[16:36.00]找到更多新的节目
[16:38.00]找到更多新的节目
[16:40.00]找到更多新的节目
[16:42.00]找到更多新的节目
[16:44.00]找到更多新的节目
[16:46.00]找到更多新的节目
[16:48.00]找到更多新的节目
[16:50.00]找到更多新的节目
[16:52.00]找到更多新的节目
[16:54.00]找到更多新的节目
[16:56.00]找到更多新的节目
[16:58.00]找到更多新的节目
[17:00.00]找到更多新的节目
[17:02.00]找到更多新的节目
[17:04.00]找到更多新的节目
[17:06.00]找到更多新的节目
[17:08.00]找到更多新的节目
[17:10.00]找到更多新的节目
[17:12.00]找到更多新的节目
[17:14.00]找到更多新的节目
[17:16.00]找到更多新的节目
[17:18.00]找到更多新的节目
[17:20.00]找到更多新的节目
[17:22.00]找到更多新的节目
[17:24.00]找到更多新的节目
[17:26.00]找到更多新的节目
[17:28.00]找到更多新的节目
[17:30.00]找到更多新的节目
[17:32.00]找到更多新的节目
[17:34.00]找到更多新的节目
[17:36.00]找到更多新的节目
[17:38.00]找到更多新的节目
[17:40.00]找到更多新的节目
[17:42.00]找到更多新的节目
[17:44.00]找到更多新的节目
[17:46.00]找到更多新的节目
[17:48.00]找到更多新的节目
[17:50.00]找到更多新的节目
[17:52.00]找到更多新的节目
[17:54.00]找到更多新的节目
[17:56.00]找到更多新的节目
[17:58.00]找到更多新的节目
[18:00.00]找到更多新的节目
[18:02.00]找到更多新的节目
[18:04.00]找到更多新的节目
[18:06.00]找到更多新的节目
[18:08.00]找到更多新的节目
[18:10.00]找到更多新的节目
[18:12.00]找到更多新的节目
[18:14.00]找到更多新的节目
[18:16.00]找到更多新的节目
[18:18.00]找到更多新的节目
[18:20.00]找到更多新的节目
[18:22.00]找到更多新的节目
[18:24.00]找到更多新的节目
[18:26.00]找到更多新的节目
[18:28.00]找到更多新的节目
[18:30.00]找到更多新的节目
[18:32.00]找到更多新的节目
[18:34.00]找到更多新的节目
[18:36.00]找到更多新的节目
[18:38.00]找到更多新的节目
[18:40.00]找到更多新的节目
[18:42.00]找到更多新的节目
[18:44.00]找到更多新的节目
[18:46.00]找到更多新的节目
[18:48.00]找到更多新的节目
[18:50.00]找到更多新的节目
[18:52.00]找到更多新的节目
[18:54.00]找到更多新的节目
[18:56.00]找到更多新的节目
[18:58.00]找到更多新的节目
[19:00.00]找到更多新的节目
[19:02.00]找到更多新的节目
[19:04.00]找到更多新的节目
[19:06.00]找到更多新的节目
[19:08.00]找到更多新的节目
[19:10.00]找到更多新的节目
[19:12.00]找到更多新的节目
[19:14.00]找到更多新的节目
[19:16.00]找到更多新的节目
[19:18.00]找到更多新的节目
[19:20.00]找到更多新的节目
[19:22.00]找到更多新的节目
[19:24.00]找到更多新的节目
[19:26.00]找到更多新的节目
[19:28.00]找到更多新的节目
[19:30.00]找到更多新的节目
[19:32.00]找到更多新的节目
[19:34.00]找到更多新的节目
[19:36.00]找到更多新的节目
[19:38.00]找到更多新的节目
[19:40.00]找到更多新的节目
[19:42.00]找到更多新的节目
[19:44.00]找到更多新的节目
[19:46.00]找到更多新的节目
[19:48.00]找到更多新的节目
[19:50.00]找到更多新的节目
[19:52.00]找到更多新的节目
[19:54.00]找到更多新的节目
[19:56.00]找到更多新的节目
[19:58.00]找到更多新的节目
[20:00.00]找到更多新的节目
[20:02.00]找到更多新的节目
[20:04.00]找到更多新的节目
[20:06.00]找到更多新的节目
[20:08.00]找到更多新的节目
[20:10.00]找到更多新的节目
[20:12.00]找到更多新的节目
[20:14.00]找到更多新的节目
[20:16.00]找到更多新的节目
[20:18.00]找到更多新的节目
[20:20.00]找到更多新的节目
[20:22.00]找到更多新的节目
[20:24.00]找到更多新的节目
[20:26.00]找到更多新的节目
[20:28.00]找到更多新的节目
[20:30.00]找到更多新的节目
[20:32.00]找到更多新的节目
[20:34.00]找到更多新的节目
[20:36.00]找到更多新的节目
[20:38.00]找到更多新的节目
[20:40.00]找到更多新的节目
[20:42.00]找到更多新的节目
[20:44.00]找到更多新的节目
[20:46.00]找到更多新的节目
[20:48.00]找到更多新的节目
[20:50.00]找到更多新的节目
[20:52.00]找到更多新的节目
[20:54.00]找到更多新的节目
[20:56.00]找到更多新的节目
[20:58.00]找到更多新的节目
[21:00.00]找到更多新的节目
[21:02.00]找到更多新的节目
[21:04.00]找到更多新的节目
[21:06.00]找到更多新的节目
[21:08.00]找到更多新的节目
[21:10.00]找到更多新的节目
[21:12.00]找到更多新的节目
[21:14.00]找到更多新的节目
[21:16.00]找到更多新的节目
[21:18.00]找到更多新的节目
[21:20.00]找到更多新的节目
[21:22.00]找到更多新的节目
[21:24.00]找到更多新的节目
[21:26.00]找到更多新的节目
[21:28.00]找到更多新的节目
[21:30.00]找到更多新的节目
[21:32.00]找到更多新的节目
[21:34.00]找到更多新的节目
[21:36.00]找到更多新的节目
[21:38.00]找到更多新的节目
[21:40.00]找到更多新的节目
[21:42.00]找到更多新的节目
[21:44.00]找到更多新的节目
[21:46.00]找到更多新的节目
[21:48.00]找到更多新的节目
[21:50.00]找到更多新的节目
[21:52.00]找到更多新的节目
[21:54.00]找到更多新的节目
[21:56.00]找到更多新的节目
[21:58.00]找到更多新的节目
[22:00.00]找到更多新的节目
[22:02.00]找到更多新的节目
[22:04.00]找到更多新的节目
[22:06.00]找到更多新的节目
[22:08.00]找到更多新的节目
[22:10.00]找到更多新的节目
[22:12.00]找到更多新的节目
[22:14.00]找到更多新的节目
[22:16.00]找到更多新的节目
[22:18.00]找到更多新的节目
[22:20.00]找到更多新的节目
[22:22.00]找到更多新的节目
[22:24.00]找到更多新的节目
[22:26.00]找到更多新的节目
[22:28.00]找到更多新的节目
[22:30.00]找到更多新的节目
[22:32.00]找到更多新的节目
[22:34.00]找到更多新的节目
[22:36.00]找到更多新的节目
[22:38.00]找到更多新的节目
[22:40.00]找到更多新的节目
[22:42.00]找到更多新的节目
[22:44.00]找到更多新的节目
[22:46.00]找到更多新的节目
[22:48.00]找到更多新的节目
[22:50.00]找到更多新的节目
[22:52.00]找到更多新的节目
[22:54.00]找到更多新的节目
[22:56.00]找到更多新的节目
[22:58.00]找到更多新的节目
[23:00.00]找到更多新的节目
[23:02.00]找到更多新的节目
[23:04.00]找到更多新的节目
[23:06.00]找到更多新的节目
[23:08.00]找到更多新的节目
[23:10.00]找到更多新的节目
[23:12.00]找到更多新的节目
[23:14.00]找到更多新的节目
[23:16.00]找到更多新的节目
[23:18.00]找到更多新的节目
[23:20.00]找到更多新的节目
[23:22.00]找到更多新的节目
[23:24.00]找到更多新的节目
[23:26.00]找到更多新的节目
[23:28.00]找到更多新的节目
[23:30.00]找到更多新的节目
[23:32.00]找到更多新的节目
[23:34.00]找到更多新的节目
[23:36.00]找到更多新的节目
[23:38.00]找到更多新的节目
[23:40.00]找到更多新的节目
[23:42.00]找到更多新的节目
[23:44.00]找到更多新的节目
[23:46.00]找到更多新的节目
[23:48.00]找到更多新的节目
[23:50.00]找到更多新的节目
[23:52.00]找到更多新的节目
[23:54.00]找到更多新的节目
[23:56.00]找到更多新的节目
[23:58.00]找到更多新的节目
[24:00.00]找到更多新的节目
[24:02.00]找到更多新的节目
[24:04.00]找到更多新的节目
[24:06.00]找到更多新的节目
[24:08.00]找到更多新的节目
[24:10.00]找到更多新的节目
[24:12.00]找到更多新的节目
[24:14.00]找到更多新的节目
[24:16.00]找到更多新的节目
[24:18.00]找到更多新的节目
[24:20.00]找到更多新的节目
[24:22.00]找到更多新的节目
[24:24.00]找到更多新的节目
[24:26.00]找到更多新的节目
[24:28.00]找到更多新的节目
[24:30.00]找到更多新的节目
[24:32.00]找到更多新的节目
[24:34.00]找到更多新的节目
[24:36.00]找到更多新的节目
[24:38.00]找到更多新的节目
[24:40.00]找到更多新的节目
[24:42.00]找到更多新的节目
[24:44.00]找到更多新的节目
[24:46.00]找到更多新的节目
[24:48.00]找到更多新的节目
[24:50.00]找到更多新的节目
[24:52.00]找到更多新的节目
[24:54.00]找到更多新的节目
[24:56.00]找到更多新的节目
[24:58.00]找到更多新的节目
[25:00.00]找到更多新的节目
[25:02.00]找到更多新的节目
[25:04.00]找到更多新的节目
[25:06.00]找到更多新的节目
[25:08.00]找到更多新的节目
[25:10.00]找到更多新的节目
[25:12.00]找到更多新的节目
[25:14.00]找到更多新的节目
[25:16.00]找到更多新的节目
[25:18.00]找到更多新的节目
[25:20.00]找到更多新的节目
[25:22.00]找到更多新的节目
[25:24.00]找到更多新的节目
[25:26.00]找到更多新的节目
[25:28.00]找到更多新的节目
[25:30.00]找到更多新的节目
[25:32.00]找到更多新的节目
[25:34.00]找到更多新的节目
[25:36.00]找到更多新的节目
[25:38.00]找到更多新的节目
[25:40.00]找到更多新的节目
[25:42.00]找到更多新的节目
[25:44.00]找到更多新的节目
[25:46.00]找到更多新的节目
[25:48.00]找到更多新的节目
[25:50.00]找到更多新的节目
[25:52.00]找到更多新的节目
[25:54.00]找到更多新的节目
[25:56.00]找到更多新的节目
[25:58.00]找到更多新的节目
[26:00.00]找到更多新的节目
[26:02.00]找到更多新的节目
[26:04.00]找到更多新的节目
[26:06.00]找到更多新的节目
[26:08.00]找到更多新的节目
[26:10.00]找到更多新的节目
[26:12.00]找到更多新的节目
[26:14.00]找到更多新的节目
[26:16.00]找到更多新的节目
[26:18.00]找到更多新的节目
[26:20.00]找到更多新的节目
[26:22.00]找到更多新的节目
[26:24.00]找到更多新的节目
[26:26.00]找到更多新的节目
[26:28.00]找到更多新的节目
[26:30.00]找到更多新的节目
[26:32.00]找到更多新的节目
[26:34.00]找到更多新的节目
[26:36.00]找到更多新的节目
[26:38.00]找到更多新的节目
[26:40.00]找到更多新的节目
[26:42.00]找到更多新的节目
[26:44.00]找到更多新的节目
[26:46.00]找到更多新的节目
[26:48.00]找到更多新的节目
[26:50.00]找到更多新的节目
[26:52.00]找到更多新的节目
[26:54.00]找到更多新的节目
[26:56.00]找到更多新的节目
[26:58.00]找到更多新的节目
[27:00.00]找到更多新的节目
[27:02.00]找到更多新的节目
[27:04.00]找到更多新的节目
[27:06.00]找到更多新的节目
[27:08.00]找到更多新的节目
[27:10.00]找到更多新的节目
[27:12.00]找到更多新的节目
[27:14.00]找到更多新的节目
[27:16.00]找到更多新的节目
[27:18.00]找到更多新的节目
[27:20.00]找到更多新的节目
[27:22.00]找到更多新的节目
[27:24.00]找到更多新的节目
[27:26.00]找到更多新的节目
[27:28.00]找到更多新的节目
[27:30.00]找到更多新的节目
[27:32.00]找到更多新的节目
[27:34.00]找到更多新的节目
[27:36.00]找到更多新的节目
[27:38.00]找到更多新的节目
[27:40.00]找到更多新的节目
[27:42.00]找到更多新的节目
[27:44.00]找到更多新的节目
[27:46.00]找到更多新的节目
[27:48.00]找到更多新的节目
[27:50.00]找到更多新的节目
[27:52.00]找到更多新的节目
[27:54.00]找到更多新的节目
[27:56.00]找到更多新的节目
[27:58.00]找到更多新的节目
[28:00.00]找到更多新的节目
[28:02.00]找到更多新的节目
[28:04.00]找到更多新的节目
[28:06.00]找到更多新的节目
[28:08.00]找到更多新的节目
[28:10.00]找到更多新的节目
[28:12.00]找到更多新的节目
[28:14.00]找到更多新的节目
[28:16.00]找到更多新的节目
[28:18.00]找到更多新的节目
[28:20.00]找到更多新的节目
[28:22.00]找到更多新的节目
[28:24.00]找到更多新的节目
[28:26.00]找到更多新的节目
[28:28.00]找到更多新的节目
[28:30.00]找到更多新的节目
[28:32.00]找到更多新的节目
[28:34.00]找到更多新的节目
[28:36.00]找到更多新的节目
[28:38.00]找到更多新的节目
[28:40.00]找到更多新的节目
[28:42.00]找到更多新的节目
[28:44.00]找到更多新的节目
[28:46.00]找到更多新的节目
[28:48.00]找到更多新的节目
[28:50.00]找到更多新的节目
[28:52.00]找到更多新的节目
[28:54.00]找到更多新的节目
[28:56.00]找到更多新的节目
[28:58.00]找到更多新的节目
[29:00.00]找到更多新的节目
[29:02.00]找到更多新的节目
[29:04.00]找到更多新的节目
[29:06.00]找到更多新的节目
[29:08.00]找到更多新的节目
[29:10.00]找到更多新的节目
[29:12.00]找到更多新的节目
[29:14.00]找到更多新的节目
[29:16.00]找到更多新的节目
[29:18.00]找到更多新的节目
[29:20.00]找到更多新的节目
[29:22.00]找到更多新的节目
[29:24.00]找到更多新的节目
[29:26.00]找到更多新的节目
[29:28.00]找到更多新的节目
[29:30.00]找到更多新的节目
[29:32.00]找到更多新的节目
[29:34.00]找到更多新的节目
[29:36.00]找到更多新的节目
[29:38.00]找到更多新的节目
[29:40.00]找到更多新的节目
[29:42.00]找到更多新的节目
[29:44.00]找到更多新的节目
[29:46.00]找到更多新的节目
[29:48.00]找到更多新的节目
[29:50.00]找到更多新的节目
[29:52.00]找到更多新的节目
[29:54.00]找到更多新的节目
[29:56.00]找到更多新的节目
[29:58.00]找到更多新的节目
[30:00.00]找到更多新的节目
[30:02.00]找到更多新的节目
[30:04.00]找到更多新的节目
[30:06.00]找到更多新的节目
[30:08.00]找到更多新的节目
[30:10.00]找到更多新的节目
[30:12.00]找到更多新的节目
[30:14.00]找到更多新的节目
[30:16.00]找到更多新的节目
[30:18.00]找到更多新的节目
[30:20.00]找到更多新的节目
[30:22.00]找到更多新的节目
[30:24.00]找到更多新的节目
[30:26.00]找到更多新的节目
[30:28.00]找到更多新的节目
[30:30.00]找到更多新的节目
[30:32.00]找到更多新的节目
[30:34.00]找到更多新的节目
[30:36.00]找到更多新的节目
[30:38.00]找到更多新的节目
[30:40.00]找到更多新的节目
[30:42.00]找到更多新的节目
[30:44.00]找到更多新的节目
[30:46.00]找到更多新的节目
[30:48.00]找到更多新的节目
[30:50.00]所以我们现在 slowly
[30:52.00]改变了 现在我们现在慢慢改变了
[30:54.00]我在将来归咖巴斯
[30:56.00]没有改变了 没有改变了
[30:58.00]但是现在我们
[31:00.00]有一个世界的 一个世界的
[31:02.00]有一个世界的 哪里有Claude
[31:04.00]还有Gemnite 还有Gypsy4
[31:06.00]希望更多的 会变得更多
[31:08.00]希望更多的 希望更多的
[31:10.00]所以 说说 说说
[31:12.00]我们的视频 视频 视频
[31:14.00]所以 非常大 非常大 因为
[31:16.00]我们的视频 也不可以用
[31:18.00]但我认为
[31:20.00]大陆的社区
[31:22.00]要他们继续建设
[31:24.00]然后他们必须找一些方法
[31:26.00]才能发现他们会做的
[31:28.00]所以他们明白
[31:30.00]他们必须解决他们要的
[31:32.00]但是我们的视频 也有
[31:34.00]Mistral 也有
[31:36.00]Grock 现在
[31:38.00]Grock 1 从 从 oktober
[31:40.00]是公司
[31:41.00]对对对对
[31:42.00]你以为Grock是Grock the chip company
[31:44.00]Grock the chip company
[31:46.00]当然 浪漫3是
[31:48.00]所有人都在问
[31:50.00]我的感觉是
[31:52.00]小时候 宋可伯
[31:54.00]刚才说了浪漫3
[31:56.00]说了 至少从
[31:58.00]一个想法的选择
[32:00.00]他不想 怎么做
[32:02.00]要保持
[32:04.00]能够保持 能够保持
[32:06.00]Mistral 也想想
[32:08.00]你去到 怎么
[32:10.00]他 能够发展
[32:12.00]每个人都好 任何
[32:14.00]对 从我听过
[32:16.00]在GDC 浪漫3
[32:18.00]最大的模式是
[32:20.00]260-300billion
[32:22.00]所以那是大部分的
[32:24.00]那不是一个开放模式
[32:26.00]你不能给人们
[32:28.00]300billion的模式
[32:30.00]要用它 非常有力量
[32:32.00]所以我认为
[32:34.00]它是 可能是开放模式
[32:36.00]但那是一个不同的问题
[32:38.00]对对对
[32:40.00]它是 比他们做的
[32:42.00]在开放模式
[32:44.00]在浪漫上
[32:46.00]你能够使用
[32:48.00]开放的AI
[32:50.00]开放的安逸
[32:52.00]一些公司在
[32:54.00]中央的强硬度
[32:56.00]所以我们
[32:58.00]在Buckets 上
[33:00.00]他们做了很多
[33:02.00]他们比PyDorch 更好
[33:04.00]那是 可能
[33:06.00]在弥術上
[33:08.00]可能是 可能是
[33:10.00]I love the duck destroying
[33:12.00]a lot of monopolies arc
[33:14.00]it's been very entertaining
[33:16.00]let's bridge into the
[33:18.00]big tech side of this
[33:20.00]I think when I did my episode
[33:22.00]I added this as an additional war
[33:24.00]that's something I'm paying attention to
[33:26.00]so we've got
[33:28.00]Microsoft's moves with inflection
[33:30.00]which I think potentially are
[33:32.00]being read as
[33:34.00]a shift vis-a-vis the relationship
[33:36.00]with open AI
[33:38.00]missure a large relationship
[33:40.00]seems to reinforce as well
[33:42.00]we have apple potentially
[33:44.00]entering the race finally
[33:46.00]giving up project titan
[33:48.00]and trying to spend more effort on this
[33:50.00]although counterpoint
[33:52.00]we also have them talking about
[33:54.00]there being reports of a deal with google
[33:56.00]which is interesting to see
[33:58.00]what their strategy there is
[34:00.00]and then metas been largely quiet
[34:02.00]we just talked about the main piece
[34:04.00]but there's spoilers
[34:06.00]one of those things has been most interesting
[34:08.00]to you guys as you think about
[34:10.00]what's going to shake out for the rest of this year
[34:12.00]let's take a crack
[34:14.00]the reason we don't have a fifth war
[34:16.00]for the big tech wars
[34:18.00]that's one of those things where I just feel
[34:20.00]we don't cover differently
[34:22.00]from other media channels
[34:24.00]I guess
[34:26.00]in our entire interest
[34:28.00]we try not to cover the big tech gamethrones
[34:30.00]or it's proxied through
[34:32.00]all the other four wars anyway
[34:34.00]there's just a lot of overlap
[34:36.00]but yeah I think absolutely personally
[34:38.00]the most interesting one is apple entering the race
[34:40.00]they actually release, they are announced
[34:42.00]their first large language model that they train themselves
[34:44.00]it's like a 30 billion multimodal model
[34:46.00]people weren't that impressed
[34:48.00]but it was like the first time
[34:50.00]that apple has kind of showcased that
[34:52.00]we're training large models in house as well
[34:54.00]of course they might be
[34:56.00]doing this deal with google
[34:58.00]it sounds very sort of rumourary to me
[35:00.00]and it's probably if it's on device
[35:02.00]it's going to be a smaller model
[35:03.00]it's going to be smarter auto complete
[35:05.00]I don't know what to say
[35:07.00]I'm still here dealing with
[35:09.00]Siri which hasn't
[35:11.00]probably hasn't been updated since
[35:13.00]God knows when it was introduced
[35:15.00]it's horrible and it
[35:17.00]makes me so angry
[35:19.00]one as an apple customer and user
[35:21.00]I'm just hoping for better ai on apple itself
[35:23.00]but two they are
[35:25.00]the gold standard
[35:27.00]when it comes to local devices
[35:29.00]personal compute and trust
[35:31.00]you trust them with your data
[35:33.00]and I think
[35:35.00]that's what a lot of people are looking for in ai
[35:37.00]that they love the benefits of ai
[35:39.00]they don't love the downsides
[35:41.00]which is that you have to send all your data
[35:43.00]to some clouds somewhere and some of this data
[35:45.00]that we're going to feed ai is the most personal data there is
[35:47.00]so apple being
[35:49.00]one of the most trusted personal
[35:51.00]data companies I think it's very important
[35:53.00]that they enter the ai race
[35:55.00]and I hope to see more out of them
[35:57.00]to me the biggest question
[35:59.00]it's like who's paying who
[36:01.00]because for the browsers
[36:03.00]google pays apple like 18
[36:05.00]20 billion every year
[36:07.00]to be the default browser
[36:09.00]is google going to pay you to have javanai
[36:11.00]or is apple paying google to have javanai
[36:13.00]I think that's like what I'm most interested
[36:15.00]to figure out because with the browsers
[36:17.00]it's like it's the entry point
[36:19.00]to the thing so it's really valuable
[36:21.00]to be the default that's what google pays
[36:23.00]but I wonder if the perception in ai
[36:25.00]is going to be like hey
[36:27.00]you have a good local model on my phone
[36:29.00]to be worth me purchasing your device
[36:31.00]and that's going to drive apple
[36:33.00]to be the one buying the model
[36:35.00]but then like Sean said
[36:37.00]they're doing the mm1 themselves
[36:39.00]are they saying we do models
[36:41.00]but they're not as good as the google ones
[36:43.00]I don't know the whole thing is really confusing
[36:45.00]but it makes for a great meme
[36:47.00]material on twitter
[36:49.00]I think like
[36:51.00]they are possibly more than
[36:53.00]open ai and mic microsoft and amazon
[36:55.00]they are the most full stack company there is
[36:57.00]in computing
[36:59.00]and so
[37:01.00]like they own the chips man
[37:03.00]like they manufacture everything
[37:05.00]so if there was a company
[37:07.00]that could seriously challenge
[37:09.00]the other ai players it would be apple
[37:11.00]and it's
[37:13.00]I don't think it's as hard as self-driving
[37:15.00]so like maybe they've just been
[37:17.00]investing in the wrong thing this whole time
[37:19.00]Wallstreet certainly thinks so
[37:21.00]Wallstreet love that move man
[37:23.00]there's a big sigh of relief
[37:25.00]well let's move away
[37:27.00]from sort of the big stuff
[37:29.00]I think to both of your points
[37:31.00]can I drop one factoid
[37:33.00]about this wallstreet thing
[37:35.00]I went and looked at
[37:37.00]when from being a VR company
[37:39.00]to an ai company
[37:41.00]and I think
[37:43.00]the stock
[37:45.00]I'm trying to look up the details now
[37:47.00]the stock has gone up 187%
[37:49.00]since limo one
[37:51.00]$830 billion in market value
[37:53.00]created in the past year
[37:55.00]if you haven't seen that chart
[37:59.00]it's actually remarkable if you draw
[38:01.00]a little arrow on it
[38:03.00]it's likeno we're an ai company now
[38:05.00]forget the VR thing
[38:07.00]it isn't interesting
[38:11.00]no I think unless you called it
[38:13.00]zuck's disruptor arc or whatever
[38:15.00]he really does
[38:17.00]he is in the midst of a total
[38:19.00]it's a redemption arc or it's just
[38:21.00]it's something different where
[38:23.00]he's sort of the spoiler like
[38:25.00]people loved him
[38:27.00]just freestyle talking about why he thought
[38:29.00]they had a better headset than apple
[38:31.00]even if they didn't agree they just loved
[38:33.00]he was going direct to camera and talking about it
[38:35.00]for five minutes or whatever
[38:37.00]that's a fascinating shift that I don't think
[38:39.00]anyone had on their bingo card
[38:41.00]whatever two years ago
[38:43.00]it's still there in cn5 d-long
[38:45.00]don't write it off
[38:47.00]we need to see him fight in the coliseum
[38:49.00]no I think in terms of
[38:51.00]self
[38:53.00]management life leadership
[38:55.00]there's a lot of lessons to learn from him
[38:57.00]you might kind of quibble
[38:59.00]with the social impact of facebook
[39:01.00]but just himself
[39:03.00]in terms of personal growth
[39:05.00]and perseverance through
[39:07.00]a lot of change
[39:09.00]everyone throwing stuff his way
[39:11.00]I think there's a lot to say about
[39:13.00]to learn from zuck
[39:15.00]he's my age
[39:17.00]awesome
[39:19.00]so one of the big things
[39:21.00]that I think you guys have
[39:23.00]distinct and unique insight into
[39:25.00]being where you are and where you work on
[39:27.00]iswhat developers
[39:29.00]are getting really excited about right now
[39:31.00]and by that I mean on the one hand
[39:33.00]certainly start ups who are actually
[39:35.00]formalized and formed to start ups
[39:37.00]but also just in terms of
[39:39.00]what people are spending their nights and weekends on
[39:41.00]what they're coming to hackathons to do
[39:43.00]and you know I think it's a
[39:45.00]it's such a fascinating indicator
[39:47.00]for where things are headed like
[39:49.00]if you zoom back a year
[39:51.00]right now was right when everyone was getting
[39:53.00]so so excited about
[39:55.00]ai agent stuff
[39:57.00]auto gpt and baby agi and these things were like
[39:59.00]if you dropped anything on youtube about those
[40:01.00]like instantly tens of thousands of views
[40:03.00]I know because I had like
[40:05.00]a 50,000 view video
[40:07.00]like the second day that I was doing
[40:09.00]the show on youtube you know because I was talking about
[40:11.00]auto gpt and so anyways
[40:13.00]you know obviously that's sort of not totally
[40:15.00]come to fruition yet but what are some of the
[40:17.00]trends and what you guys are seeing in terms of
[40:19.00]people's interest and what people are building
[40:21.00]I can start maybe with the agents part
[40:23.00]and then I know Sean is doing a
[40:25.00]diffusion meetup tonight there's
[40:27.00]a lot of different things
[40:29.00]the agent wave has been the most
[40:31.00]interesting kind of like dream
[40:33.00]to reality
[40:35.00]arc so auto gpt I think
[40:37.00]they went from zero to like
[40:39.00]125,000 get up stars in six weeks
[40:41.00]and then one year later
[40:43.00]they have 150,000
[40:45.00]stars so there's kind of been a big
[40:47.00]plot so I mean you might say
[40:49.00]there's just not that many people that can
[40:51.00]start it you know everybody already started
[40:53.00]but the promise of
[40:55.00]hey I'll just give you a goal
[40:57.00]and you do it I think it's like
[40:59.00]amazing to get people's
[41:01.00]imagination going you know
[41:03.00]they're like oh wow this is
[41:05.00]this is awesome everybody
[41:07.00]can try this to do anything
[41:09.00]but then as technologists
[41:11.00]you're like well that's
[41:13.00]that's just like not possible you know
[41:15.00]we would have like solved everything and
[41:17.00]I think it takes a little bit to go from
[41:19.00]the promise and the hope
[41:21.00]that people show you to then
[41:23.00]trinate yourself and going back to say
[41:25.00]okay this is not really working for me and
[41:27.00]David won from adept you know
[41:29.00]they in our episode he specifically said
[41:31.00]we don't want to do a bottom sub product
[41:33.00]you knowwe don't want something that everybody
[41:35.00]could try because it's really hard to get it
[41:37.00]to be reliable so
[41:39.00]we're seeing a lot of companies
[41:41.00]doing vertical agents that
[41:43.00]are narrow for a specific
[41:45.00]domain and they're very good at something
[41:47.00]Myconover who was at Databricks before
[41:49.00]is also a friend of Layton space
[41:51.00]he's doing this new company go bright with
[41:53.00]doing AI agents for financial research
[41:55.00]and that's it you know and
[41:57.00]they're doing very well there are
[41:59.00]other companies doing it in
[42:01.00]security doing it in
[42:03.00]compliance doing it in legal
[42:05.00]all of these things that like
[42:07.00]people nobody
[42:09.00]just wakes up and sayoh I
[42:11.00]cannot wait to go on auto gpd and ask
[42:13.00]it to do a compliance review of my thing
[42:15.00]you know just not what inspires people
[42:17.00]so think the gap on the developer
[42:19.00]side has been the more bottom sub
[42:21.00]hacker mentality is trying to build
[42:23.00]this like very generic
[42:25.00]agents that can do a lot of open
[42:27.00]ended task and then the more business
[42:29.00]side of things is like hey if I want
[42:31.00]to raise my next round I cannot
[42:33.00]just like sit around the mess
[42:35.00]mess around with like super generic
[42:37.00]stuff I need to find a use case that
[42:39.00]really works and I think that that
[42:41.00]is worth for a lot of folks in
[42:43.00]parallel you have a lot of companies
[42:45.00]doing evals there are dozens
[42:47.00]of them that just want to help you
[42:49.00]measure how good your models are
[42:51.00]doing again if you build evals
[42:53.00]you need to also have a restrained
[42:55.00]surface area to actually figure out
[42:57.00]whether or not it's good right because
[42:59.00]there's a lot of stuff going on
[43:01.00]and everything under the sun so
[43:03.00]that's another category where I've
[43:05.00]seen from the sort of
[43:07.00]pitches that I've seen there's a
[43:09.00]lot of interest in the enterprise
[43:11.00]it's just like really fragmented
[43:13.00]because the production use cases
[43:15.00]are just coming like now you know
[43:17.00]there are not a lot of long established
[43:19.00]ones to test against and so
[43:21.00]that's kind of on the virtual agents
[43:23.00]and then the robotic side
[43:25.00]it's probably been the thing that
[43:27.00]the amount of robots that were there
[43:29.00]there were just like robots everywhere
[43:31.00]both in the keynote and then on the show floor
[43:33.00]you would haveboston dynamics
[43:35.00]dogs running around
[43:37.00]there was like this like fox
[43:39.00]robot that had like a virtual face
[43:41.00]that like talked to you and like moved
[43:43.00]in realtime there were industrial
[43:45.00]robots.imbedia did a big push
[43:47.00]on their own omniverse thing which
[43:49.00]is like thisdigital twin
[43:51.00]ofwhatever environments you're in
[43:53.00]that you can use to train the robots agents
[43:55.00]so that kind of takes people back to the
[43:57.00]reinforcement learning days but
[43:59.00]yeah agents people want them
[44:01.00]you know people want them I give a talk
[44:03.00]about the rise of the full stack employees
[44:05.00]and kind of this future the same way
[44:07.00]full stack engineers kind of work
[44:09.00]agross the stack in the future every
[44:11.00]employee is going to interact with
[44:13.00]every part of the organization through
[44:15.00]agents and AI enabled tooling
[44:17.00]this is happening it just needs to be a
[44:19.00]lot more narrow than maybe the first
[44:21.00]approach that we took which is just
[44:23.00]short of super interesting stuff going on
[44:25.00]yeah but he less you covered
[44:27.00]a lot of stuff there I'll separate
[44:29.00]the robotics piece because I feel like
[44:31.00]that's so different from the software world
[44:33.00]but yeah we do we do talk to a lot of
[44:35.00]engineers and you know that that this is
[44:37.00]our sort of bread and butter and I do agree
[44:39.00]that vertical agents have worked out a
[44:41.00]lot better than the horizontal ones
[44:43.00]I think you know the point I'll make
[44:45.00]here is just the reason auto GPT
[44:47.00]and maybe AGI you know it's in the
[44:49.00]name like they were promising AGI
[44:51.00]but I think people are discovering that you cannot
[44:53.00]engineer your way to AGI it has to be
[44:55.00]done at the model level and all these
[44:57.00]engineer and prompt engineering
[44:59.00]hacks on top of it weren't really going to
[45:01.00]get us there in a meaningful way
[45:03.00]without much further
[45:05.00]improvements in the models
[45:07.00]I would say I'll go so far as to say
[45:09.00]even Devin which is I would
[45:11.00]I think the most advanced agents
[45:13.00]that we've ever seen still requires a
[45:15.00]lot of engineering and still probably
[45:17.00]falls apart a lot in terms of like
[45:19.00]practical usage or it's just way too slow
[45:21.00]and expensive for you know what it's
[45:23.00]what is promised in comparison to video
[45:25.00]so yeah that's what happened
[45:27.00]of agents from last year
[45:29.00]but I do see like vertical agents
[45:31.00]being very popular and sometimes
[45:33.00]I think the word agent might even be
[45:35.00]overused sometimes like people don't
[45:37.00]really care whether or not you call it an
[45:39.00]AI agent right like does it replace
[45:41.00]boring menial tasks that I do
[45:43.00]that I might hire human to do or
[45:45.00]that the human who is hired to do it
[45:47.00]doesn't really want to do
[45:49.00]and I think there's absolutely ways
[45:51.00]in sort of a vertical context
[45:53.00]that you can actually go after
[45:55.00]very routine tasks that can be scaled out
[45:57.00]to a lot of you know AI assistance
[45:59.00]so yeah I would
[46:01.00]basically plus one what I said there
[46:03.00]I think it's very very promising
[46:05.00]and I think more people should work on it
[46:07.00]not less like there's not enough people
[46:09.00]like this should be the main thrust
[46:11.00]of the AI engineers to look out
[46:13.00]look for use cases and go to production
[46:15.00]instead of just always working on some
[46:17.00]AI promising thing that never arrives
[46:19.00]I can only add that
[46:21.00]so I've been fiercely making
[46:23.00]tutorials behind the scenes
[46:25.00]around basically everything you can imagine
[46:27.00]with AI we've probably done about 300
[46:29.00]tutorials over the last couple months
[46:31.00]and the verticalized
[46:33.00]anything right like this is a solution
[46:35.00]for your particular job
[46:37.00]or role even if it's way less
[46:39.00]interesting or
[46:41.00]kind of sexy it's like so radically more
[46:43.00]useful to people in terms of intersecting
[46:45.00]with how like those are the ways that people are
[46:47.00]actually adopting AI
[46:49.00]in a lot of cases it's just a
[46:51.00]thing that I do over and over again
[46:53.00]by the way I think that's the same way that even the
[46:55.00]generalized models are getting adopted you know
[46:57.00]it's like I use mid-journey
[46:59.00]for lots of stuff but the main thing I use it for
[47:01.00]is youtube thumbnails every day like day in
[47:03.00]day out I will always do a youtube thumbnail
[47:05.00]you know or two with with mid-journey
[47:07.00]and it's like you can you can start to extrapolate
[47:09.00]that across a lot of things and all of a sudden
[47:11.00]you know AI doesn't
[47:13.00]it looks revolutionary because of
[47:15.00]a million small changes rather than
[47:17.00]one sort of big dramatic change
[47:19.00]and I think that the verticalization of agents
[47:21.00]is sort of a great example of
[47:23.00]how that's going to play out too
[47:25.00]So I'll have one caveat here
[47:27.00]which is I think that because
[47:29.00]multimodal models are now commonplace
[47:31.00]like claw,generalized,openEI
[47:33.00]all very very easily
[47:35.00]multimodal,apples easily
[47:37.00]multimodal,all this stuff
[47:39.00]the switch for agents for sort of general desktop
[47:41.00]browsing
[47:43.00]that I think people need to keep an eye on
[47:45.00]it's not mature yet
[47:47.00]but it is absolutely coming on the way
[47:49.00]and so just as we're starting to talk
[47:51.00]about this verticalization piece
[47:53.00]because that is mature
[47:55.00]that is ready for people to work on
[47:57.00]that is that a lot of people are making really good money doing that
[47:59.00]the thing that's on the rise
[48:01.00]is this sort of drive by vision
[48:03.00]version of the agent where
[48:05.00]they're not specifically taking in text or anything
[48:07.00]but just watching your screen just like someone else would
[48:09.00]and piloting it
[48:11.00]by vision
[48:13.00]in the episode with David
[48:15.00]that will have dropped by the time that this airs
[48:17.00]I think that is the promise of adept
[48:19.00]that is the promise of what
[48:21.00]a lot of these sort of desktop agents are
[48:23.00]and that is the more general purpose
[48:25.00]system that could be as big as
[48:27.00]the browser
[48:29.00]the operating system
[48:31.00]people really want to build that
[48:33.00]foundational piece of software in AI
[48:35.00]and I would see the potential
[48:37.00]therefor desktop agents being that
[48:39.00]that you can have self-driving
[48:41.00]computers. Don't write the horizontal
[48:43.00]piece out. I just think we took a while
[48:45.00]to get there. What else are you guys
[48:47.00]seeing?That's interesting to you.
[48:49.00]I'm looking at your notes and seeing a ton
[48:51.00]ofcategories. I'll take
[48:53.00]the next two as one category
[48:55.00]which is basically alternative architectures.
[48:57.00]The two main things that everyone
[48:59.00]following AI kind of knows now is
[49:01.00]one, the diffusion architecture
[49:03.00]and two, the
[49:05.00]let's just say the decoder only
[49:07.00]transformer architecture that is popular
[49:09.00]by GPT. You can read, you can look on
[49:11.00]YouTube for thousands and thousands of tutorials
[49:13.00]on each of those things. What we are talking about
[49:15.00]here is what's next, what people are
[49:17.00]researching and what could be on the horizon
[49:19.00]that takes the place of those other two things.
[49:21.00]So first we'll talk about transformer architectures
[49:23.00]and then diffusion. So transform the two leading
[49:25.00]candidates are effectively RWKV
[49:27.00]and the state space models. The most recent
[49:29.00]one of which is Mamba but there's others
[49:31.00]and the S4H3
[49:33.00]stuff coming out of Hazy Research in Stanford
[49:35.00]and all of those are
[49:37.00]non-quadratic
[49:39.00]language models that scale
[49:41.00]that promise to scale a lot better
[49:43.00]than the traditional transformer
[49:45.00]that this might be too theoretical
[49:47.00]for most people right now
[49:49.00]but it's going to be
[49:51.00]it's going to come out in weird ways
[49:53.00]where imagine if like right now
[49:55.00]the talk of the town is that
[49:57.00]Claude and Gemini have a million tokens of context
[49:59.00]and you can put in like
[50:01.00]two hours of video now
[50:03.00]but what if you could throw in
[50:05.00]200,000 hours of video
[50:07.00]how does that change your
[50:09.00]usage of AI
[50:11.00]what if you could throw in the entire
[50:13.00]genetic sequence of a human
[50:15.00]and synthesize new drugs
[50:17.00]how does that change things
[50:19.00]we don't know because we haven't had access to this capability
[50:21.00]being so cheap before
[50:23.00]and that's the ultimate promise of these two models
[50:25.00]they're not there yet
[50:27.00]it's a very very good progress
[50:29.00]RWKV and Mamba are probably the two leading examples
[50:31.00]both of which are open source
[50:33.00]that you can try them today
[50:35.00]and have a lot of progress there
[50:37.00]the main thing I'll highlight for other UKV
[50:39.00]is that at the 7B level
[50:41.00]they seem to have beat
[50:43.00]Lama2 in all benchmarks
[50:45.00]that matter
[50:47.00]at the same size for the same amount of training
[50:49.00]as an open source model so that's exciting
[50:51.00]there are 7B now
[50:53.00]they're not at 7TB we don't know if it'll scale
[50:55.00]the other thing is diffusion
[50:57.00]diffusions and transformers
[50:59.00]are kind of on the collision course
[51:01.00]the original stable diffusion already used
[51:03.00]transformers in parts of its architecture
[51:05.00]it seems that transformers are
[51:07.00]eating more and more of those layers
[51:09.00]particularly the VAE layer
[51:11.00]so that's the diffusion transformer
[51:13.00]is what Sora is built on
[51:15.00]the guy who wrote the diffusion transformer
[51:17.00]paper
[51:19.00]bilt pebbles is the lead tech guy
[51:21.00]on Sora
[51:23.00]but there's more sort of experimentation
[51:25.00]to diffusion I'm holding a meetup
[51:27.00]actually here in San Francisco that's going to be like the state of diffusion
[51:29.00]which I'm pretty excited about
[51:31.00]stability's doing a lot of good work
[51:33.00]and if you look at the architecture
[51:35.00]of how they're creating
[51:37.00]stable diffusion 3, our glass diffusion
[51:39.00]and the inconsistency models
[51:41.00]or SDXL turbo
[51:43.00]all of these are like very very interesting innovations
[51:45.00]on like the original idea
[51:47.00]of what stable diffusion was so if you think
[51:49.00]that it is expensive to create or slow to create
[51:51.00]stable diffusion or AI-generated art
[51:53.00]you are not up to date with the latest models
[51:55.00]if you think it is hard to create
[51:57.00]texted images you are not up to date with the latest models
[51:59.00]and people still are kind of far behind
[52:01.00]the last piece of which
[52:03.00]is the wall cut I always kind of hold out
[52:05.00]which is text diffusion
[52:07.00]so instead of using auto generative
[52:09.00]or auto regressive transformers
[52:11.00]can you use text to diffuse
[52:13.00]so you can use diffusion models to diffuse
[52:15.00]and create entire chunks of text
[52:17.00]all at once instead of token by token
[52:19.00]and that is something that mid-journey confirmed today
[52:21.00]because it was only rumored
[52:23.00]the past few months but they confirmed today
[52:25.00]that they were looking into
[52:27.00]all those things are like very exciting new model
[52:29.00]architectures that are maybe something
[52:31.00]that you will see in production 2-3 years from now
[52:33.00]so the couple of the trends that I want to just
[52:37.00]get your takes on because they're sort of something
[52:39.00]that seems like they're coming up are
[52:41.00]one sort of these wearable
[52:43.00]kind of passive
[52:45.00]AI experiences where
[52:47.00]they're absorbing a lot of what's going on around you
[52:49.00]and then kind of bringing things back
[52:51.00]and then the other one that I
[52:53.00]wanted to see if you guys had thoughts on were
[52:55.00]sort of this next generation of chip companies
[52:57.00]obviously there's a huge amount of emphasis
[52:59.00]on hardware and silicon
[53:01.00]and different ways of doing things but
[53:03.00]love your take on neither or both of those
[53:05.00]so wearables
[53:07.00]I'm very excited about it
[53:09.00]I want wearables on me at all times
[53:11.00]I have two right here to quantify my health
[53:13.00]and I'm all for them
[53:15.00]but society is not ready for wearables
[53:17.00]no one's comfortable with
[53:19.00]a device on recording every single
[53:21.00]conversation we have even
[53:23.00]all three of us here as
[53:25.00]podcasters we don't record everything
[53:27.00]that we say and I think
[53:29.00]there's a social shift that needs to happen
[53:31.00]I'm an investor in tab
[53:33.00]they are renaming to a broader
[53:35.00]vision but they are one of the
[53:37.00]three or four leading wearables in this
[53:39.00]space instead of the AI pendants
[53:41.00]or AI OS
[53:43.00]I have seen two humans
[53:45.00]in a while in San Francisco
[53:47.00]I'm very very excited to report
[53:49.00]that there are people walking around
[53:51.00]with those things on their chest
[53:53.00]and it is as goofy as it sounds
[53:55.00]it absolutely is going to fail
[53:57.00]but god bless them for trying
[53:59.00]and I've also bought a rabbit
[54:01.00]so I'm very excited for all those things to arrive
[54:03.00]but yeah people
[54:05.00]are very keen on hardware
[54:07.00]I think the idea that you can have physical objects
[54:09.00]that embody an AI
[54:11.00]do specific things for you
[54:13.00]is as old as
[54:15.00]the sort of golem
[54:17.00]in sort of medieval times
[54:19.00]in terms of like how much we want
[54:21.00]our objects to be smart
[54:23.00]and do things for us
[54:25.00]and I think it's absolutely
[54:27.00]a great play
[54:29.00]the funny thing is people are much more willing
[54:31.00]to pay you up front
[54:33.00]for a hardware device
[54:35.00]than they are willing to pay $8 a month
[54:37.00]suscription recurring for software
[54:39.00]and so the interesting economics
[54:41.00]of these wearable companies
[54:43.00]is they have negative float
[54:45.00]in the sense that people pay deposits up front
[54:47.00]like I paid
[54:49.00]$200 for the rabbit up front
[54:51.00]and I don't get it for another six months
[54:53.00]I paid $600 for the tablet
[54:55.00]I don't get it for another six months
[54:57.00]and then they can take that money
[54:59.00]and sort of invest it in their next
[55:01.00]events or their next properties
[55:03.00]or ventures
[55:05.00]and I think that's a very interesting
[55:07.00]comics from other types of
[55:09.00]AI companies that I see
[55:11.00]and I think just the tactile feel
[55:13.00]of an AI I think is very promising
[55:15.00]I don't know if you have other
[55:17.00]thoughts on the wearable stuff
[55:19.00]open interpreter
[55:21.00]just announced their product four hours ago
[55:23.00]which is not really a wearable
[55:25.00]but it's still like a
[55:27.00]physical device
[55:29.00]it's a push to talk mic
[55:31.00]to a device on your laptop
[55:33.00]it's a $99 push
[55:35.00]but again
[55:37.00]go back to your point
[55:39.00]people are interested in
[55:41.00]spending money for things that they can hold
[55:43.00]I don't know what that means overall
[55:45.00]where things are going but
[55:47.00]making more of this AI
[55:49.00]be a physical part of your life
[55:51.00]I think people are interested in that
[55:53.00]but I agree with Sean
[55:55.00]I talked to Avi about this
[55:57.00]but Avi's point is like
[55:59.00]most consumers care about utility
[56:01.00]more than they care about privacy
[56:03.00]I'm not like you've seen with social media
[56:05.00]but I also think there's a big
[56:07.00]social reaction
[56:09.00]to AI that is like much more
[56:11.00]rooted than the social media one
[56:13.00]but we'll see
[56:15.00]but a lot again a lot of work a lot of developers
[56:17.00]a lot of money going into it so
[56:19.00]there's bound to be experiments
[56:21.00]being run on the chip
[56:23.00]sorry I'll just
[56:25.00]ship it one more thing and then we transition to the chips
[56:27.00]the thing I'll caution people on is
[56:29.00]don't overly focus on the form factor
[56:31.00]the form factor is a delivery mode
[56:33.00]there will be many form factors
[56:35.00]it doesn't matter so much as
[56:37.00]where in the data war does it sit
[56:39.00]it actually is context acquisition
[56:41.00]because and maybe a little bit of multi-modality
[56:43.00]context is king
[56:45.00]if you have access to data
[56:47.00]that no one else has
[56:49.00]then you will be able to create AI that no one else can create
[56:51.00]and so what is the most personal context
[56:53.00]it is your everyday conversation
[56:55.00]it is as close to mapping your
[56:57.00]mental train of thought as possible
[56:59.00]without physically you writing down notes
[57:01.00]so that is the promise
[57:03.00]the ultimate goal here
[57:05.00]which is like personal context
[57:07.00]it's always available on you
[57:09.00]a little ncl that stuff
[57:11.00]that's the frame I want to give people
[57:13.00]that the form factors will change
[57:15.00]and there will be multiple form factors
[57:17.00]but it's the software behind that
[57:19.00]in the personal context
[57:21.00]that you cannot get anywhere else
[57:23.00]that'll win
[57:25.00]so that was wearables
[57:27.00]but it's not even a new release
[57:29.00]because the company I think was started in 2016
[57:31.00]so it's actually quite old
[57:33.00]but now recently captured
[57:35.00]people's imagination with their mixrall
[57:37.00]500 tokens a second demo
[57:39.00]yeah I think so far
[57:41.00]the battle on the GPU side
[57:43.00]has beeneither you go
[57:45.00]kind of like massive chip
[57:47.00]like this irreversible of the world
[57:49.00]where one chip front
[57:51.00]reversible is about 2 million dollars
[57:53.00]you know that's compared
[57:55.00]so you cannot compare one chip
[57:57.00]versus one chip but h100
[57:59.00]it's like 40,000
[58:01.00]the problem with those architectures
[58:03.00]has been
[58:05.00]they want to be very general
[58:07.00]but they wanted to put a lot
[58:09.00]of the SRAM on the chip
[58:11.00]it's much more convenient
[58:13.00]when you're using larger language models
[58:15.00]but the models outpace the size
[58:17.00]of the chips and chips have a much longer
[58:19.00]turnaround cycle
[58:21.00]Grock today is great for the current architecture
[58:23.00]it's a lot more expensive also
[58:25.00]as far as dollar per flop
[58:27.00]but there are the SIK
[58:29.00]when you have very high concurrency
[58:31.00]we actually were much cheaper
[58:33.00]you shouldn't just be looking at the compute power
[58:35.00]for most people this doesn't really matter
[58:37.00]you know like I think that's like the most
[58:39.00]the most interesting thing to me is like
[58:41.00]we've now gone back
[58:43.00]with AI to a world where
[58:45.00]developers care
[58:47.00]about what hardware is running
[58:49.00]which was not the case in traditional software
[58:51.00]maybe 20 years since
[58:53.00]as the cloud has gotten really big
[58:55.00]my thinking is that
[58:57.00]in the next 2-3 years
[58:59.00]we're going to go back to that
[59:01.00]we're like people are not going to be sweating
[59:03.00]what GPU do you have in your cloud
[59:05.00]what do you have
[59:07.00]you want to run this model
[59:09.00]we can run it at the same speed as everybody else
[59:11.00]and then everybody will make different choices
[59:13.00]whether they want to have
[59:15.00]higher front end capital investment
[59:17.00]and then better utilization
[59:19.00]and then upgrade later
[59:21.00]there are a lot of parameters
[59:23.00]and then there's the dark horses
[59:25.00]that is some of the smaller companies
[59:27.00]like Lemurian Labs, MedEx
[59:29.00]that are working on
[59:31.00]maybe not a chip alone
[59:33.00]but also some of the actual math
[59:35.00]infrastructure and the instructions
[59:37.00]that make them run
[59:39.00]there's a lot going on but
[59:41.00]yeah I think the
[59:43.00]the episode with Dylan will be
[59:45.00]interesting for people but
[59:47.00]hey everybody has pros and cons
[59:49.00]it's different than the models
[59:51.00]where you're like oh this one is definitely better for me
[59:53.00]and I'm going to use it
[59:55.00]I think for most people
[59:57.00]it's like fun twitter meming
[59:59.00]99% of people
[60:01.00]that tweet about this stuff
[60:03.00]are never going to buy any of these chips anyway
[60:05.00]so it's really more for entertainment
[60:07.00]wow I mean
[60:09.00]this is serious business here
[60:11.00]the potential new Nvidia
[60:13.00]if anyone can take
[60:15.00]i'm more talking about
[60:17.00]how should people think about it
[60:19.00]i think the end user
[60:21.00]is not impacted as much
[60:23.00]so I disagree
[60:25.00]I love disagreements
[60:27.00]who likes the podcast
[60:29.00]where all 3 people always agree with each other
[60:31.00]you will see the impact of this
[60:33.00]in the tokens per second over time
[60:35.00]this year
[60:37.00]I have very very credible sources
[60:39.00]all telling me
[60:41.00]that the average tokens per second
[60:43.00]right now we have
[60:45.00]somewhere between 50 to 100
[60:47.00]that's the norm for people
[60:49.00]average tokens per second
[60:51.00]will go to 500 to 2000
[60:53.00]this year
[60:55.00]from a number of chip suppliers that I cannot name
[60:57.00]that will cause
[60:59.00]step change in the use cases
[61:01.00]every time you have an order of magnitude improvement
[61:03.00]in the speed of something
[61:05.00]you unlock new use cases
[61:07.00]that become fun instead of a chore
[61:09.00]and so that's what I would
[61:11.00]caution this audience to think about
[61:13.00]what can you do in much higher AI speed
[61:15.00]it's not just things streaming out faster
[61:17.00]it is
[61:19.00]things working in the background a lot more seamlessly
[61:21.00]and therefore being a lot more useful
[61:23.00]than previously imagined
[61:25.00]so that would be my two cents on that
[61:27.00]yeah
[61:29.00]I mean the new
[61:31.00]imbedia chips are also much faster
[61:33.00]to me that's true
[61:35.00]when it comes about startups
[61:37.00]are the startups pushing
[61:39.00]on the incumbents
[61:41.00]or are the incumbents still leading
[61:43.00]and then the startups are riding the same wave
[61:45.00]I don't have yet a good sense of that
[61:47.00]it's next year's
[61:49.00]imbedia release just gonna be better than everything
[61:51.00]that gets released this year
[61:53.00]if that's the case
[61:55.00]it's like okay
[61:57.00]damn jensen
[61:59.00]it's like I'm gonna fight imbedia
[62:01.00]damn jensen got hands
[62:03.00]he really does
[62:05.00]well
[62:07.00]one conversation guys
[62:09.00]just by way of wrapping up
[62:11.00]call it over the next three months
[62:13.00]between now and sort of the beginning of summer
[62:15.00]what's one prediction that each of you has
[62:17.00]could be about anything
[62:19.00]could be big company, could be startup
[62:21.00]could be something you have privileged information that you know
[62:23.00]and you just won't tell us that you actually know
[62:25.00]does it have to be something that we think
[62:27.00]it's gonna be true or like something that we think
[62:29.00]because for me it's like
[62:31.00]is sundar gonna be the CEO of google
[62:33.00]maybe not in three months but maybe in like six months
[62:35.00]in nine months
[62:37.00]people were like oh maybe that miss is gonna be the new CEO
[62:39.00]that was kinda like
[62:41.00]I was busyfishing some deep mind people
[62:43.00]and google people for like a good
[62:45.00]guest for the pot and I was like
[62:47.00]oh what about jeptine and they're like
[62:49.00]well that miss is really like the person that runs everything
[62:51.00]anyway and the stuff
[62:53.00]and it's like interesting
[62:55.00]so I don't know
[62:57.00]serge could come back I don't know
[62:59.00]like he's making more appearances these days
[63:01.00]yeah
[63:03.00]but we can just put it as like
[63:05.00]my thing is like
[63:07.00]CEO change potential
[63:09.00]but again
[63:11.00]three months is too short
[63:13.00]to make a prediction
[63:15.00]time scale might be off
[63:17.00]yeah I mean
[63:21.00]for me I think the progression
[63:23.00]in vertical agent companies
[63:25.00]will keep going
[63:27.00]we just had the other day
[63:29.00]Klarna talking about how they replaced like
[63:31.00]customer support agents with the
[63:33.00]AI agents
[63:35.00]that's to the beginning guys
[63:37.00]imagine this rolling out across most of the fortune
[63:39.00]500
[63:41.00]and I'm not saying this is like a utopian
[63:43.00]scenario there will be very very
[63:45.00]embarrassing and bad outcomes
[63:47.00]of this where like humans would never
[63:49.00]make this mistake but AIS did and
[63:51.00]we'll all laugh at it or it will be very offended
[63:53.00]by whatever you know bad outcome
[63:55.00]it did so we have to be responsible
[63:57.00]and careful in the rollout but yeah this is
[63:59.00]the rolling out you know Alessio likes to say
[64:01.00]that this year is the year of AI production
[64:03.00]let's see it let's see all these vertical
[64:05.00]full stack employees
[64:07.00]come out into the workforce
[64:09.00]love it alright guys well thank you so much
[64:11.00]for sharing your thoughts and insights
[64:13.00]here and can't wait to do it again
[64:15.00]welcome back again
[64:17.00]it's Charlie your AI cohost
[64:19.00]we're now in part two
[64:21.00]of the special weekend episode
[64:23.00]collating some of SWIX and Alessio's
[64:25.00]recent appearances
[64:27.00]if you are not active in the latent space discord
[64:29.00]you might not be aware of the
[64:31.00]many many many in person
[64:33.00]events we host gathering our
[64:35.00]listener community all over the world
[64:37.00]you can see the latent space
[64:39.00]community page for how to join
[64:41.00]and subscribe to our event calendar
[64:43.00]for future meetups
[64:45.00]we're going to share some of our recent live
[64:47.00]appearances in this next part
[64:49.00]starting with the Thursday nights in AI meetup
[64:51.00]a regular fixture in the SF AI scene
[64:53.00]run by imbu and outset capital
[64:55.00]primarily our former guest
[64:57.00]Kanjen Kyu
[64:59.00]Allie Roder and Josh Albrecht
[65:01.00]here's SWIX
[65:03.00]today for those of you who have been here before
[65:07.00]you know the general format
[65:09.00]so we'll do a quick fireside Q&A with SWIX
[65:11.00]where we're asking him the questions
[65:13.00]then we'll actually go to our rapid fire Q&A
[65:15.00]where we're asking really fast
[65:17.00]hopefully spicy questions
[65:19.00]and then we'll open it up to the audience
[65:21.00]for your questions so you guys
[65:23.00]around the room submit your questions
[65:25.00]and we'll go through as many of them as possible
[65:27.00]during that period
[65:29.00]and then actually SWIX brought
[65:31.00]a gift for us which is two
[65:33.00]latent space t-shirts
[65:35.00]AI engineer t-shirts
[65:37.00]and those will be awarded to the
[65:39.00]two spiciest questions
[65:41.00]askers
[65:43.00]and I'll let Josh decide on that
[65:45.00]so we want to get your spiciest takes
[65:47.00]please send them in during the event
[65:49.00]as we're talking and then also at the end
[65:51.00]alright with that
[65:53.00]let's get going
[65:55.00]ok
[65:57.00]welcome SWIX
[65:59.00]how does it feel to be interviewed
[66:01.00]rather than the interviewer
[66:03.00]weird I don't know what to do in this chair
[66:05.00]where should I put my hands
[66:07.00]you look good
[66:09.00]and I also love asking follow up questions
[66:11.00]and I tend to take over panels a lot
[66:13.00]if you ever see me on a panel
[66:15.00]I tend to ask the other panelists questions
[66:17.00]so we should be ready
[66:19.00]this is like a free interview
[66:21.00]so why not
[66:23.00]so you interviewed Ken June
[66:25.00]the CEO of Impeo4 but you didn't interview Josh
[66:27.00]so maybe tonight
[66:29.00]we will look for different questions
[66:31.00]and look for alignment
[66:33.00]I love it
[66:35.00]I just want to hear this story
[66:37.00]you've completely exploded with latent space
[66:39.00]and AI engineer and I know
[66:41.00]you also before all of that
[66:43.00]had exploded in popularity for your
[66:45.00]learning in public movement and your dev tools work
[66:47.00]dav relations work
[66:49.00]so who are you and how did you get here
[66:51.00]let's start with that
[66:53.00]quick story is I'm Sean I'm from Singapore
[66:55.00]SWIX is my initials for those who don't know
[66:57.00]a lot of Singaporeans are ethically Chinese
[66:59.00]and we have Chinese names and English names
[67:01.00]so it's just my initials
[67:03.00]came to the US for college
[67:05.00]and have been here for about 15 years
[67:07.00]but half of that was in finance
[67:09.00]and then the other half was in tech
[67:11.00]tech is where I was most known
[67:13.00]just because I realized that
[67:15.00]I was much more
[67:17.00]aligned towards learning in public
[67:19.00]whereas in finance everything is a trade secret
[67:21.00]everything is zero sum
[67:23.00]whereas in tech you're allowed to come
[67:25.00]to meetups and conferences and share your
[67:27.00]learnings and share your mistakes even
[67:29.00]and that's totally fine you like
[67:31.00]open source your code it's totally fine
[67:33.00]and even better you like contribute
[67:35.00]pr to other people's code which is even better
[67:37.00]and I found that I thrives in that
[67:39.00]learning public environments and
[67:41.00]that kind of got me started
[67:43.00]early higher
[67:45.00]early directly higher at Netlify
[67:47.00]and then did the same at EWS
[67:49.00]Temporal and Airbyte
[67:51.00]and so that's like the whole story
[67:53.00]I can talk more about like developer tooling
[67:55.00]and developer relations if that's something
[67:57.00]that people are interested in
[67:59.00]but I think the more recent thing is AI
[68:01.00]and I started really being interested
[68:03.00]in it mostly because
[68:05.00]the approximate cause of starting learning
[68:07.00]in space was stable diffusion
[68:09.00]when you could run a large model
[68:11.00]on your desktop
[68:13.00]where I was like okay this is
[68:15.00]something qualitatively very different
[68:17.00]and then we started
[68:19.00]in space and
[68:21.00]we have to talk about it on a podcast
[68:23.00]there we go
[68:25.00]it wasn't a podcast for like four months
[68:27.00]and then I had been running a discord
[68:29.00]for DevTools investors
[68:31.00]I also invested in DevTools
[68:33.00]and I advised companies on DevTools
[68:35.00]definition things
[68:37.00]and I think it was the start of 2023
[68:39.00]Alessio and I were both like
[68:41.00]I think we need to get more tokens out of
[68:43.00]people and I was running out of original sources
[68:45.00]to write about
[68:47.00]so I was like okay I'll go get those original sources
[68:49.00]and I think that's when we started the podcast
[68:51.00]and I think it's just the chemistry between us
[68:53.00]the way we spike in different ways
[68:55.00]and also honestly the
[68:57.00]kind participation of the guests
[68:59.00]to give us their time
[69:01.00]getting George Hoss was a big deal
[69:03.00]and also shoutout to Alessio for cold emailing him
[69:05.00]for booking some of our
[69:07.00]biggest guests
[69:09.00]and just working really hard to try to tell the story
[69:11.00]that people can use at work
[69:13.00]I think that there's a lot of AI podcasts out there
[69:15.00]and a lot of AI forums
[69:17.00]or fireside chats with no fire
[69:19.00]that always talk about
[69:21.00]what's your AGI timeline, what's your PDoom
[69:23.00]very very nice hallway conversations
[69:25.00]for freshman year but not very useful
[69:27.00]for work
[69:29.00]and practically making money
[69:31.00]and thinking about
[69:33.00]changing everyday lives
[69:35.00]what's interesting is obviously
[69:37.00]you care about the existential
[69:39.00]safety of the human race
[69:41.00]but in the meantime we got to eat
[69:43.00]so I think that's kind of
[69:45.00]the inspaces niche
[69:47.00]we explicitly don't really talk about AGI
[69:49.00]we explicitly don't talk about
[69:51.00]things that we're a little bit too far out
[69:53.00]we don't do a ton of robotics
[69:55.00]we don't do a ton of high frequency trading
[69:57.00]there's tons of machine learning in there
[69:59.00]but we just don't do that
[70:01.00]we're like what are most software engineers going to need
[70:03.00]because that's our background
[70:05.00]and that's the audience that we serve
[70:07.00]and I think just being really clear on that audience
[70:09.00]has resonated with people
[70:11.00]you would never expect a technical podcast
[70:13.00]to reach a general audience
[70:15.00]like top 10 on the tech charts
[70:17.00]but I've been surprised by that before
[70:19.00]and it's been successful
[70:21.00]I don't know what to say about that
[70:23.00]I think honestly I kind of
[70:25.00]have this negative reaction towards being
[70:27.00] being classified as a podcast
[70:29.00]because the podcast is downstream of ideas
[70:31.00]and it's one mode of conversation
[70:33.00]it's one mode of idea delivery
[70:35.00]but you can deliver ideas
[70:37.00]on a newsletter in a person like this
[70:39.00]there's so many different ways
[70:41.00]and so I think I think about it more as
[70:43.00]we are trying to start or serve an industry
[70:45.00]and that industry is the AI
[70:47.00]engineer industry
[70:49.00]which we can talk about more
[70:51.00]yes let's go into that
[70:53.00]so the AI engineer you penned a piece
[70:55.00]all the rise of the AI engineer
[70:57.00]you tweeted about it
[70:59.00]you also responded
[71:01.00]largely agreeing with what you said
[71:03.00]what is an AI engineer
[71:05.00]the AI engineer is the software engineer
[71:07.00]building with AI
[71:09.00]enhanced by AI
[71:11.00]and eventually it will be a non-human
[71:13.00]engineers writing code for you
[71:15.00]which I know impu is all about
[71:17.00]you're saying eventually the AI engineer
[71:19.00]will become a non-human engineer
[71:21.00]that will be one kind of AI engineer
[71:23.00]that people are trying to build
[71:25.00]and is probably the most furthest away
[71:27.00]because it's so hard
[71:29.00]but there are three types of AI engineer
[71:31.00]one is AI enhanced
[71:33.00]where you use AI products like co-pilot
[71:35.00]and two is AI products engineer
[71:37.00]where you use the exposed AI capabilities
[71:39.00]to the end user
[71:41.00]as a software engineer like not doing pre-training
[71:43.00]not being an ML researcher
[71:45.00]not being an ML engineer
[71:47.00]but just interacting with foundation models
[71:49.00]probably APIs from foundation model labs
[71:51.00]what's the third one
[71:53.00]and the third one is the non-human AI engineer
[71:55.00]the fully autonomous
[71:57.00]dream coder
[71:59.00]how long do you think it is till we get to
[72:01.00]early
[72:03.00]this is my equivalent of AGI timelines
[72:05.00]I know I know
[72:07.00]lots of active
[72:09.00]I have supported companies
[72:11.00]actively working on that
[72:13.00]I think it's more useful to think about levels of autonomy
[72:15.00]and so my answer to that
[72:17.00]is perpetually five years away
[72:19.00]until it figures it out
[72:21.00]no but my actual antidote
[72:23.00]the closest comparison we have to that is self-driving
[72:25.00]we're doing this in San Francisco
[72:27.00]for those who are watching on the live stream
[72:29.00]if you haven't come to San Francisco
[72:31.00]and seen and taken a waymo ride
[72:33.00]just come get a friend and take a waymo ride
[72:35.00]I remember 2014
[72:37.00]we covered a little bit of autos in my hedge fund
[72:39.00]and I remember telling a friend
[72:41.00]I was self-driving cars around the corner
[72:43.00]this is it
[72:45.00]parking will be a thing of the past
[72:47.00]and it didn't happen for the next ten years
[72:49.00]but now most of us in San Francisco
[72:51.00]can take it for granted
[72:53.00]I think you just have to
[72:55.00]be mindful that
[72:57.00]the rough edges take a long time
[72:59.00]and yes it's going to work in demos
[73:01.00]then it's going to work a little bit further out
[73:03.00]and it's just going to take a long time
[73:05.00]the more useful mental model I have
[73:07.00]is levels of autonomy
[73:09.00]in self-driving you have level one, two, three, four, five
[73:11.00]just the amount of human attention
[73:13.00]that you getat first
[73:15.00]your hands are always on ten and two
[73:17.00]and you have to pay attention to the driving
[73:19.00]every thirty seconds
[73:21.00]and eventually you can sleep in the car
[73:23.00]there's a whole spectrum of that
[73:25.00]what's the equivalent for coding
[73:27.00]keep your hands on the keyboard
[73:29.00]and then eventually you have to accept everything
[73:31.00]that's good
[73:33.00]approve the PR
[73:35.00]approve this looks good
[73:37.00]that's the dream that people want
[73:39.00]because really you unlock a lot of coding
[73:41.00]when people, non-technical people can file issues
[73:43.00]and then the
[73:45.00]AI engineer can sort of automatically
[73:47.00]write code, pass your tests
[73:49.00]and if it kind of works as
[73:51.00]as advertised, then you can just kind of merge it
[73:53.00]and then you
[73:55.00]10x, 100x, the number of developers in your company
[73:57.00]immediately
[73:59.00]so that's the goal, that's the holy grail
[74:01.00]we're not there yet, but sweep, code gen
[74:03.00]there's a bunch of companies, magic probably
[74:05.00]are all working towards that
[74:07.00]and so the TLDR
[74:09.00]the thing that we covered, Leslie and I covered
[74:11.00]in the January recap
[74:13.00]that we did was that
[74:15.00]the people should have in their minds is the inner loop
[74:17.00]versus the outer loop for the developer
[74:19.00]inner loop is everything that happens
[74:21.00]in your IDE between git commits
[74:23.00]and outer loop is what happens
[74:25.00]when you push up your git commit to
[74:27.00]github, for example, or gitlab
[74:29.00]and that's a nice split, which means
[74:31.00]everything local, everything that needs to be fast
[74:33.00]everything that's kind of very hands on for developers
[74:35.00]is probably easier to automate
[74:37.00]or easier to have code assistance
[74:39.00]that's what copilot is, that's what all those things are
[74:41.00]and then everything that happens autonomously
[74:43.00]or effectively away from the keyboard
[74:45.00]with a github issue or something
[74:47.00]that is more outer loop where
[74:49.00]you're relying a lot more on autonomy
[74:51.00]and we are maybe not smart enough
[74:53.00]to do that yet
[74:55.00]Do you have any thoughts on the user experience
[74:57.00]and how that will change? One of the things that
[74:59.00]has happened for me, looking at some of these products
[75:01.00]and playing around with things ourselves
[75:03.00]it sounds good to have an automated PR
[75:05.00]then you get an automated PR and you're like
[75:07.00]I really don't want to review 300 lines
[75:09.00]of generated code and find the bug
[75:11.00]and then you have another agent that's a reviewer
[75:13.00]and then they just come up to you
[75:15.00]and then you like tell it, go fix it
[75:17.00]and it comes back with 400 lines
[75:19.00]yes, there is a length bias to code
[75:21.00]and
[75:23.00]you do have higher passing rates
[75:25.00]in PRs, this is a documented human behavior
[75:27.00]thing, send me two lines of code
[75:29.00]I will review the shit out of that
[75:31.00]I don't know if I can swear on this
[75:33.00]send me 200 lines of code, looks good to me
[75:35.00]guess what, the agents are going to
[75:37.00]perfectly happy to copy that behavior from us
[75:39.00]when we actually want them to do the opposite
[75:41.00]so, yeah, I think
[75:43.00]the GAN model of code generation
[75:45.00]is probably not going to work super well
[75:47.00]I do think we probably need just
[75:49.00]better planning from the start
[75:51.00]which is, I'm just repeating the
[75:53.00]MBU thesis, by the way
[75:55.00]just go listen to Kanjin talk about this
[75:57.00]she's much better at it than I am
[75:59.00]but yeah, I think
[76:01.00]the code review thing is going to be
[76:03.00]I think that what codium
[76:05.00]the two codiums, the Israeli one
[76:07.00]Israeli codium
[76:09.00]with the E
[76:11.00]Yeah, codium with the E
[76:13.00]They still have refused to rename
[76:15.00]I'm friends with both of them
[76:17.00]Every month I'm like
[76:19.00]Guys, lets all come to one room
[76:21.00]Someone's got a fold
[76:23.00]Codium with the E has gone
[76:25.00]You got to write the test first
[76:27.00]It's like a sort of tripartite
[76:29.00]relationship, again this is also covered on a
[76:31.00]podcast with them which is fantastic
[76:33.00]You interview me, you sort of through me
[76:35.00]So, codium is like
[76:37.00]They've already thought this all the way through
[76:39.00]They're like, okay you write the user story
[76:41.00]From the user story you generate all the tests
[76:43.00]You also generate the code
[76:45.00]And you update any one of those
[76:47.00]They all have to update together
[76:49.00]And probably the critical factor
[76:51.00]Is the test generation from the story
[76:53.00]Because everything else
[76:55.00]Can just kind of bounce the hits off
[76:57.00]Of those things until they pass
[76:59.00]So you have to write good tests
[77:01.00]It's kind of like the eat your vegetables of coding
[77:03.00]Which nobody really wants to do
[77:05.00]And so I think it's a really smart tactic
[77:07.00]To go to market
[77:09.00]By saying we automatically generate
[77:11.00]Test for you and start not great
[77:13.00]But then get better
[77:15.00]And eventually you get to
[77:17.00]The weakest point in the chain
[77:19.00]For the entire loop of code generation
[77:21.00]What do you think their weakest link is
[77:25.00]The weakest link
[77:27.00]It's test generation
[77:29.00]Do you think there's a way to
[77:31.00]To make that actually better
[77:33.00]For making it better
[77:37.00]You have to have good isolation
[77:39.00]And I think
[77:41.00]Proper, serverless cloud environment
[77:43.00]Is integral to that
[77:45.00]It could be like a fly I/O
[77:47.00]It could be like
[77:49.00]A cloud fair worker
[77:51.00]It depends how many resources
[77:53.00]Your test environment needs
[77:55.00]And effectively I was talking about this
[77:57.00]I think with maybe Rob earlier in the audience
[77:59.00]Every agent needs a sandbox
[78:01.00]If you're a code agent you need a coding sandbox
[78:03.00]But if you're whatever
[78:05.00]Like MBU used to have this
[78:07.00]Minecrafts clone that was much faster
[78:09.00]If you have a model of the real world
[78:11.00]You have to go generate some plan
[78:13.00]Or some code or some whatever
[78:15.00]Test it against that real world
[78:17.00]So that you can get this iterative feedback
[78:19.00]And then get the final result back
[78:21.00]That is somewhat validated against the real world
[78:23.00]And so you need a really good sandbox
[78:25.00]I don't think people
[78:27.00]This is an infrastructure need
[78:29.00]Humans have had for a long time
[78:31.00]We've never solved it for ourselves
[78:33.00]And now we have to solve it for about
[78:35.00]A thousand times larger quantity of agents
[78:37.00]Then actually exist
[78:39.00]And so I think we actually have to
[78:41.00]Involve a lot more infrastructure
[78:43.00]In order to serve these things
[78:45.00]So for those who don't know
[78:47.00]I also have
[78:49.00]So we're talking about the rise of AI engineer
[78:51.00]I also have various conversations
[78:53.00]About immutable infrastructure
[78:55.00]And this is all of the kinds
[78:57.00]In order to solve
[78:59.00]Agents and coding agents
[79:01.00]We're going to have to solve the other stuff too along the way
[79:03.00]And it's really neat for me
[79:05.00]To see all that tied together in my deaf tools work
[79:07.00]That all these themes kind of reemerge
[79:09.00]Just naturally just because
[79:11.00]Everything we needed for humans
[79:13.00]We just need a hundred times more for agents
[79:15.00]Let's talk about the AI engineer
[79:17.00]AI engineer has become a whole thing
[79:19.00]It's become a term and also a conference
[79:21.00]And tell us more
[79:23.00]And a job title
[79:25.00]Tell us more about that
[79:27.00]What's going on there
[79:29.00]That is a very big cloud of things
[79:31.00]I would just say
[79:33.00]I think it's an emergent industry
[79:35.00]I've seen this happen repeatedly
[79:37.00]So the general term
[79:39.00]So the general term is software engineer
[79:41.00]Or programmer
[79:43.00]In the 70s and 80s
[79:45.00]There would not be a senior engineer
[79:47.00]There would just be an engineer
[79:49.00]I don't think you would even call
[79:51.00]What about a member of the technical staff
[79:53.00]Oh yeah MTS
[79:55.00]Very very elite
[79:57.00]So these striations appear
[79:59.00]When the population grows
[80:01.00]And the technical depth grows
[80:03.00]Over time
[80:05.00]When it starts
[80:07.00]Not that important
[80:09.00]It's just going to specialize
[80:11.00]I've seen this happen for front end
[80:13.00]For DevOps, for data
[80:15.00]I can't remember what else I listed in that piece
[80:17.00]But those are the main three that I was around for
[80:19.00]Now a lot of people are arguing
[80:21.00]That there is the ML researcher
[80:23.00]The ML engineer
[80:25.00]Who sort of pairs with the researcher
[80:27.00]Sometimes they also call research engineer
[80:29.00]And then on the other side of the fence
[80:31.00]It's just software engineers
[80:33.00]And that's how it was until about last year
[80:35.00]And now there's this specializing
[80:37.00]And rising class of people
[80:39.00]Building AI specific software
[80:41.00]That are not any of those previous titles
[80:43.00]That I just mentioned
[80:45.00]And that's the thesis of the AI engineer
[80:47.00]In the emerging category of start-ups
[80:49.00]Of jobs
[80:51.00]I've had people from Meta, IBM, Microsoft
[80:53.00]Open AI tell me that their title
[80:55.00]Is now AI engineer
[80:57.00]So like I can see that this is a trend
[80:59.00]And I think that's what Andre called out
[81:01.00]In his post that like just mathematically
[81:03.00]Just the limitations in terms of talents
[81:05.00]Research talents and GPUs
[81:07.00]That all these will tend to concentrate
[81:09.00]In a few labs
[81:11.00]And everyone else
[81:13.00]Are just going to have to rely on them
[81:15.00]Or build differentiation of products
[81:17.00]In other ways, and those will be AI engineers
[81:19.00]So mathematically there will be more AI engineers
[81:21.00]Than ML engineers, it's just the truth
[81:23.00]Right now it's the other way
[81:25.00]Right now the number of AI engineers
[81:27.00]Is maybe 10x less
[81:29.00]So I think that the ratio will invert
[81:31.00]And I think the goal of it in space
[81:33.00]And the goal of the conference
[81:35.00]And anything else I do is to serve
[81:37.00]That growing audience
[81:39.00]To make the distinction clear
[81:41.00]If I'm a software engineer
[81:43.00]What do I have to learn
[81:45.00]What additional capabilities does that
[81:47.00]Type of engineer have
[81:49.00]Funny you say that
[81:51.00]I don't actually have a specific blog
[81:53.00]Post on how to like
[81:55.00]Change classes
[81:57.00]I do think I always think about this
[81:59.00]In terms of Baldur's Gate
[82:01.00]DND ruleset number
[82:03.00]5.1 or whatever
[82:05.00]So I kind of intentionally left that open
[82:07.00]To leave space for others
[82:09.00]I think when you start an industry
[82:11.00]That's the only way to guarantee
[82:13.00]That it will fail
[82:15.00]I do have a take
[82:17.00]Obviously because a lot of people
[82:19.00]Are asking me where to start
[82:21.00]And I think basically
[82:23.00]So what we have is
[82:25.00]Late and Space University
[82:27.00]We just finished working on
[82:29.00]Day 7 today
[82:31.00]It's a seven-day project
[82:33.00]And I think we've done a great job
[82:35.00]We've done a great job
[82:37.00]We've done a great job
[82:39.00]And we've just finished working on
[82:41.00]Day 7 today
[82:43.00]It's a seven-day email course
[82:45.00]Where it basically like
[82:47.00]It is completely designed to answer
[82:49.00]The question of like
[82:51.00]I'm an existing software engineer
[82:53.00]I know how to code
[82:55.00]But I don't get all this AI stuff
[82:57.00]I've been living under a rock
[82:59.00]Or like it's just too overwhelming for me
[83:01.00]You have to pick for me
[83:03.00]Or curate for me as a trusted friend
[83:05.00]And I have one hour a day for seven days
[83:07.00]It's image generation
[83:09.00]It's code generation
[83:11.00]It's audio
[83:13.00]ASR
[83:15.00]Audio speech recognition
[83:17.00]And then I forget
[83:19.00]What the fifth and sixth one is
[83:21.00]But the last day is agents
[83:23.00]And so basically I'm just like
[83:25.00]Here are seven projects that you should do
[83:27.00]To feel like you can do anything in AI
[83:29.00]You can't really do everything in AI
[83:31.00]Just from that small list
[83:33.00]But I think it's just like anything
[83:35.00]Go through like a set list
[83:37.00]Of things that are basic skills
[83:39.00]That I think everyone in this industry should have
[83:41.00]To be at least conversant
[83:43.00]In if someone, if like a boss comes to you
[83:45.00]And goes like hey can we build this
[83:47.00]You don't even know if the answer is no
[83:49.00]So I want you to move towards
[83:51.00]From like unknown unknowns to at least known unknowns
[83:53.00]And I think that's where you start
[83:55.00]Being competent as an engineer
[83:57.00]So yeah that's LSU
[83:59.00]In Space University just to trigger the tigers
[84:03.00]So do you think in the future that people
[84:05.00]An AI engineer is going to be someone's
[84:07.00]Full-time job like people are just going to be
[84:09.00]AI engineers or do you think it's going to be
[84:11.00]More of a world where I'm a software engineer
[84:13.00]And like 20% of my time
[84:15.00]I'm using open ai's, api's
[84:17.00]And I'm working on prompt engineering
[84:19.00]And stuff like that and using code pilot
[84:21.00]You just reminded me of day six's
[84:23.00]Open source models and fine tuning
[84:25.00]I think it will be a spectrum. That's why I don't want to be
[84:27.00]Like too definitive about it. Like we have
[84:29.00]Full-time front-end engineers and we have part-time
[84:31.00]And you dip into that community whenever you want
[84:33.00]But wouldn't it be nice if there was a
[84:35.00]Collective name for that community
[84:37.00]So you could go find it, you could find each other
[84:39.00]And like honestly that's really it
[84:41.00]Like a lot of people, a lot of companies are pinging me
[84:43.00]For like hey I want to hire this kind of person
[84:45.00]But you can't hire that person
[84:47.00]But I want someone like that
[84:49.00]And then people on the labor side
[84:51.00]Were pinging me going like okay I want to do more
[84:53.00]In this space but where do I go
[84:55.00]And I think just having that shelling point
[84:57.00]Of what an industry title or name is
[84:59.00]And sort of building out that mythology
[85:01.00]And community and conference
[85:03.00]I think is helpful hopefully
[85:05.00]And I don't have any prescriptions
[85:07.00]On whether or not it's a full-time job
[85:09.00]I do think over time it's going to become
[85:11.00]More of a full-time job
[85:13.00]And that's great for the people who want to do that
[85:15.00]And the companies that want to employ that
[85:17.00]But it's absolutely like you can take it part-time
[85:19.00]Like jobs come in many formats
[85:21.00]Yep that makes sense
[85:23.00]And then you have a huge world fair
[85:25.00]Coming up
[85:27.00]Tell me about that
[85:29.00]So part of I think
[85:31.00]What creating industry requires
[85:33.00]Is to let people gather in one place
[85:35.00]And also for me
[85:37.00]To get high quality talks out of people
[85:39.00]You have to create an event out of it
[85:41.00]Otherwise they don't do the work
[85:43.00]So last year we did
[85:45.00]The AI engineer summit which went very well
[85:47.00]And people can see that online
[85:49.00]And we're very happy with how that turned out
[85:51.00]This year we want to go four times bigger
[85:53.00]With the world fair
[85:55.00]To try to reflect AI engineering
[85:57.00]As it is in 2024
[85:59.00]I always admired
[86:01.00]Two conferences in this respect
[86:03.00]One is NEURUPs which I went to last year
[86:05.00]And documented on the pod which was fantastic
[86:07.00]And two which is KubeCon
[86:09.00]From the other side of my life
[86:11.00]Which is the sort of cloud registration
[86:13.00]In DevOps world
[86:15.00]So NEURUPs is the one place that you go to
[86:17.00]To I think it's the top conference
[86:19.00]I mean there's others
[86:21.00]That you can kind of consider
[86:23.00]So NEURUPs
[86:25.00]NEURUPs is where the research sciences are the stars
[86:27.00]The researchers are the stars, the PhDs are the stars
[86:29.00]Mostly it's just PhDs on the job market
[86:31.00]It's really funny to go
[86:33.00]Especially these things
[86:35.00]It's really funny to go to NEURUPs
[86:37.00]And the PhDs trying to back them
[86:39.00]There are lots of VCs trying to back there
[86:41.00]This year
[86:43.00]So NEURUPs research sciences are the stars
[86:45.00]And I wanted for AI engineers
[86:47.00]Engineer to be the star
[86:49.00]To show off their tooling
[86:51.00]And their techniques
[86:53.00]And their difficulty
[86:55.00]Moving all these ideas from research into production
[86:57.00]The other one was KubeCon
[86:59.00]Where you could honestly just go
[87:01.00]And not attend any of the talks
[87:03.00]And just walk the floor
[87:05.00]And figure out what's going on in DevOps
[87:07.00]Which is fantastic
[87:09.00]So that curation
[87:11.00]And that bringing together of an industry
[87:13.00]Is what I'm going for for the conference
[87:15.00]And it's coming in June
[87:17.00]The most important thing to be honest
[87:19.00]The most important thing was to buy the domain
[87:21.00]So we got AI.engineer
[87:23.00]People are like engineer is a domain
[87:25.00]And funny enough
[87:27.00].engineer was cheaper than.engineering
[87:29.00]I don't understand why
[87:31.00]But that's up to the domain people
[87:33.00]All right
[87:35.00]Josh, any questions on agents
[87:37.00]Yeah, I think maybe you have a lot of
[87:39.00]Experience and exposure
[87:41.00]Talking to all these companies and founders
[87:43.00]And researchers and everyone that's on your podcast
[87:45.00]Do you have like
[87:47.00]Do you feel like you have a good kind of perspective
[87:49.00]On some of the things that like
[87:51.00]Some of the kind of technical issues having seen
[87:53.00]Like we were just talking about like for
[87:55.00]Coding agents like oh how you know
[87:57.00]The value of test is really important
[87:59.00]There are other things like for you know retrieval
[88:01.00]Like now, you know, we have these models
[88:03.00]Coming out with a million context, you know
[88:05.00]Are a million tokens of context like 30 million
[88:07.00]Is retrieval going to matter anymore
[88:09.00]The huge context matter like what do you think
[88:11.00]Specific about the long context thing
[88:13.00]Sure, yeah
[88:15.00]I was going to ask a few other ones after that
[88:17.00]So go for that one first
[88:19.00]That's what I was going to ask first
[88:21.00]Yeah, let's talk about the long context
[88:23.00]So for those who don't know
[88:25.00]Long context was kind of
[88:27.00]In the air last year but really
[88:29.00]Really really really came into focus this year
[88:31.00]With Gemini1.5 having
[88:33.00]Million token context and saying that
[88:35.00]It was in research for 10 million tokens
[88:37.00]And that means that
[88:39.00]You can put you like
[88:41.00]No longer have to really think about
[88:43.00]What you retrieve
[88:45.00]No longer really think about
[88:47.00]What you have to put into context
[88:49.00]You can just kind of throw the entire
[88:51.00]Knowledge base in there or books or film
[88:53.00]Anything like that and that's fantastic
[88:55.00]A lot of people are thinking that it kills
[88:57.00]Rag and I think like one that's not true
[88:59.00]Because for any kind of cost reason
[89:01.00]You know, you still pay per token
[89:03.00]So basically Google is like perfectly happy
[89:05.00]To let you pay a million tokens
[89:07.00]Every single time you make an API call
[89:09.00]But good luck, you know, having $100 API call
[89:11.00]You don't want to be slow, no explanation needed
[89:13.00]And then finally my criticism of
[89:15.00]Long context is that it's also not debuggable
[89:17.00]Like if something goes wrong with the result
[89:19.00]You can't do like the rags decomposition
[89:21.00]Of where the source of error
[89:23.00]Like you just have to like go like
[89:25.00]It's the way it's bro like it's somewhere in there
[89:27.00]I'm sorry, I pretty strongly agree with this
[89:29.00]Why do you think people are making such
[89:31.00]Crazy Long context windows
[89:32.00]People love to kill rag
[89:33.00]It's so much because it's too expensive
[89:35.00]It's so expensive like you said
[89:37.00]Yeah, I just think I'm just calling it
[89:39.00]It's a different dimension
[89:40.00]I think it's an option that's great when it's there
[89:42.00]Like when I'm prototyping
[89:43.00]I do not ever want to worry about context
[89:45.00]And I'm going to call stuff a few times
[89:47.00]And I don't want to run the errors
[89:48.00]I don't want to have it set up a complex retrieval system
[89:50.00]Just to prototype something
[89:52.00]But once I done prototyping
[89:53.00]Then I'll worry about all the other rags stuff
[89:55.00]And yes, I'm going to buy some system
[89:57.00]Or build some system or whatever to go do that
[90:00.00]So I think it's just like an improvement
[90:03.00]In like one dimension that you need
[90:05.00]But the improvements in the other dimensions
[90:07.00]And it's all needed
[90:08.00]Like this space isn't going to keep growing
[90:10.00]In unlimited fashion
[90:12.00]I do think that this combined with multi-modality
[90:16.00]Does unlock new things
[90:18.00]So that's what I was going to ask about next
[90:19.00]It's like how important is multimodal
[90:21.00]Like great, you know, generating videos
[90:23.00]Sure, whatever
[90:24.00]Okay, how many of us need to generate videos that often
[90:26.00]It'd be cool for TV shows, sure, but like, yeah
[90:28.00]I think it's pretty important
[90:30.00]The one thing that in
[90:31.00]When we launched the in space podcast
[90:33.00]We listed a bunch of interest areas
[90:35.00]One thing I love about being explicit
[90:37.00]Or intentional about our work
[90:40.00]Is that you list the things that you're interested in
[90:42.00]And you list the things that you're not interested in
[90:44.00]And people are very unwilling
[90:46.00]To have an entire interest list
[90:48.00]One of the things that we are not interested in
[90:50.00]Was multimodality last year
[90:52.00]Because everyone was
[90:54.00]I was just like, okay, you can generate images
[90:56.00]And they're pretty, but like, not a giant business
[90:58.00]I was wrong
[90:59.00]Journey is a giant, giant, massive business
[91:01.00]That no one can understand or get into
[91:03.00]But also, I think
[91:05.00]Being able to natively understand
[91:07.00]Audio and video and code
[91:09.00]I consider code a special modality
[91:11.00]All that is
[91:13.00]Very, like, qualitatively different
[91:15.00]Then translating it into English first
[91:17.00]And using English as, you know, like a bottleneck
[91:19.00]Or pipe
[91:20.00]And then, you know, applying it in LLMs
[91:22.00]The ability of LLMs to reason across modalities
[91:25.00]Gives you something more than you could
[91:27.00]Individually by using Texas
[91:29.00]The universal interface
[91:30.00]So I think that's useful
[91:32.00]So concretely, what does that mean?
[91:34.00]It means that
[91:35.00]So, I think the reference post for everyone
[91:37.00]That you should have in your head
[91:39.00]Is Simon Willison's post on Gemini1.5's
[91:41.00]Video capability
[91:43.00]Where he basically shot a video of
[91:45.00]His bookshelf, just kind of scanning through it
[91:47.00]And he was able to give back a complete
[91:49.00]JSON list of the books and the authors
[91:51.00]And all the details that were visible there
[91:53.00]Helucinated some of it
[91:55.00]Which is, you know, another issue
[91:57.00]But I think it's just like unlocks this use case
[91:59.00]That you just would not even try to code
[92:01.00]Without the native video understanding capability
[92:04.00]And obviously, like, on a technical level
[92:07.00]Video is just a bunch of frames
[92:08.00]So it actually is just image understanding
[92:10.00]But image within the temporal dimension
[92:12.00]Which this month, I think
[92:14.00]Became much more of a important thing
[92:16.00]Like the integration of space and time
[92:18.00]In transformers
[92:19.00]I don't think anyone was really talking about that
[92:21.00]Until this month
[92:22.00]And now it's the only thing anyone can ever think about
[92:24.00]For Sora and for all the other stuff
[92:26.00]The last thing I'll say
[92:28.00]Which is against this trend of
[92:30.00]Every modality is important
[92:32.00]They just do all the modalities
[92:34.00]I kind of agree with Nat Friedman
[92:36.00]Who actually kind of pointed out
[92:37.00]Just before the Gemini thing blew up
[92:39.00]This month
[92:41.00]Which was, like, why is it that
[92:43.00]OpenAI is pushing Dolly so hard
[92:45.00]Why is Bing pushing Bing Image Creator?
[92:47.00]Like, it's not apparent
[92:49.00]That you have to create images to create AGI
[92:51.00]But every lab just seems to want to do this
[92:54.00]And I kind of agree
[92:55.00]That it's not on the critical path
[92:57.00]Especially for image generation
[92:59.00]Maybe image understanding, video understanding
[93:01.00]Yeah, consumption
[93:02.00]But generation
[93:04.00]Maybe we'll be wrong next year
[93:06.00]Just catches you a bunch of flak with, like, you know
[93:08.00]Culture war things
[93:10.00]It's true
[93:11.00]All right, we're going to move into
[93:13.00]Rapidfire Q&A
[93:14.00]So, we're going to ask you a question
[93:16.00]I don't want to overthink it, baby
[93:18.00]We're going to audience the Q&A
[93:20.00]So, I'll tell you
[93:22.00]All right
[93:23.00]We've cut the Q&A section for time
[93:25.00]So, if you want to hear the spicy questions
[93:27.00]Head over to the Thursday nights in A.I. video
[93:29.00]for the full discussion
[93:31.00]Next up, we have another former guest
[93:33.00]Dylan Patel of Semi-analysis
[93:36.00]the inventor of the GPU rich poor divide
[93:39.00]who did a special live show with us in March
[93:41.00]But that means you can finally
[93:45.00]side-to-side A.V. test your favorite boba shops
[93:48.00]We got Gongcha, we got boba guys
[93:50.00]We got the lemon, whatever it's called
[93:53.00]So, let us know what's your favorite
[93:55.00]We also have Slido up to submit questions
[93:58.00]We already had Dylan on the podcast
[94:00.00]And this guy tweets and writes about all kinds of stuff
[94:02.00]So, we want to know what people want to know more about
[94:05.00]Rather than just being self-truined
[94:07.00]But we'll do a state of the union, maybe
[94:10.00]Everybody wants to know about GROC
[94:12.00]Everybody wants to know whether or not
[94:14.00]MVD is going to zero after GROC
[94:16.00]Everybody wants to know what's going on with AMD
[94:18.00]We got some AMD folks in the crowd, too
[94:20.00]So, feel free to interact at any time
[94:23.00]We have HECCLE, HECCLE, please
[94:25.00]Good comedians show their color
[94:28.00]The way they can handle the crowd when they're HECCLE
[94:31.00]Do not throw boba
[94:33.00]Do not throw boba at the stand
[94:35.00]We cannot afford another podcast thing set up
[94:37.00]Awesome, welcome everybody
[94:39.00]To the seminalysis and latest space crossover
[94:41.00]Dylan texted me on signal
[94:43.00]He was like, dude, how do I easily set up a meetup
[94:46.00]And here we are today
[94:48.00]As you might have seen, there's no name tags
[94:50.00]There's a bunch of things that are missing
[94:52.00]But we did our best
[94:53.00]It was extremely easy, right?
[94:55.00]Like, I texted Alessio, he's like, yo, I got the spot
[94:58.00]Okay, cool, here's a link, send it to people
[95:00.00]Send it, and then show it up
[95:03.00]And like, there was zero other organization that I required
[95:07.00]So, everybody's here
[95:09.00]A lot of seminalysis fans, we get in the crowd
[95:12.00]Everybody wants to know more about
[95:14.00]What's going on today and GROC has definitely been the hottest thing
[95:16.00]We just recorded our monthly podcast today
[95:18.00]And we didn't talk that much about GROC
[95:20.00]Because we wanted you to talk more about
[95:22.00]And then we'll splice you into our monthly recap
[95:24.00]So, let's start there
[95:26.00]So, you guys are the two GROC spreadsheetors
[95:29.00]So, we broke out some GROC numbers
[95:32.00]Because everyone was wondering
[95:33.00]There's two things going on, right?
[95:34.00]One, you know, how important, you know
[95:36.00]How does it achieve the inference speed that it does
[95:38.00]That it has been demonstrated by GROC chat
[95:41.00]And two, how does it achieve its price promise
[95:44.00]That is promised, that is sort of the public pricing
[95:46.00]Of 27 cents per million token
[95:48.00]And there's been a lot of speculation
[95:50.00]Or, you know, some numbers thrown out there
[95:52.00]I put out some tentative numbers
[95:54.00]And you put out different numbers
[95:55.00]But I'll just kind of lay that as the groundwork
[95:58.00]Like, everyone's very excited about
[96:00.00]Essentially like five times faster
[96:02.00]Token generation than any other LLM currently
[96:05.00]And that unlocks interesting downstream possibilities
[96:08.00]If it's sustainable
[96:10.00]If it's affordable
[96:11.00]And so I think your question
[96:13.00]Or reading your piece on GROC
[96:15.00]Which is on the screen right now
[96:16.00]Is it sustainable
[96:18.00]So, like many things
[96:20.00]This is VC funded, including this Boba
[96:23.00]No, I'm just kidding
[96:24.00]I'm paying for the Boba
[96:25.00]Thank you, Semunas, the subscribers
[96:27.00]I hope he pays for it
[96:29.00]I pay for it right now
[96:30.00]That's true, Alessio has the IOU
[96:33.00]Right?
[96:34.00]And that's all it is
[96:35.00]But yeah, like many things, you know
[96:37.00]They're not making money off of their inference service
[96:40.00]They're just throwing it out there for cheap
[96:42.00]And hoping to get business
[96:43.00]And maybe raise money off of that
[96:45.00]And I think that's a that's a fine use case
[96:48.00]But the question is like
[96:49.00]How much money are they losing, right?
[96:51.00]And that's sort of what I went through
[96:52.00]Breaking down in this article
[96:54.00]That's on the screen
[96:55.00]And that's it's pretty clear
[96:56.00]They're like seven to ten X off
[97:00.00]Like break even on their inference API
[97:02.00]Which is like horrendous
[97:04.00]Like far worse than any other
[97:06.00]Sort of inference API provider
[97:08.00]So this is like a simple simple
[97:10.00]Cost thing that was pulled up
[97:12.00]You can either inference at
[97:13.00]Very high throughput
[97:14.00]Or you can inference at
[97:15.00]Very high very low latency
[97:17.00]With GPUs you can do both
[97:18.00]With GROC you can only do one
[97:20.00]Of course with GROC
[97:21.00]You can do that one faster
[97:22.00]Marginally faster than a
[97:24.00]Inference latency optimized GPU server
[97:26.00]But no one offers inference latency
[97:28.00]Optimized GPU servers
[97:29.00]Because you would just burn money
[97:31.00]Makes no economic sense to do so
[97:33.00]Until maybe someone's willing
[97:34.00]To pay for that
[97:35.00]So GROC service
[97:36.00]You know the surface looks awesome
[97:37.00]Compared to everyone else's service
[97:39.00]Which is throughput optimized
[97:41.00]And then when you compare to
[97:43.00]Input optimized scenario
[97:44.00]GPUs look quite slow
[97:46.00]But the reality is
[97:47.00]Is they're serving 64
[97:49.00]128 users at once
[97:51.00]They have a batch size
[97:52.00]How many users are being
[97:53.00]Served at once
[97:54.00]Where as GROC is taking
[97:55.00]576 chips
[97:56.00]And they're not really
[97:58.00]Doing that efficiently
[97:59.00]They're serving a far far
[98:01.00]Fewer number of users
[98:02.00]But extremely fast
[98:03.00]Now that could be worthwhile
[98:05.00]If they can get there
[98:07.00]The number of users
[98:08.00]They're serving at once up
[98:10.00]But that's extremely hard
[98:11.00]Because they don't have
[98:12.00]Their chip
[98:13.00]So they can't store
[98:14.00]KVcash
[98:15.00]KVcash for all the
[98:16.00]Various different users
[98:17.00]So their crux of the issue
[98:19.00]Is just like hey
[98:20.00]Can they get that performance
[98:22.00]Up as much as they claim they will
[98:24.00]They need to get it up
[98:25.00]More than 10x
[98:26.00]To make this like a reasonable
[98:28.00]Benefit
[98:29.00]In the meantime
[98:30.00]NVIDIA is launching a new
[98:31.00]GPU in two weeks
[98:32.00]That'll be fun at GTC
[98:34.00]And they're constantly
[98:35.00]Pushing software as well
[98:36.00]So we'll see
[98:37.00]If GROC can catch up to that
[98:38.00]The current verdict is
[98:40.00]They're quite far behind
[98:41.00]But hopefully
[98:42.00]That maybe they can
[98:43.00]Get there by scaling
[98:44.00]Their system larger
[98:45.00]Yeah
[98:46.00]I was listening back
[98:47.00]To our original episode
[98:48.00]And you were talking
[98:49.00]About how NVIDIA
[98:50.00]Basically adopted
[98:51.00]This different strategy
[98:52.00]Of just leaning on
[98:54.00]Networking GPUs together
[98:55.00]And it seems like
[98:56.00]GROC has some minor
[98:58.00]Version of that going on
[98:59.00]Here with the GROC rack
[99:00.00]Is it enough?
[99:03.00]What's GROC's next
[99:05.00]Step here strategically?
[99:07.00]Yeah, that's
[99:09.00]The next step is of course
[99:10.00]So right now
[99:12.00]They connect 10 racks
[99:13.00]Of chips together
[99:14.00]And that's the system
[99:15.00]That's running
[99:16.00]On their API today
[99:17.00]Whereas most people
[99:18.00]Who are running
[99:19.00]Mischro are running
[99:20.00]In on two GPUs
[99:22.00]One fourth of a server
[99:24.00]And that rack
[99:25.00]Is not
[99:26.00]Obviously 10 racks
[99:27.00]Is pretty crazy
[99:28.00]But they think
[99:29.00]That they can
[99:30.00]Scale performance
[99:31.00]If they have
[99:32.00]This individual system
[99:33.00]Be 20 racks
[99:34.00]They think they can
[99:35.00]Continue to scale performance
[99:36.00]Extra linearly
[99:37.00]So that'd be amazing
[99:38.00]If they could
[99:39.00]And I'm doubtful
[99:40.00]That's going to be something
[99:42.00]That's scalable
[99:43.00]Especially for
[99:45.00]You know, larger models
[99:47.00]So there's the chip itself
[99:51.00]But there's also a lot
[99:52.00]Of work they're doing
[99:53.00]At the compiler level
[99:54.00]Do you have any good sense
[99:55.00]Of like how easy it is
[99:57.00]To actually work with LPU
[99:59.00]Is that something
[100:00.00]That's going to be
[100:01.00]About on that for them
[100:02.00]So Ollie's in the front
[100:03.00]Right there
[100:04.00]And he knows a ton
[100:05.00]About VLIW architectures
[100:07.00]But to summarize
[100:08.00]In his opinion
[100:09.00]And I think many folks
[100:10.00]Is it's extremely hard
[100:11.00]To program
[100:12.00]These sorts of architectures
[100:14.00]Which is why they have
[100:15.00]Their compiler
[100:16.00]And so on and so forth
[100:17.00]But it's an incredible amount
[100:19.00]Of work for them
[100:20.00]To stand up individual models
[100:22.00]And to get the performance
[100:23.00]Up on them
[100:24.00]Which is what they've been
[100:25.00]Working on
[100:26.00]Whereas GPUs
[100:27.00]Are far more flexible
[100:28.00]Of course
[100:29.00]And so the question is
[100:30.00]Can this compiler
[100:31.00]Continue to extract
[100:32.00]Performance
[100:33.00]Well theoretically
[100:34.00]There's a lot more
[100:35.00]Performance to run
[100:36.00]On the hardware
[100:37.00]And they don't have
[100:38.00]Many things that people
[100:40.00]Generally associate with
[100:42.00]Programmable hardware
[100:44.00]They don't have buffers
[100:45.00]And many other things
[100:46.00]So it makes it very tough
[100:47.00]To do that
[100:48.00]But that's what their
[100:49.00]They're relatively large
[100:51.00]Compiler team is working on
[100:53.00]So I'm not a GPU compiler guy
[100:55.00]But I do want to
[100:56.00]Clarify my understanding
[100:57.00]From what I read
[100:58.00]Which is a lot of
[100:59.00]Catching up to do
[101:00.00]It is
[101:01.00]The crux of it
[101:02.00]Is some kind of speculative
[101:04.00]The word that comes
[101:05.00]The routing
[101:06.00]Of weights
[101:08.00]And work
[101:09.00]That needs to
[101:10.00]Be done or scheduling
[101:11.00]Of work across
[101:12.00]The ten racks
[101:14.00]Of GPUs
[101:15.00]Is that like
[101:17.00]The bulk of the benefit
[101:18.00]That you get
[101:19.00]From the compilation
[101:20.00]So with the grok chips
[101:22.00]What's really
[101:23.00]Interesting is like
[101:25.00]With GPUs
[101:26.00]You can issue
[101:27.00]Certain instructions
[101:29.00]And you will get
[101:30.00]A different result
[101:31.00]Like depending on
[101:32.00]The time I know
[101:33.00]A lot of people
[101:34.00]I'll have
[101:35.00]Have had that experience
[101:36.00]Or like the GPU
[101:37.00]Literally doesn't return
[101:38.00]The numbers it should be
[101:39.00]As basically called
[101:40.00]Non-determinism
[101:41.00]With grok
[101:42.00]Their chip is completely
[101:43.00]Deterministic
[101:44.00]The moment you compile it
[101:45.00]You know exactly how long
[101:46.00]It will take to operate
[101:47.00]Right there is no
[101:48.00]There is no like
[101:49.00]Deviation at all
[101:51.00]And so
[101:52.00]You know they've
[101:53.00]They're planning everything
[101:55.00]Ahead of time
[101:56.00]Like every instruction
[101:57.00]Like it will
[101:58.00]Complete in the time
[101:59.00]That they've planned it for
[102:00.00]And there is no
[102:01.00]I don't know what
[102:02.00]The best way to state this is
[102:03.00]No variance there
[102:04.00]Which is
[102:05.00]Interesting from like
[102:06.00]When you look historically
[102:07.00]They tried to push this
[102:08.00]In automotive
[102:09.00]Because automotive
[102:10.00]You probably want
[102:11.00]Your car to do
[102:12.00]Exactly what you
[102:13.00]I issued it to do
[102:14.00]And not have
[102:15.00]Unpredictability
[102:16.00]But yeah
[102:17.00]Sorry I lost track
[102:18.00]Of the question
[102:19.00]It's okay
[102:20.00]I just wanted to
[102:21.00]Understand a little bit
[102:22.00]More about like
[102:23.00]What people should
[102:24.00]Should know about
[102:25.00]The compiler magic
[102:26.00]That goes on with grok
[102:27.00]Like you know
[102:28.00]Like I think
[102:30.00]From a software
[102:31.00]Like hardware point of view
[102:32.00]That intersection of
[102:34.00]I guess
[102:35.00]So chips have like
[102:36.00]Like I'm still this
[102:37.00]From someone
[102:38.00]Here in the crowd
[102:39.00]But chips have like
[102:40.00]Five you know
[102:41.00]Sort of there's
[102:42.00]Like when you're designing a chip
[102:43.00]There's there's
[102:44.00]It's called PPA right
[102:45.00]Power performance in area
[102:47.00]It's kind of a triangle
[102:48.00]That you optimize around
[102:49.00]And the one thing
[102:50.00]People don't realize
[102:51.00]Is there's a
[102:52.00]There's a third P
[102:53.00]That's like PPA P
[102:55.00]And the last P
[102:56.00]Is pain in the ass
[102:57.00]To program
[102:58.00]And that's
[102:59.00]That is very important
[103:00.00]For like
[103:01.00]High hardware
[103:02.00]Like TPU
[103:04.00]Without the
[103:05.00]Hundreds of people
[103:06.00]That work on the compiler
[103:07.00]And jacks
[103:08.00]And XLA
[103:09.00]And all these sorts of things
[103:10.00]Would be a pain in the ass
[103:11.00]To program
[103:12.00]But Google's got that like
[103:13.00]Plumbing
[103:14.00]Now if you look across
[103:15.00]The ecosystem
[103:16.00]Everything else is a pain
[103:17.00]In the ass to program
[103:18.00]Compared to NVIDIA
[103:19.00]And this applies
[103:20.00]To the grok chip as well
[103:22.00]So yeah
[103:23.00]Question is like
[103:24.00]Can the compiler team
[103:25.00]Get performance up
[103:26.00]Anywhere close to theoretical
[103:28.00]And then can they make
[103:29.00]It not a pain in the ass
[103:30.00]To program
[103:31.00]And then can they make
[103:32.00]It not a pain in the ass
[103:33.00]To program
[103:34.00]And then can they make
[103:35.00]It not a pain in the ass
[103:36.00]To program
[103:37.00]And then can they make
[103:38.00]It not a pain in the ass
[103:39.00]And then can they make
[103:40.00]It not a pain in the ass
[103:41.00]And then can they make
[103:42.00]It not a pain in the ass
[103:43.00]And then can they make
[103:44.00]It not a pain in the ass
[103:45.00]And then can they make
[103:46.00]It not a pain in the ass
[103:47.00]And then can they make
[103:48.00]It not a pain in the ass
[103:49.00]And then can they make
[103:50.00]It not a pain in the ass
[103:51.00]And then can they make
[103:52.00]It not a pain in the ass
[103:53.00]And then can they make
[103:54.00]It not a pain in the ass
[103:55.00]And then can they make
[103:56.00]It not a pain in the ass
[103:57.00]And then can they make
[103:58.00]It not a pain in the ass
[103:59.00]And then can they make
[104:00.00]It not a pain in the ass
[104:01.00]And then can they make
[104:02.00]It not a pain in the ass
[104:03.00]And then can they make
[104:04.00]It not a pain in the ass
[104:05.00]And then can they make
[104:06.00]It not a pain in the ass
[104:07.00]And then can they make
[104:08.00]It not a pain in the ass
[104:09.00]And then can they make
[104:10.00]It not a pain in the ass
[104:11.00]And then can they make
[104:12.00]It not a pain in the ass
[104:13.00]And then can they make
[104:14.00]It not a pain in the ass
[104:15.00]And then can they make
[104:16.00]It not a pain in the ass
[104:17.00]And then can they make
[104:18.00]It not a pain in the ass
[104:19.00]And then can they make
[104:20.00]It not a pain in the ass
[104:21.00]And then can they make
[104:22.00]It not a pain in the ass
[104:23.00]And then can they make
[104:24.00]It not a pain in the ass
[104:25.00]And then can they make
[104:26.00]It not a pain in the ass
[104:27.00]And then can they make
[104:28.00]It not a pain in the ass
[104:29.00]And then can they make
[104:30.00]It not a pain in the ass
[104:31.00]And then can they make
[104:32.00]It not a pain in the ass
[104:33.00]And then can they make
[104:34.00]It not a pain in the ass
[104:35.00]And then can they make
[104:36.00]It not a pain in the ass
[104:37.00]And then can they make
[104:38.00]It not a pain in the ass
[104:39.00]And then can they make
[104:40.00]It not a pain in the ass
[104:41.00]And then can they make
[104:42.00]It not a pain in the ass
[104:43.00]And then can they make
[104:44.00]It not a pain in the ass
[104:45.00]And then can they make
[104:46.00]It not a pain in the ass
[104:47.00]And then can they make
[104:48.00]It not a pain in the ass
[104:49.00]And then can they make
[104:50.00]It not a pain in the ass
[104:51.00]And then can they make
[104:52.00]It not a pain in the ass
[104:53.00]And then can they make
[104:54.00]It not a pain in the ass
[104:55.00]And then can they make
[104:56.00]It not a pain in the ass
[104:57.00]And then can they make
[104:58.00]It not a pain in the ass
[104:59.00]And then can they make
[105:00.00]It not a pain in the ass
[105:01.00]And then can they make
[105:02.00]It not a pain in the ass
[105:03.00]And then can they make
[105:04.00]It not a pain in the ass
[105:05.00]And then can they make
[105:06.00]It not a pain in the ass
[105:07.00]And then can they make
[105:08.00]It not a pain in the ass
[105:09.00]And then can they make
[105:10.00]It not a pain in the ass
[105:11.00]And then can they make
[105:12.00]It not a pain in the ass
[105:13.00]And then can they make
[105:14.00]It not a pain in the ass
[105:15.00]And then can they make
[105:16.00]It not a pain in the ass
[105:17.00]And then can they make
[105:18.00]It not a pain in the ass
[105:19.00]And then can they make
[105:20.00]It not a pain in the ass
[105:21.00]And then can they make
[105:22.00]It not a pain in the ass
[105:23.00]And then can they make
[105:24.00]It not a pain in the ass
[105:25.00]And then can they make
[105:26.00]It not a pain in the ass
[105:27.00]And then can they make
[105:28.00]It not a pain in the ass
[105:29.00]And then can they make
[105:30.00]It not a pain in the ass
[105:31.00]And then can they make
[105:32.00]It not a pain in the ass
[105:33.00]And then can they make
[105:34.00]It not a pain in the ass
[105:35.00]And then can they make
[105:36.00]It not a pain in the ass
[105:37.00]And then can they make
[105:38.00]It not a pain in the ass
[105:39.00]And then can they make
[105:40.00]It not a pain in the ass
[105:41.00]And then can they make
[105:42.00]It not a pain in the ass
[105:43.00]And then can they make
[105:44.00]It not a pain in the ass
[105:45.00]And then can they make
[105:46.00]It not a pain in the ass
[105:47.00]And then can they make
[105:48.00]It not a pain in the ass
[105:49.00]And then can they make
[105:50.00]It not a pain in the ass
[105:51.00]And then can they make
[105:52.00]It not a pain in the ass
[105:53.00]And then can they make
[105:54.00]It not a pain in the ass
[105:55.00]And then can they make
[105:56.00]It not a pain in the ass
[105:57.00]And then can they make
[105:58.00]It not a pain in the ass
[105:59.00]And then can they make
[106:00.00]It not a pain in the ass
[106:01.00]And then can they make
[106:02.00]It not a pain in the ass
[106:03.00]And then can they make
[106:04.00]It not a pain in the ass
[106:05.00]And then can they make
[106:06.00]It not a pain in the ass
[106:07.00]And then can they make
[106:08.00]It not a pain in the ass
[106:09.00]And then can they make
[106:10.00]It not a pain in the ass
[106:11.00]And then can they make
[106:12.00]It not a pain in the ass
[106:13.00]And then can they make
[106:14.00]It not a pain in the ass
[106:15.00]And then can they make
[106:16.00]It not a pain in the ass
[106:17.00]And then can they make
[106:18.00]It not a pain in the ass
[106:19.00]And then can they make
[106:20.00]It not a pain in the ass
[106:21.00]And then can they make
[106:22.00]It not a pain in the ass
[106:23.00]And then can they make
[106:24.00]It not a pain in the ass
[106:25.00]And then can they make
[106:26.00]It not a pain in the ass
[106:27.00]And then can they make
[106:28.00]It not a pain in the ass
[106:29.00]And then can they make
[106:30.00]It not a pain in the ass
[106:31.00]And then can they make
[106:32.00]It not a pain in the ass
[106:33.00]And then can they make
[106:34.00]It not a pain in the ass
[106:35.00]And then can they make
[106:36.00]It not a pain in the ass
[106:37.00]And then can they make
[106:38.00]It not a pain in the ass
[106:39.00]And then can they make
[106:40.00]It not a pain in the ass
[106:41.00]And then can they make
[106:42.00]It not a pain in the ass
[106:43.00]And then can they make
[106:44.00]It not a pain in the ass
[106:45.00]And then can they make
[106:46.00]It not a pain in the ass
[106:47.00]And then can they make
[106:48.00]It not a pain in the ass
[106:49.00]And then can they make
[106:50.00]It not a pain in the ass
[106:51.00]And then can they make
[106:52.00]It not a pain in the ass
[106:53.00]And then can they make
[106:54.00]It not a pain in the ass
[106:55.00]And then can they make
[106:56.00]It not a pain in the ass
[106:57.00]And then can they make
[106:58.00]It not a pain in the ass
[106:59.00]And then can they make
[107:00.00]It not a pain in the ass
[107:01.00]And then can they make
[107:02.00]It not a pain in the ass
[107:03.00]And then can they make
[107:04.00]It not a pain in the ass
[107:05.00]And then can they make
[107:06.00]It not a pain in the ass
[107:07.00]And then can they make
[107:08.00]It not a pain in the ass
[107:09.00]And then can they make
[107:10.00]It not a pain in the ass
[107:11.00]And then can they make
[107:12.00]It not a pain in the ass
[107:13.00]And then can they make
[107:14.00]It not a pain in the ass
[107:15.00]And then can they make
[107:16.00]It not a pain in the ass
[107:17.00]And then can they make
[107:18.00]It not a pain in the ass
[107:19.00]And then can they make
[107:20.00]It not a pain in the ass
[107:21.00]And then can they make
[107:22.00]It not a pain in the ass
[107:23.00]And then can they make
[107:24.00]It not a pain in the ass
[107:25.00]And then can they make
[107:26.00]It not a pain in the ass
[107:27.00]And then can they make
[107:28.00]It not a pain in the ass
[107:29.00]And then can they make
[107:30.00]It not a pain in the ass
[107:31.00]And then can they make
[107:32.00]It not a pain in the ass
[107:33.00]And then can they make
[107:34.00]It not a pain in the ass
[107:35.00]And then can they make
[107:36.00]It not a pain in the ass
[107:37.00]And then can they make
[107:38.00]It not a pain in the ass
[107:39.00]And then can they make
[107:40.00]It not a pain in the ass
[107:41.00]And then can they make
[107:42.00]It not a pain in the ass
[107:43.00]And then can they make
[107:44.00]It not a pain in the ass
[107:45.00]And then can they make
[107:46.00]It not a pain in the ass
[107:47.00]And then can they make
[107:48.00]It not a pain in the ass
[107:49.00]And then can they make
[107:50.00]It not a pain in the ass
[107:51.00]And then can they make
[107:52.00]It not a pain in the ass
[107:53.00]And then can they make
[107:54.00]It not a pain in the ass
[107:55.00]And then can they make
[107:56.00]It not a pain in the ass
[107:57.00]And then can they make
[107:58.00]It not a pain in the ass
[107:59.00]And then can they make
[108:00.00]It not a pain in the ass
[108:01.00]And then can they make
[108:02.00]And then can they make
[108:03.00]And then can they make
[108:04.00]And then can they make
[108:05.00]It not a pain in the ass
[108:06.00]And then can they make
[108:07.00]It not a pain in the ass
[108:08.00]And then can they make
[108:09.00]It not a pain in the ass
[108:10.00]And then can they make
[108:11.00]It not a pain in the ass
[108:12.00]And then can they make
[108:13.00]It not a pain in the ass
[108:14.00]And then can they make
[108:15.00]It not a pain in the ass
[108:16.00]And then can they make
[108:17.00]It not a pain in the ass
[108:18.00]And then can they make
[108:19.00]It not a pain in the ass
[108:20.00]And then can they make
[108:21.00]It not a pain in the ass
[108:22.00]And then can they make
[108:23.00]It not a pain in the ass
[108:24.00]And then can they make
[108:25.00]It not a pain in the ass
[108:26.00]And then can they make
[108:27.00]It not a pain in the ass
[108:28.00]And then can they make
[108:29.00]And then can they make
[108:30.00]It not a pain in the ass
[108:31.00]And then can they make
[108:32.00]It not a pain in the ass
[108:33.00]And then can they make
[108:34.00]It not a pain in the ass
[108:35.00]And then can they make
[108:36.00]It not a pain in the ass
[108:37.00]And then can they make
[108:38.00]It not a pain in the ass
[108:39.00]And then can they make
[108:40.00]It not a pain in the ass
[108:41.00]And then can they make
[108:42.00]It not a pain in the ass
[108:43.00]And then can they make
[108:44.00]It not a pain in the ass
[108:45.00]And then can they make
[108:46.00]It not a pain in the ass
[108:47.00]And then can they make
[108:48.00]It not a pain in the ass
[108:49.00]And then can they make
[108:50.00]It not a pain in the ass
[108:51.00]And then can they make
[108:52.00]It not a pain in the ass
[108:53.00]And then can they make
[108:54.00]It not a pain in the ass
[108:55.00]And then can they make
[108:56.00]It not a pain in the ass
[108:57.00]And then can they make
[108:58.00]It not a pain in the ass
[108:59.00]And then can they make
[109:00.00]It not a pain in the ass
[109:01.00]And then can they make
[109:02.00]It not a pain in the ass
[109:03.00]And then can they make
[109:04.00]It not a pain in the ass
[109:05.00]And then can they make
[109:06.00]It not a pain in the ass
[109:07.00]And then can they make
[109:08.00]It not a pain in the ass
[109:09.00]And then can they make
[109:10.00]It not a pain in the ass
[109:11.00]And then can they make
[109:12.00]It not a pain in the ass
[109:13.00]And then can they make
[109:14.00]It not a pain in the ass
[109:15.00]And then can they make
[109:16.00]It not a pain in the ass
[109:17.00]And then can they make
[109:18.00]It not a pain in the ass
[109:19.00]And then can they make
[109:20.00]It not a pain in the ass
[109:21.00]And then can they make
[109:22.00]It not a pain in the ass
[109:23.00]And then can they make
[109:24.00]It not a pain in the ass
[109:25.00]And then can they make
[109:26.00]It not a pain in the ass
[109:27.00]And then can they make
[109:28.00]It not a pain in the ass
[109:29.00]And then can they make
[109:30.00]It not a pain in the ass
[109:31.00]And then can they make
[109:32.00]It not a pain in the ass
[109:33.00]And then can they make
[109:34.00]It not a pain in the ass
[109:35.00]And then can they make
[109:36.00]It not a pain in the ass
[109:37.00]And then can they make
[109:38.00]It not a pain in the ass
[109:39.00]And then can they make
[109:40.00]It not a pain in the ass
[109:41.00]And then can they make
[109:42.00]It not a pain in the ass
[109:43.00]And then can they make
[109:44.00]It not a pain in the ass
[109:45.00]And then can they make
[109:46.00]It not a pain in the ass
[109:47.00]And then can they make
[109:48.00]It not a pain in the ass
[109:49.00]And then can they make
[109:50.00]It not a pain in the ass
[109:51.00]And then can they make
[109:52.00]It not a pain in the ass
[109:53.00]And then can they make
[109:54.00]It not a pain in the ass
[109:55.00]And then can they make
[109:56.00]It not a pain in the ass
[109:57.00]And then can they make
[109:58.00]It not a pain in the ass
[109:59.00]And then can they make
[110:00.00]It not a pain in the ass
[110:01.00]And then can they make
[110:02.00]It not a pain in the ass
[110:03.00]And then can they make
[110:04.00]It not a pain in the ass
[110:05.00]And then can they make
[110:06.00]It not a pain in the ass
[110:07.00]And then can they make
[110:08.00]It not a pain in the ass
[110:09.00]And then can they make
[110:10.00]It not a pain in the ass
[110:11.00]And then can they make
[110:12.00]It not a pain in the ass
[110:13.00]And then can they make
[110:14.00]It not a pain in the ass
[110:15.00]And then can they make
[110:16.00]It not a pain in the ass
[110:17.00]And so then I just
[110:18.00]I just started editing them
[110:19.00]And so then I just started editing them
[110:20.00]So I have stopped
[110:21.00]Comparing RAG with long
[110:22.00]Context of Fine Tuning
[110:23.00]Hold on, you said I retweeted
[110:24.00]You defending it
[110:25.00]I thought you were hating on it
[110:26.00]And that's why I retweeted it
[110:28.00]It's not one of a defense
[110:30.00]Because everyone was like
[110:31.00]Long context is killing RAG
[110:32.00]And then I had a future
[110:33.00]Oh that should be so quadratic
[110:34.00]That's another one
[110:36.00]And I actually
[110:37.00]Messed the fine print as well
[110:38.00]Let's see
[110:39.00]Power benefits of
[110:40.00]Astrum dominant
[110:41.00]Yeah, so that's a good question
[110:43.00]Astrum is on chip memory
[110:45.00]Everyone's just using HBM
[110:46.00]If you don't have to go
[110:47.00]To off-chip memory
[110:48.00]That'd be really efficient
[110:49.00]Right?
[110:50.00]Because you're
[110:51.00]You're not moving bits around
[110:52.00]But there's always
[110:54.00]The issue of
[110:55.00]You don't have enough memory
[110:56.00]So you still have to move
[110:57.00]Bits around constantly
[110:58.00]And so that's the
[110:59.00]That's the question
[111:00.00]So yeah, sure
[111:01.00]If you can not move data
[111:02.00]Around as you compute
[111:03.00]It's going to be fantastically
[111:04.00]Efficient
[111:05.00]But that isn't really
[111:06.00]It's not really just
[111:07.00]Easier simple to do
[111:08.00]What do you think is going to be
[111:09.00]Harder in the future
[111:10.00]Like getting more energy
[111:11.00]At cheaper cost
[111:12.00]Like getting more of this hardware
[111:14.00]To run
[111:15.00]Yeah, I wonder
[111:16.00]So someone was talking about this earlier
[111:18.00]But it's like
[111:19.00]Here in the crowd
[111:20.00]And I'm looking right at him
[111:21.00]But he's complaining
[111:22.00]That journalists keep saying that
[111:24.00]You know that
[111:25.00]Like misreporting about how data centers
[111:27.00]Or what data centers
[111:28.00]Are doing to the environment
[111:29.00]Right?
[111:30.00]Which I thought was quite funny
[111:31.00]Because they're inundated by
[111:33.00]Journalists talking about data centers
[111:35.00]Like destroying the world
[111:36.00]Anyways, you know
[111:37.00]That's not quite the case
[111:38.00]But yeah, I don't know
[111:39.00]Like the power is certainly
[111:42.00]Gonna be hard to get
[111:44.00]But you know
[111:45.00]I think
[111:46.00]If you just look at history
[111:47.00]Right?
[111:48.00]Like humanity
[111:49.00]Especially America
[111:50.00]Like power
[111:51.00]Power production and usage
[111:52.00]Kept skyrocketing
[111:53.00]From like the 1700s
[111:55.00]To like 1970s
[111:57.00]And then it's kind of
[111:58.00]Flat line from there
[111:59.00]So why can't we like
[112:00.00]Go back to the like growth stage
[112:02.00]I guess it's like the
[112:03.00]The whole like mantra
[112:04.00]Of like accelerationists
[112:06.00]I guess
[112:07.00]This is IAC, yep
[112:08.00]Well, I don't think it's IAC
[112:09.00]I think it's like
[112:10.00]Same Altman
[112:11.00]Hotly believes this too
[112:12.00]And I don't think he's IAC
[112:13.00]So but yeah
[112:14.00]Like I don't think like
[112:15.00]It's like think
[112:16.00]It's like something to
[112:17.00]Think about like
[112:18.00]The US is going back
[112:19.00]To growing an energy usage
[112:21.00]Where as for the last like
[112:22.00]Forty years kind of
[112:24.00]We're flat on energy usage
[112:25.00]And what does that mean
[112:26.00]Like yeah
[112:29.00]There was another question
[112:31.00]On Marvel but kind of the
[112:32.00]I think that's
[112:33.00]It's definitely like
[112:34.00]One of these three guys
[112:35.00]We're on the buy side
[112:36.00]That are asking this question
[112:39.00]Wanna know if Marvell's stock
[112:41.00]Is gonna go up
[112:43.00]So Marvell
[112:44.00]They're doing the
[112:46.00]Customizing for GROC
[112:48.00]They also do the
[112:49.00]Trinium too
[112:50.00]And the Google CPU
[112:51.00]Yeah any other
[112:52.00]Any other chip
[112:53.00]That they're working on
[112:54.00]That people should
[112:55.00]Should keep in mind
[112:56.00]It's like yeah
[112:57.00]Any needle moving
[112:58.00]Any stock moving
[112:59.00]Yeah exactly
[113:01.00]They're working on
[113:02.00]Some more stuff
[113:03.00]Yeah I'll refrain from
[113:05.00]Yeah all right
[113:06.00]Let's see
[113:07.00]Other GROC stuff
[113:08.00]We want to get it
[113:09.00]Get through
[113:10.00]I don't think so
[113:11.00]All right
[113:12.00]Most of other ones
[113:13.00]You're going edge compute hardware
[113:16.00]Any real use cases
[113:17.00]For it
[113:18.00]Yeah I mean
[113:20.00]I have like a
[113:21.00]Really like anti edge view
[113:23.00]So many people were like
[113:25.00]Oh I'm gonna run
[113:26.00]This model on my phone
[113:27.00]Or on my laptop
[113:28.00]And I love
[113:30.00]I love how much
[113:31.00]It's raining so now
[113:32.00]I can be horrible
[113:33.00]And you people won't leave
[113:35.00]Like I want you
[113:36.00]To try and leave
[113:37.00]This building
[113:38.00]Captive audience
[113:40.00]Should I start singing
[113:41.00]Like there's
[113:42.00]Nothing you can do
[113:43.00]You definitely
[113:44.00]I'll stop you from that
[113:45.00]Sorry
[113:46.00]Edge hardware
[113:47.00]Like you know
[113:48.00]People are like
[113:49.00]I'm going to run
[113:50.00]This model on my phone
[113:51.00]Or on my laptop
[113:52.00]It makes no sense to me
[113:53.00]Current hardware
[113:54.00]Is not really capable of it
[113:55.00]So you're going to buy
[113:56.00]Any hardware
[113:57.00]To run
[113:58.00]Whatever on the edge
[113:59.00]Or you're going to
[114:00.00]Just run
[114:01.00]Very very small models
[114:02.00]But in either case
[114:03.00]You're going to end up
[114:05.00]With like
[114:06.00]The performance is really low
[114:07.00]And then whatever you spent
[114:08.00]To run it locally
[114:09.00]In the cloud
[114:10.00]It could service 10x the users
[114:12.00]So it kind of like
[114:14.00]SOL in terms of like
[114:17.00]Economics of
[114:18.00]Running things on the edge
[114:20.00]And then like latency is like
[114:23.00]For LLMs
[114:24.00]Right for LLMs
[114:25.00]It's like
[114:26.00]Not that big of a deal
[114:27.00]Relative
[114:28.00]To like
[114:29.00]Internet latency
[114:30.00]Is not that big of a deal
[114:31.00]Relative to the
[114:32.00]Use of the model
[114:33.00]Right like the actual model
[114:34.00]Operating
[114:35.00]Whether it's on edge hardware
[114:36.00]Or cloud hardware
[114:37.00]And cloud hardware
[114:38.00]Like edge hardware
[114:39.00]Is not really
[114:40.00]Able to like
[114:41.00]Have a measurable
[114:43.00]Appreciable
[114:44.00]Like advantage
[114:45.00]Over cloud hardware
[114:48.00]This applies
[114:49.00]To diffusion models
[114:50.00]This applies to LLMs
[114:52.00]Of course small models
[114:53.00]Will be able to run
[114:54.00]But not all
[114:55.00]Yeah
[114:56.00]What chances
[114:57.00]The startups
[114:58.00]Like medax etched
[114:59.00]Or 5,600
[115:00.00]I think you all interviewed them
[115:01.00]Why don't you answer
[115:02.00]Yeah we have connections
[115:03.00]With medax and lemur
[115:04.00]And we haven't know
[115:06.00]But Gavin is friendly
[115:07.00]They didn't
[115:08.00]Yeah they said
[115:09.00]They don't want to talk publicly
[115:10.00]Yeah
[115:11.00]What they're doing
[115:12.00]It's something like
[115:13.00]When they open up
[115:14.00]We can
[115:15.00]Sure sure
[115:16.00]Yeah
[115:17.00]But do you think like
[115:18.00]I think the two three
[115:19.00]We're going to answer the question
[115:20.00]What do you think of them
[115:21.00]There's a couple things
[115:23.00]It's like
[115:24.00]How do the other companies
[115:26.00]Innovate against them
[115:27.00]I think when you do a new
[115:28.00]Silicon you're like
[115:29.00]Oh we're going to be
[115:30.00]So much better at this thing
[115:31.00]You're like much faster
[115:32.00]Much cheaper
[115:33.00]But there's all the other curves
[115:34.00]Going down
[115:35.00]On the macro environment
[115:36.00]So if it takes you
[115:37.00]Like five years
[115:38.00]Before you were
[115:39.00]Like a lot better
[115:40.00]Five years later
[115:41.00]Once you take the chip out
[115:42.00]You're only comparing yourself
[115:43.00]To the five year advancement
[115:44.00]That the major companies had to
[115:46.00]So then it's like
[115:47.00]Okay that
[115:48.00]We're going to have like
[115:49.00]The C300 whatever
[115:50.00]From from nvidia
[115:51.00]By the time
[115:52.00]Some of these chips come up
[115:53.00]What's after Z
[115:55.00]What do you think is after Z
[115:56.00]In the roadmap
[115:57.00]It's X Y Z
[116:00.00]No
[116:01.00]Anyways
[116:03.00]Yeah yeah
[116:04.00]It's like the age old problem
[116:05.00]You build a chip
[116:06.00]It has some cool thing
[116:07.00]Cool feature
[116:08.00]And then like
[116:09.00]A year later nvidia
[116:10.00]Has it in hardware
[116:11.00]It has implemented
[116:12.00]Some flavor of that in hardware
[116:14.00]Or two generations out
[116:16.00]Like what idea
[116:17.00]Are you going to have
[116:18.00]That nvidia can't implement
[116:19.00]Is like really the question
[116:21.00]It's like you have to be
[116:22.00]Fundamentally different
[116:23.00]In some way
[116:24.00]That holds through
[116:25.00]For four or five years
[116:26.00]That's kind of the big issue
[116:28.00]But you know like
[116:30.00]Like those people have
[116:31.00]Some ideas that are interesting
[116:32.00]And yeah maybe
[116:33.00]It'll work out right
[116:34.00]It's going to be hard
[116:35.00]To fight nvidia
[116:36.00]Who one doesn't
[116:37.00]Consider them competition
[116:38.00]I worried about like
[116:39.00]Google and Amazon's chip
[116:40.00]Right they're not
[116:41.00]And I guess
[116:42.00]To some extent AMD's chip
[116:43.00]But like
[116:44.00]They're not really worried
[116:45.00]About you know
[116:46.00]Maddox or etched
[116:47.00]Or or grok
[116:48.00]Or you know
[116:49.00]Positron or any of these folks
[116:51.00]How much of an advantage
[116:53.00]Do they have
[116:54.00]By working closely
[116:55.00]With like open ai
[116:56.00]And some of these other folks
[116:57.00]And then already knowing
[116:58.00]Where some of the
[116:59.00]Architecture decisions
[117:00.00]Are going and since
[117:01.00]Those companies are like
[117:02.00]The biggest buyers
[117:03.00]Of the chips
[117:04.00]Yeah I mean like
[117:05.00]You see like
[117:06.00]Like the most important
[117:07.00]Sort of ai companies
[117:09.00]Are obviously going to
[117:10.00]Tell hardware vendors
[117:11.00]What they want
[117:12.00]You know open ai
[117:13.00]And you know
[117:14.00]So on and so forth
[117:15.00]Right they can obviously
[117:16.00]Tell them what they want
[117:17.00]And the startups
[117:18.00]Aren't actually going to
[117:19.00]Get anywhere close to
[117:20.00]As much feedback on
[117:21.00]What to do on like
[117:22.00]You know very
[117:23.00]My new low level stuff
[117:24.00]So that is a difficult here
[117:26.00]Some startups
[117:27.00]Like Maddox
[117:28.00]Obviously have
[117:29.00]People who built
[117:30.00]Or worked on
[117:31.00]The largest models
[117:32.00]And other startups
[117:33.00]Might not have
[117:34.00]That advantage
[117:35.00]And so they're always
[117:36.00]Gonna have that issue
[117:37.00]Of like
[117:38.00]Hey how do I get
[117:39.00]The feedback
[117:40.00]Or what's changing
[117:41.00]What do they see
[117:42.00]Down the pipeline
[117:43.00]That's
[117:44.00]That I really need
[117:45.00]To be aware of
[117:46.00]And ready for
[117:47.00]When I design
[117:48.00]My hardware
[117:49.00]All right
[117:50.00]Every hardware shortage
[117:51.00]As eventually
[117:52.00]Turn into a glut
[117:53.00]Will that be
[117:54.00]Through of NVIDIA chips
[117:55.00]It's so when
[117:56.00]But also why
[117:57.00]Absolutely
[117:58.00]And I'm so excited
[117:59.00]To buy like
[118:00.00]H100s
[118:01.00]Not a thousand
[118:02.00]But yeah
[118:03.00]Everyone's
[118:04.00]Gonna buy chips
[118:05.00]It's just the way
[118:06.00]Semiconductors work
[118:07.00]Because the supply chain
[118:08.00]Takes forever to build out
[118:09.00]And it's like
[118:10.00]A really weird thing
[118:11.00]So if the backlog
[118:12.00]Of chips is a year
[118:15.00]People will order
[118:16.00]Two years worth
[118:17.00]Of what they want
[118:18.00]For the next year
[118:19.00]It is like
[118:20.00]A very common thing
[118:21.00]It's not just like
[118:22.00]This AI cycle
[118:23.00]But like
[118:24.00]Like microcontroller
[118:25.00]Like the automotive companies
[118:26.00]They order
[118:27.00]Two years worth
[118:28.00]Of what they needed
[118:29.00]For one year
[118:30.00]What happens
[118:31.00]In semiconductors
[118:32.00]When lead times
[118:33.00]Lengthen
[118:34.00]The purchases
[118:35.00]And inventory
[118:36.00]Is sort of like
[118:37.00]Double
[118:38.00]So these
[118:39.00]The NVIDIA GPU shortage
[118:44.00]Obviously is going to be
[118:45.00]Rectified
[118:46.00]And when it is
[118:47.00]Everyone's sort of
[118:48.00]Double orders
[118:49.00]Will become
[118:50.00]Extremely apparent
[118:51.00]Right
[118:52.00]And you know
[118:53.00]You see like
[118:54.00]Random companies
[118:55.00]Out of nowhere being
[118:56.00]Like yeah
[118:57.00]We've got 32,000
[118:58.00]H100s on order
[118:59.00]And 25,000
[119:00.00]And trust
[119:01.00]They're not all
[119:02.00]They're not all
[119:03.00]Real orders
[119:04.00]For one
[119:05.00]But I think
[119:06.00]The bubble will
[119:07.00]Continue on
[119:08.00]For a long time
[119:09.00]Right like it's not
[119:10.00]It's not going to end
[119:11.00]Like this year
[119:12.00]Right like people
[119:13.00]People need AI
[119:14.00]Right like I think
[119:15.00]Everyone in this audience
[119:16.00]Would agree right like
[119:17.00]There's no
[119:18.00]There's no
[119:19.00]Like immediate
[119:20.00]Like end to the
[119:21.00]To the bubble
[119:22.00]Right
[119:23.00]What's next
[119:24.00]Why
[119:25.00]I think it's just
[119:26.00]Because the supply chain
[119:27.00]Expands so much
[119:28.00]Some companies
[119:29.00]Will continue to buy
[119:30.00]Like an
[119:31.00]Open AI
[119:32.00]Or meta
[119:33.00]Will continue to buy
[119:34.00]But then like
[119:35.00]All these random
[119:36.00]Startups will
[119:37.00]Or a lot of them
[119:38.00]Will not be able to
[119:39.00]Continue to buy
[119:40.00]So then
[119:41.00]That kind of leads to like
[119:42.00]They'll pause
[119:43.00]For a little bit
[119:44.00]Or like
[119:45.00]I think in 2018
[119:46.00]Right like
[119:47.00]Memory pricing was
[119:48.00]Extremely high
[119:49.00]Then all of a sudden
[119:50.00]Google, Microsoft
[119:51.00]And Amazon
[119:52.00]All agreed
[119:53.00]I don't you know
[119:54.00]They won't
[119:55.00]They won't say
[119:56.00]It's together
[119:57.00]It's a week
[119:58.00]To stop ordering memory
[119:59.00]And within like
[120:00.00]A month
[120:01.00]The price of memory
[120:02.00]Started tanking
[120:03.00]Like insane amounts
[120:04.00]Right and like
[120:05.00]People will claim
[120:06.00]You know all sorts of
[120:07.00]Reasons why
[120:08.00]That was timed
[120:09.00]Extremely well
[120:10.00]It was like very clear
[120:11.00]And people in the
[120:12.00]Financial markets
[120:13.00]Were able to make trades
[120:14.00]And everything
[120:15.00]People stopped buying
[120:16.00]And it's not like
[120:17.00]Their demand just dried up
[120:18.00]It's just like
[120:19.00]They had a little bit
[120:20.00]Of a demand slowdown
[120:21.00]And then they had enough
[120:22.00]Inventory that they could
[120:23.00]Like weather until
[120:24.00]Like prices tanked
[120:25.00]Because it's such
[120:26.00]Thank you very much
[120:27.00]Is it
[120:28.00]Hey everyone
[120:33.00]And so
[120:34.00]Today we have a special guest
[120:35.00]Millions from Capital One
[120:37.00]But I would tend to
[120:38.00]Like to introduce people
[120:39.00]With a bit of their background
[120:41.00]And then learn a little bit
[120:42.00]More about you
[120:43.00]On the personal side
[120:44.00]You call your PhD
[120:45.00]In a probabilistic framework
[120:47.00]From mapping audio
[120:48.00]Visual features to semantics
[120:49.00]I feel like that
[120:50.00]Is like the beginnings
[120:51.00]Of like a multimodal
[120:52.00]AI model in some sense
[120:54.00]Do you have any sort
[120:55.00]Factions on your PhD
[120:56.00]Vis this cut
[120:57.00]Thanks for having me
[120:59.00]And so
[121:00.00]Let me say this that
[121:01.00]It almost feels like
[121:02.00]Things go around in circles
[121:03.00]Right
[121:04.00]In research and development
[121:05.00]And so
[121:06.00]At the right time
[121:07.00]And the right place
[121:08.00]You kind of intersect
[121:09.00]Back with some of the topics
[121:11.00]And then
[121:12.00]Some other conditions
[121:13.00]That have happened
[121:14.00]Suddenly make
[121:15.00]A big difference
[121:16.00]Between that ticking off
[121:17.00]Vorses you know
[121:18.00]It may not be
[121:19.00]You know
[121:20.00]As intently pursued
[121:21.00]At any given point of time
[121:22.00]Right so
[121:23.00]I have been
[121:24.00]In AI for now
[121:25.00]Three decades
[121:26.00]You know
[121:27.00]You talked about
[121:28.00]My PhD thesis
[121:29.00]My bachelor's thesis
[121:30.00]Was on implementing
[121:32.00]Neural networks
[121:33.00]On India's
[121:34.00]You know
[121:35.00]Homegrown supercomputers
[121:36.00]Back then
[121:37.00]And so
[121:38.00]You know
[121:39.00]This whole notion of
[121:40.00]Message passing
[121:41.00]And distributed computing
[121:42.00]And computing weights
[121:44.00]You know
[121:45.00]And then bringing
[121:46.00]All of them back
[121:47.00]Distributing
[121:48.00]The computations
[121:49.00]Of the neural network
[121:50.00]You know
[121:51.00]Forward pass
[121:52.00]Those things
[121:53.00]For what we used to do
[121:54.00]You know
[121:55.00]For that
[121:56.00]Particular supercomputing
[121:57.00]Architecture
[121:58.00]We had back then
[121:59.00]And then
[122:00.00]My PhD
[122:01.00]Of course was
[122:02.00]How to understand
[122:03.00]What's going on
[122:04.00]In a video
[122:05.00]Right and use
[122:06.00]Multi-modal cues
[122:07.00]To your point
[122:08.00]You know
[122:09.00]What has happened
[122:10.00]In the last couple of
[122:11.00]Decades
[122:12.00]One we have
[122:13.00]Tremendous amount
[122:14.00]Of data explosion
[122:15.00]So when I was doing
[122:16.00]My PhD
[122:17.00]I used to
[122:18.00]Actually go
[122:19.00]To blockbuster
[122:20.00]And rent
[122:21.00]Explosions
[122:22.00]And how do you
[122:23.00]Actually
[122:24.00]Models in the audio stream
[122:25.00]And the visual stream
[122:26.00]For something
[122:27.00]Like an explosion
[122:28.00]So I remember
[122:29.00]Going and doing
[122:30.00]All this digitization
[122:31.00]Of tape and then
[122:32.00]Cleaning up the data
[122:33.00]And then
[122:34.00]You know
[122:35.00]Having some kind
[122:36.00]Of a labeling tool
[122:37.00]That I actually cooked up
[122:38.00]And having a spouse
[122:40.00]Of a friend of mine
[122:41.00]To do the labeling
[122:42.00]For me
[122:43.00]So look at
[122:44.00]Where we were back then
[122:45.00]And now you have
[122:46.00]Scale.ai
[122:47.00]That basically goes
[122:48.00]And does labeling
[122:49.00]For a lot of these models
[122:50.00]And so forth
[122:51.00]So scale
[122:52.00]Of data has changed
[122:53.00]That's one
[122:54.00]The second thing
[122:55.00]That has changed
[122:56.00]Is we were looking
[122:57.00]At computing
[122:58.00]Architectures
[122:59.00]That were
[123:00.00]Much much
[123:01.00]Much less rich
[123:02.00]In terms of what
[123:03.00]We had back then
[123:04.00]And the 2012
[123:06.00]Breakthrough
[123:07.00]By Hinton and his students
[123:09.00]And using GPUs
[123:11.00]Really when
[123:12.00]The ImageNet competition
[123:14.00]Really helped
[123:15.00]Take this field off
[123:16.00]In a completely
[123:17.00]New direction
[123:18.00]And at
[123:19.00]Very very large scale
[123:20.00]And the third thing
[123:21.00]Is of course
[123:22.00]The GPU computing
[123:23.00]Back then
[123:24.00]I did not have access
[123:25.00]When I was doing my PhD
[123:26.00]To some of the amazing things
[123:28.00]That Nvidia hadn't yet built
[123:30.00]And so
[123:31.00]It's I think really
[123:32.00]The confluence of those three things
[123:34.00]Which make all the difference
[123:35.00]Between a lot of this research
[123:36.00]That happened
[123:37.00]In the late 90s
[123:38.00]And what's happening
[123:39.00]Between the 2010
[123:41.00]To now
[123:42.00]Kind of time frame
[123:43.00]But if you look at the intent
[123:45.00]The intent was the same
[123:47.00]What's in the video
[123:48.00]How do we understand
[123:49.00]The multimodal cues
[123:50.00]You know
[123:51.00]That come together
[123:52.00]To give us
[123:53.00]That semantic understanding
[123:54.00]Of what's in the video
[123:55.00]And so
[123:56.00]To that extent
[123:57.00]The problems
[123:58.00]We were trying
[123:59.00]To solve the same
[124:00.00]But the tools
[124:01.00]That we have now
[124:02.00]Are amazingly
[124:03.00]Amazingly different
[124:04.00]And amazingly
[124:05.00]More powerful
[124:06.00]Are there any
[124:07.00]Maybe research approaches
[124:08.00]Or ML patterns
[124:10.00]That you tried
[124:11.00]That didn't work
[124:12.00]That you think
[124:13.00]Will work today
[124:14.00]Would you abuse
[124:15.00]That people haven't tried
[124:16.00]I think
[124:17.00]There are many people
[124:18.00]That have done serious
[124:19.00]ML research before
[124:20.00]The GPU era
[124:21.00]I would say
[124:22.00]If you think about
[124:23.00]All the ML researchers
[124:24.00]Working today
[124:25.00]Most of them are post-GPU
[124:26.00]Any like story
[124:27.00]That you remember
[124:28.00]That you were like
[124:29.00]Oh, this seems really promising
[124:30.00]But like
[124:31.00]There wasn't enough compute
[124:32.00]Or anything like that
[124:33.00]The whole concept
[124:34.00]Of modeling context
[124:35.00]Right
[124:36.00]So my thesis was about
[124:37.00]Not
[124:38.00]How do you just
[124:39.00]Detect isolated things
[124:40.00]In a video
[124:41.00]Right
[124:42.00]Like this is a car
[124:43.00]That's an explosion
[124:44.00]And so on and so forth
[124:45.00]If I see these
[124:46.00]End things together
[124:47.00]Do they actually
[124:48.00]Contextually make sense
[124:49.00]Like do I see
[124:50.00]The sky
[124:51.00]About the land
[124:52.00]And if I do that
[124:53.00]Then I have a higher confidence
[124:55.00]That this indeed is sky
[124:56.00]And that indeed is land
[124:57.00]Right
[124:58.00]And so on and so forth
[124:59.00]So when we were trying
[125:00.00]To model that context
[125:01.00]We had extremely limited
[125:03.00]Libling
[125:04.00]And extremely limited
[125:05.00]Corpus
[125:06.00]In terms of how we could do it
[125:08.00]Now when I think back
[125:10.00]I think that
[125:11.00]What better way to describe
[125:12.00]Context than a multimodal LLM
[125:14.00]Right
[125:15.00]Which strain on
[125:16.00]As much of the data
[125:18.00]As there is on the internet
[125:20.00]In terms of multimodality
[125:22.00]And that to me
[125:23.00]Would have been an amazing thing
[125:24.00]For me to have back then
[125:26.00]So I would say
[125:27.00]How to model context
[125:28.00]Is a problem
[125:29.00]That is going to be evergreen
[125:30.00]It never goes out of fashion
[125:32.00]But how we are able
[125:33.00]To do it now
[125:34.00] Versus then
[125:35.00]I see as
[125:36.00]One of the steps towards
[125:38.00]Truly understanding
[125:39.00]You know what's happening
[125:40.00]Right in the multiple modalities
[125:42.00]The other part
[125:43.00]Is reasoning
[125:44.00]Right
[125:45.00]I think we are still
[125:46.00]In the very early innings
[125:47.00]Of reasoning
[125:48.00]You know we see this
[125:49.00]Interesting evolution
[125:51.00]Of you know
[125:52.00]How do you actually
[125:53.00]Build a model of the world
[125:55.00]And Yanlacan's work
[125:56.00]You know very interesting
[125:58.00]In this sense to me
[125:59.00]He has been talking
[126:00.00]About it for a while
[126:01.00]Right so
[126:02.00]But now I think
[126:03.00]We are getting to a point
[126:04.00]Where a lot of those pieces
[126:05.00]May start coming together
[126:07.00]And I think solving
[126:09.00]That reasoning piece
[126:10.00]Is a very very critical step
[126:12.00]Before we can actually
[126:14.00]Build
[126:15.00]Truly intelligent machines
[126:17.00]And do you have any
[126:18.00]Intuition on the part
[126:19.00]The video
[126:20.00]Is going to play in it
[126:21.00]Because a lot of Yan's
[126:22.00]Also talking points
[126:23.00]Are around you know
[126:24.00]With vjapa
[126:25.00]And some of those models
[126:26.00]If you show
[126:27.00]What's going to happen next
[126:28.00]Like that's part
[126:29.00]Of a war model
[126:30.00]Like it's video going to be
[126:32.00]Like a big part in it
[126:33.00]Like do we need to
[126:34.00]Get there to actually
[126:35.00]Get a real war model
[126:36.00]Like do you think text
[126:37.00]Is enough to
[126:38.00]Get a good shot at it
[126:39.00]And maybe biased
[126:40.00]In answering that question
[126:41.00]Given that
[126:42.00]I you know
[126:43.00]Cut my teeth
[126:44.00]In multi modality
[126:45.00]And given that
[126:46.00]The video modality
[126:47.00]In general
[126:48.00]Right is a lot
[126:49.00]More challenging
[126:50.00]You know whether
[126:51.00]It's just because
[126:52.00]Of the sheer size of data
[126:53.00]Right in terms of
[126:54.00]The number of pixels
[126:55.00]That you need to process
[126:56.00]Whether it is because
[126:57.00]You are actually
[126:58.00]Capturing the real world
[127:00.00]Which tends to
[127:01.00]Be far more complex
[127:02.00]You know in the
[127:03.00]Case of language
[127:04.00]You know humans
[127:05.00]Over millennia
[127:07.00]Have evolved this
[127:09.00]Longly concise
[127:11.00]Codebook
[127:12.00]Of how to describe
[127:13.00]Things
[127:14.00]So there is a humongous
[127:15.00]Amount of abstraction
[127:18.00]And rationalization
[127:20.00]And concise definition
[127:22.00]That has gone on
[127:23.00]In how languages
[127:25.00]Evolved
[127:26.00]And so you know
[127:27.00]The number of words
[127:28.00]In the vocabulary
[127:29.00]Of a language
[127:30.00]When you look at that
[127:31.00]We are able to
[127:32.00]Tell beautiful stories
[127:33.00]Right with
[127:34.00]Just those many
[127:36.00]You know words
[127:37.00]But if you look at
[127:38.00]In the real world
[127:39.00]If you look at
[127:40.00]Capturing that
[127:41.00]Right whether it is
[127:42.00]Through
[127:43.00]The eyes of a robot
[127:44.00]As it's looking
[127:45.00]You know and trying
[127:46.00]To help you around
[127:47.00]In a room setting
[127:48.00]Or a building setting
[127:49.00]Right or whether
[127:50.00]It's the traffic
[127:51.00]Right which
[127:52.00]Vacues have to look at
[127:53.00]When they're
[127:54.00]Driving on the roads
[127:55.00]The amount of
[127:56.00]Variability
[127:57.00]The amount of
[127:58.00]Distortion of signal
[127:59.00]That comes with
[128:00.00]That right
[128:01.00]Is just remarkably
[128:03.00]Difficult
[128:04.00]And remarkably rich
[128:05.00]To really analyze
[128:06.00]So I would say
[128:07.00]Alicio
[128:08.00]It's both
[128:09.00]I feel that
[128:10.00]The video modality
[128:11.00]Has to be understood
[128:13.00]For a true understanding
[128:14.00]Of the world model
[128:15.00]I would also say
[128:16.00]That's harder
[128:17.00]In some sense
[128:19.00]Because of its inherent complexity
[128:20.00]Then the language part
[128:22.00]Which there's already
[128:24.00]Consize representation
[128:25.00]That people have come up with
[128:26.00]Awesome
[128:27.00]Sorry Sean
[128:28.00]I know we hijacked the intro
[128:30.00]But it was a good rabbit hole
[128:32.00]To go into
[128:33.00]No it's okay
[128:34.00]I also just like
[128:35.00]Stunned by
[128:36.00]Background and history
[128:37.00]That millennium is
[128:38.00]Bringing to
[128:39.00]AI you know
[128:40.00]I guess the speed run through
[128:41.00]In the resume
[128:42.00]You know 14 years
[128:43.00]At IBM
[128:44.00]Finally ending
[128:45.00]As chief scientist at
[128:46.00]IBM research
[128:47.00]And then since
[128:48.00]Cisco
[128:49.00]With cognitive systems
[128:50.00]And CTO
[128:51.00]Of metropolis at NVIDIA
[128:52.00]What should people know
[128:53.00]About like
[128:54.00]How your
[128:55.00]Interest in your trajectory
[128:56.00]Has progressed through
[128:57.00]Your career
[128:58.00]You know when I reflect
[128:59.00]Part of what I see
[129:00.00]Is there is a constant
[129:02.00]And the constant is
[129:04.00]How do you actually
[129:05.00]AI work
[129:06.00]For you know
[129:07.00]Name your favorite
[129:08.00]Problem
[129:09.00]That favorite problem
[129:10.00]Changes maybe
[129:11.00]From decade to decade
[129:12.00]Or in the context of
[129:13.00]Even my state
[129:14.00]IBM research
[129:15.00]But it's always been
[129:16.00]How do we build
[129:17.00]AI solutions
[129:18.00]AI platforms
[129:19.00]That solve
[129:20.00]Real world problems
[129:21.00]So when
[129:22.00]We started out
[129:23.00]We were the first
[129:24.00]Video understanding platform
[129:26.00]That we built at
[129:27.00]IBM research right
[129:28.00]We actually got a Wall Street
[129:29.00]Multimedia
[129:30.00]Award for it
[129:31.00]Innovation Award for it
[129:32.00]In the early 2000s
[129:34.00]We helped with
[129:35.00]Setting up the benchmark
[129:36.00]That's known as
[129:38.00]Trackweed
[129:39.00]Which again
[129:40.00]To at least your
[129:41.00]Point people
[129:42.00]Only know ImageNet
[129:43.00]And thereafter
[129:44.00]Before ImageNet
[129:45.00]There was Trackweed
[129:46.00]And so
[129:47.00]The first decade
[129:48.00]Shan was really about
[129:49.00]You know how do we
[129:50.00]Understand what's in the video
[129:51.00]And how can then
[129:52.00]We turn that into
[129:53.00]Meaningful use of
[129:55.00]AI technology
[129:56.00]For media companies
[129:58.00]You know broadcasting
[129:59.00]Corporations
[130:00.00]And so on and so forth
[130:01.00]Right the business to
[130:02.00]Business kind of setting
[130:03.00]Of course
[130:04.00]Because we were at
[130:05.00]IBM
[130:06.00]We did not focus
[130:07.00]On the consumer
[130:08.00]And then
[130:09.00]You know YouTube
[130:10.00]Happened right
[130:11.00]Here in the valley
[130:12.00]And then
[130:13.00]You see the
[130:14.00]Explosive content
[130:15.00]And applications
[130:16.00]Of AI to
[130:17.00]You know those
[130:18.00]Kind of video
[130:19.00]Understanding problems
[130:20.00]Then it was
[130:21.00]How do we actually
[130:22.00]Make sense
[130:23.00]Out of our
[130:24.00]IoT world
[130:25.00]So when you have
[130:26.00]Signals coming from
[130:27.00]Sensors everywhere
[130:28.00]You know whether
[130:29.00]There are sensors
[130:30.00]Embedded in your
[130:31.00]Bridges
[130:32.00]Sensory information
[130:33.00]To make
[130:34.00]Good decisions
[130:35.00]To a
[130:36.00]You know observe
[130:37.00]The environments
[130:38.00]And we optimize
[130:39.00]Those environments
[130:40.00]Right and so a large
[130:41.00]Part of my
[130:42.00]Second half of
[130:43.00]State IBM research
[130:44.00]Was to come up with
[130:45.00]What is this
[130:46.00]Research around
[130:48.00]Smart cities
[130:49.00]Smarter planet
[130:50.00]And that
[130:51.00]Actually became
[130:52.00]An AI platform
[130:53.00]For helping
[130:55.00]Optimized traffic
[130:56.00]Right one of
[130:57.00]The proudest
[130:58.00]Things that
[130:59.00]I really
[131:00.00]Foundly remember
[131:01.00]Use the data
[131:02.00]From you know
[131:03.00]Teleco data sources
[131:04.00]To understand
[131:05.00]How people move
[131:06.00]In a city
[131:07.00]And then use
[131:08.00]That information
[131:09.00]To build
[131:10.00]Optimal planning
[131:11.00]Right whether
[131:12.00]It's for bus outs
[131:13.00]Whether it's for metro
[131:14.00]Right we did
[131:15.00]Some amazing work
[131:16.00]In Istanbul
[131:17.00]For example
[131:18.00]You know
[131:19.00]Completely different
[131:20.00]Scale
[131:21.00]And then
[131:22.00]The same
[131:23.00]Kind of platform
[131:24.00]We applied
[131:25.00]To helping
[131:26.00]Cities in
[131:27.00]American Midwest
[131:28.00]Like Dubuque
[131:29.00]Optimized their
[131:30.00]Dubuque
[131:31.00]And then
[131:32.00]And then
[131:33.00]The same
[131:34.00]Like Dubuque
[131:35.00]And then
[131:36.00]The same
[131:37.00]Like Dubuque
[131:38.00]And then
[131:39.00]The same
[131:40.00]Like Dubuque
[131:41.00]And then
[131:42.00]The same
[131:43.00]Like Dubuque
[131:44.00]And then
[131:45.00]The same
[131:46.00]Like Dubuque
[131:47.00]And then
[131:48.00]The same
[131:49.00]Like Dubuque
[131:50.00]And then
[131:51.00]The same
[131:52.00]Like Dubuque
[131:53.00]And then
[131:54.00]The same
[131:55.00]Like Dubuque
[131:56.00]And then
[131:57.00]The same
[131:58.00]Like Dubuque
[131:59.00]And then
[132:00.00]And then
[132:01.00]The same
[132:02.00]Like Dubuque
[132:03.00]And then
[132:04.00]The same
[132:05.00]Like Dubuque
[132:06.00]And then
[132:07.00]The same
[132:08.00]Like Dubuque
[132:09.00]And then
[132:10.00]The same
[132:11.00]Like Dubuque
[132:12.00]And then
[132:13.00]The same
[132:14.00]Like Dubuque
[132:15.00]And then
[132:16.00]The same
[132:17.00]Like Dubuque
[132:18.00]And then
[132:19.00]The same
[132:20.00]Like Dubuque
[132:21.00]And then
[132:22.00]The same
[132:23.00]Like Dubuque
[132:24.00]And then
[132:25.00]The same
[132:26.00]Like Dubuque
[132:27.00]And then
[132:28.00]The same
[132:29.00]Like Dubuque
[132:30.00]And then
[132:31.00]The same
[132:32.00]Like Dubuque
[132:33.00]And then
[132:34.00]The same
[132:35.00]Like Dubuque
[132:36.00]And then
[132:37.00]The same
[132:38.00]Like Dubuque
[132:39.00]And then
[132:40.00]The same
[132:41.00]Like Dubuque
[132:42.00]And then
[132:43.00]The same
[132:44.00]Like Dubuque
[132:45.00]And then
[132:46.00]The same
[132:47.00]Like Dubuque
[132:48.00]And then
[132:49.00]The same
[132:50.00]Like Dubuque
[132:51.00]And then
[132:52.00]The same
[132:53.00]Like Dubuque
[132:54.00]And then
[132:55.00]The same
[132:56.00]Like Dubuque
[132:57.00]And then
[132:58.00]The same
[132:59.00]Like Dubuque
[133:00.00]And then
[133:01.00]The same
[133:02.00]Like Dubuque
[133:03.00]And then
[133:04.00]The same
[133:05.00]Like Dubuque
[133:06.00]And then
[133:07.00]The same
[133:08.00]Like Dubuque
[133:09.00]And then
[133:10.00]The same
[133:11.00]Like Dubuque
[133:12.00]And then
[133:13.00]The same
[133:14.00]Like Dubuque
[133:15.00]And then
[133:16.00]The same
[133:17.00]Like Dubuque
[133:18.00]And then
[133:19.00]The same
[133:20.00]Like Dubuque
[133:21.00]And then
[133:22.00]The same
[133:23.00]Like Dubuque
[133:24.00]And then
[133:25.00]The same
[133:26.00]Like Dubuque
[133:27.00]And then
[133:28.00]The same
[133:29.00]Like Dubuque
[133:30.00]And then
[133:31.00]The same
[133:32.00]Like Dubuque
[133:33.00]And then
[133:34.00]The same
[133:35.00]Like Dubuque
[133:36.00]And then
[133:37.00]The same
[133:38.00]Like Dubuque
[133:39.00]And then
[133:40.00]The same
[133:41.00]Like Dubuque
[133:42.00]And then
[133:43.00]The same
[133:44.00]Like Dubuque
[133:45.00]And then
[133:46.00]The same
[133:47.00]Like Dubuque
[133:48.00]And then
[133:49.00]The same
[133:50.00]Like Dubuque
[133:51.00]And then
[133:52.00]The same
[133:53.00]Like Dubuque
[133:54.00]And then
[133:55.00]The same
[133:56.00]Like Dubuque
[133:57.00]And then
[133:58.00]The same
[133:59.00]Like Dubuque
[134:00.00]And then
[134:01.00]The same
[134:02.00]Like Dubuque
[134:03.00]And then
[134:04.00]The same
[134:05.00]Like Dubuque
[134:06.00]And then
[134:07.00]The same
[134:08.00]Like Dubuque
[134:09.00]And then
[134:10.00]The same
[134:11.00]Like Dubuque
[134:12.00]And then
[134:13.00]The same
[134:14.00]Like Dubuque
[134:15.00]And then
[134:16.00]The same
[134:17.00]Like Dubuque
[134:18.00]And then
[134:19.00]The same
[134:20.00]Like Dubuque
[134:21.00]And then
[134:22.00]The same
[134:23.00]Like Dubuque
[134:24.00]And then
[134:25.00]The same
[134:26.00]Like Dubuque
[134:27.00]And then
[134:28.00]The same
[134:29.00]Like Dubuque
[134:30.00]And then
[134:31.00]The same
[134:32.00]Like Dubuque
[134:33.00]And then
[134:34.00]The same
[134:35.00]Like Dubuque
[134:36.00]And then
[134:37.00]The same
[134:38.00]Like Dubuque
[134:39.00]And then
[134:40.00]The same
[134:41.00]Like Dubuque
[134:42.00]And then
[134:43.00]The same
[134:44.00]Like Dubuque
[134:45.00]And then
[134:46.00]The same
[134:47.00]Like Dubuque
[134:48.00]And then
[134:49.00]The same
[134:50.00]Like Dubuque
[134:51.00]And then
[134:52.00]The same
[134:53.00]Like Dubuque
[134:54.00]And then
[134:55.00]The same
[134:56.00]Like Dubuque
[134:57.00]And then
[134:58.00]The same
[134:59.00]Like Dubuque
[135:00.00]And then
[135:01.00]The same
[135:02.00]Like Dubuque
[135:03.00]And then
[135:04.00]The same
[135:05.00]Like Dubuque
[135:06.00]And then
[135:07.00]The same
[135:08.00]Like Dubuque
[135:09.00]And then
[135:10.00]The same
[135:11.00]Like Dubuque
[135:12.00]And then
[135:13.00]The same
[135:14.00]Like Dubuque
[135:15.00]And then
[135:16.00]The same
[135:17.00]Like Dubuque
[135:18.00]And then
[135:19.00]The same
[135:20.00]Like Dubuque
[135:21.00]And then
[135:22.00]The same
[135:23.00]Like Dubuque
[135:24.00]And then
[135:25.00]The same
[135:26.00]Like Dubuque
[135:27.00]And then
[135:28.00]The same
[135:29.00]Like Dubuque
[135:30.00]And then
[135:31.00]The same
[135:32.00]Like Dubuque
[135:33.00]And then
[135:34.00]The same
[135:35.00]Like Dubuque
[135:36.00]And then
[135:37.00]The same
[135:38.00]Like Dubuque
[135:39.00]And then
[135:40.00]The same
[135:41.00]Like Dubuque
[135:42.00]And then
[135:43.00]The same
[135:44.00]Like Dubuque
[135:45.00]And then
[135:46.00]The same
[135:47.00]Like Dubuque
[135:48.00]And then
[135:49.00]The same
[135:50.00]Like Dubuque
[135:51.00]And then
[135:52.00]The same
[135:53.00]Like Dubuque
[135:54.00]And then
[135:55.00]The same
[135:56.00]Like Dubuque
[135:57.00]And then
[135:58.00]The same
[135:59.00]Like Dubuque
[136:00.00]And then
[136:01.00]The same
[136:02.00]Like Dubuque
[136:03.00]And then
[136:04.00]The same
[136:05.00]Like Dubuque
[136:06.00]And then
[136:07.00]The same
[136:08.00]Like Dubuque
[136:09.00]And then
[136:10.00]The same
[136:11.00]Like Dubuque
[136:12.00]And then
[136:13.00]The same
[136:14.00]Like Dubuque
[136:15.00]And then
[136:16.00]The same
[136:17.00]Like Dubuque
[136:18.00]And then
[136:19.00]The same
[136:20.00]Like Dubuque
[136:21.00]And then
[136:22.00]The same
[136:23.00]Like Dubuque
[136:24.00]And then
[136:25.00]The same
[136:26.00]Like Dubuque
[136:27.00]And then
[136:28.00]The same
[136:29.00]Like Dubuque
[136:30.00]And then
[136:31.00]The same
[136:32.00]Like Dubuque
[136:33.00]And then
[136:34.00]The same
[136:35.00]Like Dubuque
[136:36.00]And then
[136:37.00]The same
[136:38.00]Like Dubuque
[136:39.00]And then
[136:40.00]The same
[136:41.00]Like Dubuque
[136:42.00]And then
[136:43.00]The same
[136:44.00]Like Dubuque
[136:45.00]And then
[136:46.00]The same
[136:47.00]Like Dubuque
[136:48.00]And then
[136:49.00]And then
[136:50.00]The same
[136:51.00]Like Dubuque
[136:52.00]And then
[136:53.00]The same
[136:54.00]Like Dubuque
[136:55.00]And then
[136:56.00]The same
[136:57.00]Like Dubuque
[136:58.00]And then
[136:59.00]The same
[137:00.00]Like Dubuque
[137:01.00]And then
[137:02.00]The same
[137:03.00]Like Dubuque
[137:04.00]And then
[137:05.00]The same
[137:06.00]Like Dubuque
[137:07.00]And then
[137:08.00]The same
[137:09.00]Like Dubuque
[137:10.00]And then
[137:11.00]The same
[137:12.00]Like Dubuque
[137:13.00]And then
[137:14.00]The same
[137:15.00]Like Dubuque
[137:16.00]And then
[137:17.00]The same
[137:18.00]Like Dubuque
[137:19.00]And then
[137:20.00]The same
[137:21.00]Like Dubuque
[137:22.00]And then
[137:23.00]The same
[137:24.00]Like Dubuque
[137:25.00]And then
[137:26.00]The same
[137:27.00]Like Dubuque
[137:28.00]And then
[137:29.00]The same
[137:30.00]Like Dubuque
[137:31.00]And then
[137:32.00]The same
[137:33.00]Like Dubuque
[137:34.00]And then
[137:35.00]The same
[137:36.00]Like Dubuque
[137:37.00]And then
[137:38.00]The same
[137:39.00]Like Dubuque
[137:40.00]And then
[137:41.00]The same
[137:42.00]Like Dubuque
[137:43.00]And then
[137:44.00]The same
[137:45.00]Like Dubuque
[137:46.00]And then
[137:47.00]The same
[137:48.00]Like Dubuque
[137:49.00]And then
[137:50.00]The same
[137:51.00]Like Dubuque
[137:52.00]And then
[137:53.00]The same
[137:54.00]Like Dubuque
[137:55.00]And then
[137:56.00]The same
[137:57.00]Like Dubuque
[137:58.00]And then
[137:59.00]The same
[138:00.00]Like Dubuque
[138:01.00]And then
[138:02.00]The same
[138:03.00]Like Dubuque
[138:04.00]And then
[138:05.00]The same
[138:06.00]Like Dubuque
[138:07.00]And then
[138:08.00]The same
[138:09.00]Like Dubuque
[138:10.00]And then
[138:11.00]The same
[138:12.00]Like Dubuque
[138:13.00]And then
[138:14.00]The same
[138:15.00]Like Dubuque
[138:16.00]And then
[138:17.00]The same
[138:18.00]Like Dubuque
[138:19.00]And then
[138:20.00]The same
[138:21.00]Like Dubuque
[138:22.00]And then
[138:23.00]The same
[138:24.00]Like Dubuque
[138:25.00]And then
[138:26.00]The same
[138:27.00]Like Dubuque
[138:28.00]And then
[138:29.00]The same
[138:30.00]Like Dubuque
[138:31.00]And then
[138:32.00]The same
[138:33.00]Like Dubuque
[138:34.00]And then
[138:35.00]The same
[138:36.00]Like Dubuque
[138:37.00]And then
[138:38.00]The same
[138:39.00]Like Dubuque
[138:40.00]And then
[138:41.00]The same
[138:42.00]Like Dubuque
[138:43.00]And then
[138:44.00]The same
[138:45.00]Like Dubuque
[138:46.00]And then
[138:47.00]The same
[138:48.00]Like Dubuque
[138:49.00]And then
[138:50.00]The same
[138:51.00]Like Dubuque
[138:52.00]And then
[138:53.00]The same
[138:54.00]Like Dubuque
[138:55.00]And then
[138:56.00]The same
[138:57.00]Like Dubuque
[138:58.00]And then
[138:59.00]The same
[139:00.00]Like Dubuque
[139:01.00]And then
[139:02.00]The same
[139:03.00]Like Dubuque
[139:04.00]And then
[139:05.00]The same
[139:06.00]Like Dubuque
[139:07.00]And then
[139:08.00]The same
[139:09.00]Like Dubuque
[139:10.00]And then
[139:11.00]The same
[139:12.00]Like Dubuque
[139:13.00]And then
[139:14.00]The same
[139:15.00]Like Dubuque
[139:16.00]And then
[139:17.00]The same
[139:18.00]Like Dubuque
[139:19.00]And then
[139:20.00]The same
[139:21.00]Like Dubuque
[139:22.00]And then
[139:23.00]The same
[139:24.00]Like Dubuque
[139:25.00]And then
[139:26.00]The same
[139:27.00]Like Dubuque
[139:28.00]And then
[139:29.00]The same
[139:30.00]Like Dubuque
[139:31.00]And then
[139:32.00]The same
[139:33.00]Like Dubuque
[139:34.00]And then
[139:35.00]The same
[139:36.00]Like Dubuque
[139:37.00]And then
[139:38.00]The same
[139:39.00]Like Dubuque
[139:40.00]And then
[139:41.00]The same
[139:42.00]Like Dubuque
[139:43.00]And then
[139:44.00]The same
[139:45.00]Like Dubuque
[139:46.00]And then
[139:47.00]The same
[139:48.00]Like Dubuque
[139:49.00]And then
[139:50.00]The same
[139:51.00]Like Dubuque
[139:52.00]And then
[139:53.00]The same
[139:54.00]Like Dubuque
[139:55.00]And then
[139:56.00]The same
[139:57.00]Like Dubuque
[139:58.00]And then
[139:59.00]The same
[140:00.00]Like Dubuque
[140:01.00]And then
[140:02.00]The same
[140:03.00]Like Dubuque
[140:04.00]And then
[140:05.00]The same
[140:06.00]Like Dubuque
[140:07.00]And then
[140:08.00]The same
[140:09.00]Like Dubuque
[140:10.00]And then
[140:11.00]The same
[140:12.00]Like Dubuque
[140:13.00]And then
[140:14.00]The same
[140:15.00]Like Dubuque
[140:16.00]And then
[140:17.00]The same
[140:18.00]Like Dubuque
[140:19.00]And then
[140:20.00]The same
[140:21.00]Like Dubuque
[140:22.00]And then
[140:23.00]The same
[140:24.00]Like Dubuque
[140:25.00]And then
[140:26.00]The same
[140:27.00]Like Dubuque
[140:28.00]And then
[140:29.00]The same
[140:30.00]Like Dubuque
[140:31.00]And then
[140:32.00]The same
[140:33.00]Like Dubuque
[140:34.00]And then
[140:35.00]The same
[140:36.00]Like Dubuque
[140:37.00]And then
[140:38.00]The same
[140:39.00]Like Dubuque
[140:40.00]And then
[140:41.00]The same
[140:42.00]Like Dubuque
[140:43.00]And then
[140:44.00]The same
[140:45.00]Like Dubuque
[140:46.00]And then
[140:47.00]The same
[140:48.00]Like Dubuque
[140:49.00]And then
[140:50.00]The same
[140:51.00]Like Dubuque
[140:52.00]And then
[140:53.00]The same
[140:54.00]Like Dubuque
[140:55.00]And then
[140:56.00]The same
[140:57.00]Like Dubuque
[140:58.00]And then
[140:59.00]The same
[141:00.00]Like Dubuque
[141:01.00]And then
[141:02.00]The same
[141:03.00]Like Dubuque
[141:04.00]And then
[141:05.00]The same
[141:06.00]Like Dubuque
[141:07.00]And then
[141:08.00]The same
[141:09.00]Like Dubuque
[141:10.00]And then
[141:11.00]The same
[141:12.00]Like Dubuque
[141:13.00]And then
[141:14.00]The same
[141:15.00]Like Dubuque
[141:16.00]And then
[141:17.00]The same
[141:18.00]Like Dubuque
[141:19.00]And then
[141:20.00]The same
[141:21.00]Like Dubuque
[141:22.00]And then
[141:23.00]The same
[141:24.00]Like Dubuque
[141:25.00]And then
[141:26.00]The same
[141:27.00]Like Dubuque
[141:28.00]And then
[141:29.00]The same
[141:30.00]Like Dubuque
[141:31.00]And then
[141:32.00]The same
[141:33.00]Like Dubuque
[141:34.00]And then
[141:35.00]The same
[141:36.00]Like Dubuque
[141:37.00]And then
[141:38.00]The same
[141:39.00]Like Dubuque
[141:40.00]And then
[141:41.00]The same
[141:42.00]Like Dubuque
[141:43.00]And then
[141:44.00]The same
[141:45.00]Like Dubuque
[141:46.00]And then
[141:47.00]The same
[141:48.00]Like Dubuque
[141:49.00]And then
[141:50.00]The same
[141:51.00]Like Dubuque
[141:52.00]And then
[141:53.00]The same
[141:54.00]Like Dubuque
[141:55.00]And then
[141:56.00]The same
[141:57.00]Like Dubuque
[141:58.00]And then
[141:59.00]The same
[142:00.00]Like Dubuque
[142:01.00]And then
[142:02.00]The same
[142:03.00]Like Dubuque
[142:04.00]And then
[142:05.00]The same
[142:06.00]Like Dubuque
[142:07.00]And then
[142:08.00]The same
[142:09.00]Like Dubuque
[142:10.00]And then
[142:11.00]The same
[142:12.00]Like Dubuque
[142:13.00]And then
[142:14.00]The same
[142:15.00]Like Dubuque
[142:16.00]And then
[142:17.00]The same
[142:18.00]Like Dubuque
[142:19.00]And then
[142:20.00]The same
[142:21.00]Like Dubuque
[142:22.00]And then
[142:23.00]The same
[142:24.00]Like Dubuque
[142:25.00]And then
[142:26.00]The same
[142:27.00]Like Dubuque
[142:28.00]And then
[142:29.00]The same
[142:30.00]Like Dubuque
[142:31.00]And then
[142:32.00]The same
[142:33.00]Like Dubuque
[142:34.00]And then
[142:35.00]The same
[142:36.00]Like Dubuque
[142:37.00]And then
[142:38.00]The same
[142:39.00]Like Dubuque
[142:40.00]And then
[142:41.00]The same
[142:42.00]Like Dubuque
[142:43.00]And then
[142:44.00]The same
[142:45.00]Like Dubuque
[142:46.00]And then
[142:47.00]The same
[142:48.00]Like Dubuque
[142:49.00]And then
[142:50.00]The same
[142:51.00]Like Dubuque
[142:52.00]And then
[142:53.00]The same
[142:54.00]Like Dubuque
[142:55.00]And then
[142:56.00]The same
[142:57.00]Like Dubuque
[142:58.00]And then
[142:59.00]The same
[143:00.00]Like Dubuque
[143:01.00]And then
[143:02.00]The same
[143:03.00]Like Dubuque
[143:04.00]And then
[143:05.00]The same
[143:06.00]Like Dubuque
[143:07.00]And then
[143:08.00]The same
[143:09.00]Like Dubuque
[143:10.00]And then
[143:11.00]The same
[143:12.00]Like Dubuque
[143:13.00]And then
[143:14.00]The same
[143:15.00]Like Dubuque
[143:16.00]And then
[143:17.00]The same
[143:18.00]Like Dubuque
[143:19.00]And then
[143:20.00]The same
[143:21.00]Like Dubuque
[143:22.00]And then
[143:23.00]The same
[143:24.00]Like Dubuque
[143:25.00]And then
[143:26.00]The same
[143:27.00]Like Dubuque
[143:28.00]And then
[143:29.00]The same
[143:30.00]Like Dubuque
[143:31.00]And then
[143:32.00]The same
[143:33.00]Like Dubuque
[143:34.00]And then
[143:35.00]The same
[143:36.00]Like Dubuque
[143:37.00]And then
[143:38.00]The same
[143:39.00]Like Dubuque
[143:40.00]And then
[143:41.00]The same
[143:42.00]Like Dubuque
[143:43.00]And then
[143:44.00]The same
[143:45.00]Like Dubuque
[143:46.00]And then
[143:47.00]The same
[143:48.00]Like Dubuque
[143:49.00]And then
[143:50.00]The same
[143:51.00]Like Dubuque
[143:52.00]And then
[143:53.00]The same
[143:54.00]Like Dubuque
[143:55.00]And then
[143:56.00]The same
[143:57.00]Like Dubuque
[143:58.00]And then
[143:59.00]The same
[144:00.00]Like Dubuque
[144:01.00]And then
[144:02.00]The same
[144:03.00]Like Dubuque
[144:04.00]And then
[144:05.00]The same
[144:06.00]Like Dubuque
[144:07.00]And then
[144:08.00]The same
[144:09.00]Like Dubuque
[144:10.00]And then
[144:11.00]The same
[144:12.00]Like Dubuque
[144:13.00]And then
[144:14.00]The same
[144:15.00]Like Dubuque
[144:16.00]And then
[144:17.00]The same
[144:18.00]Like Dubuque
[144:19.00]And then
[144:20.00]The same
[144:21.00]Like Dubuque
[144:22.00]And then
[144:23.00]The same
[144:24.00]Like Dubuque
[144:25.00]And then
[144:26.00]The same
[144:27.00]Like Dubuque
[144:28.00]And then
[144:29.00]The same
[144:30.00]Like Dubuque
[144:31.00]And then
[144:32.00]The same
[144:33.00]Like Dubuque
[144:34.00]And then
[144:35.00]The same
[144:36.00]Like Dubuque
[144:37.00]And then
[144:38.00]The same
[144:39.00]Like Dubuque
[144:40.00]And then
[144:41.00]The same
[144:42.00]Like Dubuque
[144:43.00]And then
[144:44.00]The same
[144:45.00]Like Dubuque
[144:46.00]And then
[144:47.00]The same
[144:48.00]Like Dubuque
[144:49.00]And then
[144:50.00]The same
[144:51.00]Like Dubuque
[144:52.00]And then
[144:53.00]The same
[144:54.00]Like Dubuque
[144:55.00]And then
[144:56.00]The same
[144:57.00]Like Dubuque
[144:58.00]And then
[144:59.00]The same
[145:00.00]Like Dubuque
[145:01.00]And then
[145:02.00]The same
[145:03.00]Like Dubuque
[145:04.00]And then
[145:05.00]The same
[145:06.00]Like Dubuque
[145:07.00]And then
[145:08.00]The same
[145:09.00]Like Dubuque
[145:10.00]And then
[145:11.00]The same
[145:12.00]Like Dubuque
[145:13.00]And then
[145:14.00]The same
[145:15.00]Like Dubuque
[145:16.00]And then
[145:17.00]The same
[145:18.00]Like Dubuque
[145:19.00]And then
[145:20.00]The same
[145:21.00]Like Dubuque
[145:22.00]And then
[145:23.00]The same
[145:24.00]Like Dubuque
[145:25.00]And then
[145:26.00]The same
[145:27.00]Like Dubuque
[145:28.00]And then
[145:29.00]The same
[145:30.00]Like Dubuque
[145:31.00]And then
[145:32.00]The same
[145:33.00]Like Dubuque
[145:34.00]And then
[145:35.00]The same
[145:36.00]Like Dubuque
[145:37.00]And then
[145:38.00]The same
[145:39.00]Like Dubuque
[145:40.00]And then
[145:41.00]The same
[145:42.00]Like Dubuque
[145:43.00]And then
[145:44.00]The same
[145:45.00]Like Dubuque
[145:46.00]And then
[145:47.00]The same
[145:48.00]Like Dubuque
[145:49.00]And then
[145:50.00]The same
[145:51.00]Like Dubuque
[145:52.00]And then
[145:53.00]The same
[145:54.00]Like Dubuque
[145:55.00]And then
[145:56.00]The same
[145:57.00]Like Dubuque
[145:58.00]And then
[145:59.00]The same
[146:00.00]Like Dubuque
[146:01.00]And then
[146:02.00]The same
[146:03.00]Like Dubuque
[146:04.00]And then
[146:05.00]The same
[146:06.00]Like Dubuque
[146:07.00]And then
[146:08.00]The same
[146:09.00]Like Dubuque
[146:10.00]And then
[146:11.00]The same
[146:12.00]Like Dubuque
[146:13.00]And then
[146:14.00]The same
[146:15.00]Like Dubuque
[146:16.00]And then
[146:17.00]The same
[146:18.00]Like Dubuque
[146:19.00]And then
[146:20.00]The same
[146:21.00]Like Dubuque
[146:22.00]And then
[146:23.00]The same
[146:24.00]Like Dubuque
[146:25.00]And then
[146:26.00]The same
[146:27.00]Like Dubuque
[146:28.00]And then
[146:29.00]The same
[146:30.00]Like Dubuque
[146:31.00]And then
[146:32.00]The same
[146:33.00]Like Dubuque
[146:34.00]And then
[146:35.00]The same
[146:36.00]Like Dubuque
[146:37.00]And then
[146:38.00]The same
[146:39.00]Like Dubuque
[146:40.00]And then
[146:41.00]The same
[146:42.00]Like Dubuque
[146:43.00]And then
[146:44.00]The same
[146:45.00]Like Dubuque
[146:46.00]And then
[146:47.00]The same
[146:48.00]Like Dubuque
[146:49.00]And then
[146:50.00]The same
[146:51.00]Like Dubuque
[146:52.00]And then
[146:53.00]The same
[146:54.00]Like Dubuque
[146:55.00]And then
[146:56.00]The same
[146:57.00]Like Dubuque
[146:58.00]And then
[146:59.00]The same
[147:00.00]Like Dubuque
[147:01.00]And then
[147:02.00]The same
[147:03.00]Like Dubuque
[147:04.00]And then
[147:05.00]The same
[147:06.00]Like Dubuque
[147:07.00]And then
[147:08.00]The same
[147:09.00]Like Dubuque
[147:10.00]And then
[147:11.00]The same
[147:12.00]Like Dubuque
[147:13.00]And then
[147:14.00]The same
[147:15.00]Like Dubuque
[147:16.00]And then
[147:17.00]The same
[147:18.00]Like Dubuque
[147:19.00]And then
[147:20.00]The same
[147:21.00]Like Dubuque
[147:22.00]And then
[147:23.00]The same
[147:24.00]Like Dubuque
[147:25.00]And then
[147:26.00]The same
[147:27.00]Like Dubuque
[147:28.00]And then
[147:29.00]The same
[147:30.00]Like Dubuque
[147:31.00]And then
[147:32.00]The same
[147:33.00]Like Dubuque
[147:34.00]And then
[147:35.00]The same
[147:36.00]Like Dubuque
[147:37.00]And then
[147:38.00]The same
[147:39.00]Like Dubuque
[147:40.00]And then
[147:41.00]The same
[147:42.00]Like Dubuque
[147:43.00]And then
[147:44.00]The same
[147:45.00]Like Dubuque
[147:46.00]And then
[147:47.00]The same
[147:48.00]Like Dubuque
[147:49.00]And then
[147:50.00]The same
[147:51.00]Like Dubuque
[147:52.00]And then
[147:53.00]The same
[147:54.00]Like Dubuque
[147:55.00]And then
[147:56.00]The same
[147:57.00]Like Dubuque
[147:58.00]And then
[147:59.00]The same
[148:00.00]Like Dubuque
[148:01.00]And then
[148:02.00]The same
[148:03.00]Like Dubuque
[148:04.00]And then
[148:05.00]The same
[148:06.00]Like Dubuque
[148:07.00]And then
[148:08.00]The same
[148:09.00]Like Dubuque
[148:10.00]And then
[148:11.00]The same
[148:12.00]Like Dubuque
[148:13.00]And then
[148:14.00]The same
[148:15.00]Like Dubuque
[148:16.00]And then
[148:17.00]The same
[148:18.00]Like Dubuque
[148:19.00]And then
[148:20.00]The same
[148:21.00]Like Dubuque
[148:22.00]And then
[148:23.00]The same
[148:24.00]Like Dubuque
[148:25.00]And then
[148:26.00]The same
[148:27.00]Like Dubuque
[148:28.00]And then
[148:29.00]The same
[148:30.00]Like Dubuque
[148:31.00]And then
[148:32.00]The same
[148:33.00]Like Dubuque
[148:34.00]And then
[148:35.00]The same
[148:36.00]Like Dubuque
[148:37.00]And then
[148:38.00]The same
[148:39.00]Like Dubuque
[148:40.00]And then
[148:41.00]The same
[148:42.00]Like Dubuque
[148:43.00]And then
[148:44.00]The same
[148:45.00]Like Dubuque
[148:46.00]And then
[148:47.00]The same
[148:48.00]Like Dubuque
[148:49.00]And then
[148:50.00]The same
[148:51.00]Like Dubuque
[148:52.00]And then
[148:53.00]The same
[148:54.00]Like Dubuque
[148:55.00]And then
[148:56.00]The same
[148:57.00]Like Dubuque
[148:58.00]And then
[148:59.00]The same
[149:00.00]Like Dubuque
[149:01.00]And then
[149:02.00]The same
[149:03.00]Like Dubuque
[149:04.00]And then
[149:05.00]The same
[149:06.00]Like Dubuque
[149:07.00]And then
[149:08.00]The same
[149:09.00]Like Dubuque
[149:10.00]And then
[149:11.00]The same
[149:12.00]Like Dubuque
[149:13.00]And then
[149:14.00]The same
[149:15.00]Like Dubuque
[149:16.00]And then
[149:17.00]The same
[149:18.00]Like Dubuque
[149:19.00]And then
[149:20.00]The same
[149:21.00]Like Dubuque
[149:22.00]And then
[149:23.00]The same
[149:24.00]Like Dubuque
[149:25.00]And then
[149:26.00]The same
[149:27.00]Like Dubuque
[149:28.00]And then
[149:29.00]The same
[149:30.00]Like Dubuque
[149:31.00]And then
[149:32.00]The same
[149:33.00]Like Dubuque
[149:34.00]And then
[149:35.00]The same
[149:36.00]Like Dubuque
[149:37.00]And then
[149:38.00]The same
[149:39.00]Like Dubuque
[149:40.00]And then
[149:41.00]The same
[149:42.00]Like Dubuque
[149:43.00]And then
[149:44.00]The same
[149:45.00]Like Dubuque
[149:46.00]And then
[149:47.00]The same
[149:48.00]Like Dubuque
[149:49.00]And then
[149:50.00]The same
[149:51.00]Like Dubuque
[149:52.00]And then
[149:53.00]The same
[149:54.00]Like Dubuque
[149:55.00]And then
[149:56.00]The same
[149:57.00]Like Dubuque
[149:58.00]And then
[149:59.00]The same
[150:00.00]Like Dubuque
[150:01.00]And then
[150:02.00]The same
[150:03.00]Like Dubuque
[150:04.00]And then
[150:05.00]The same
[150:06.00]Like Dubuque
[150:07.00]And then
[150:08.00]The same
[150:09.00]Like Dubuque
[150:10.00]And then
[150:11.00]The same
[150:12.00]Like Dubuque
[150:13.00]And then
[150:14.00]The same
[150:15.00]Like Dubuque
[150:16.00]And then
[150:17.00]The same
[150:18.00]Like Dubuque
[150:19.00]And then
[150:20.00]The same
[150:21.00]Like Dubuque
[150:22.00]And then
[150:23.00]The same
[150:24.00]Like Dubuque
[150:25.00]And then
[150:26.00]The same
[150:27.00]Like Dubuque
[150:28.00]And then
[150:29.00]The same
[150:30.00]Like Dubuque
[150:31.00]And then
[150:32.00]The same
[150:33.00]Like Dubuque
[150:34.00]And then
[150:35.00]The same
[150:36.00]Like Dubuque
[150:37.00]And then
[150:38.00]The same
[150:39.00]Like Dubuque
[150:40.00]And then
[150:41.00]The same
[150:42.00]Like Dubuque
[150:43.00]And then
[150:44.00]The same
[150:45.00]Like Dubuque
[150:46.00]And then
[150:47.00]The same
[150:48.00]Like Dubuque
[150:49.00]And then
[150:50.00]The same
[150:51.00]Like Dubuque
[150:52.00]And then
[150:53.00]The same
[150:54.00]Like Dubuque
[150:55.00]And then
[150:56.00]The same
[150:57.00]Like Dubuque
[150:58.00]And then
[150:59.00]The same
[151:00.00]Like Dubuque
[151:01.00]And then
[151:02.00]The same
[151:03.00]Like Dubuque
[151:04.00]And then
[151:05.00]The same
[151:06.00]Like Dubuque
[151:07.00]And then
[151:08.00]The same
[151:09.00]Like Dubuque
[151:10.00]And then
[151:11.00]The same
[151:12.00]Like Dubuque
[151:13.00]And then
[151:14.00]The same
[151:15.00]Like Dubuque
[151:16.00]And then
[151:17.00]The same
[151:18.00]Like Dubuque
[151:19.00]And then
[151:20.00]The same
[151:21.00]Like Dubuque
[151:22.00]And then
[151:23.00]The same
[151:24.00]Like Dubuque
[151:25.00]And then
[151:26.00]The same
[151:27.00]Like Dubuque
[151:28.00]And then
[151:29.00]The same
[151:30.00]Like Dubuque
[151:31.00]And then
[151:32.00]The same
[151:33.00]Like Dubuque
[151:34.00]And then
[151:35.00]The same
[151:36.00]Like Dubuque
[151:37.00]And then
[151:38.00]The same
[151:39.00]Like Dubuque
[151:40.00]And then
[151:41.00]The same
[151:42.00]Like Dubuque
[151:43.00]And then
[151:44.00]The same
[151:45.00]Like Dubuque
[151:46.00]And then
[151:47.00]The same
[151:48.00]Like Dubuque
[151:49.00]And then
[151:50.00]The same
[151:51.00]Like Dubuque
[151:52.00]And then
[151:53.00]The same
[151:54.00]Like Dubuque
[151:55.00]And then
[151:56.00]The same
[151:57.00]Like Dubuque
[151:58.00]And then
[151:59.00]The same
[152:00.00]Like Dubuque
[152:01.00]And then
[152:02.00]The same
[152:03.00]Like Dubuque
[152:04.00]And then
[152:05.00]The same
[152:06.00]Like Dubuque
[152:07.00]And then
[152:08.00]The same
[152:09.00]Like Dubuque
[152:10.00]And then
[152:11.00]The same
[152:12.00]Like Dubuque
[152:13.00]And then
[152:14.00]The same
[152:15.00]Like Dubuque
[152:16.00]And then
[152:17.00]The same
[152:18.00]Like Dubuque
[152:19.00]And then
[152:20.00]The same
[152:21.00]Like Dubuque
[152:22.00]And then
[152:23.00]The same
[152:24.00]Like Dubuque
[152:25.00]And then
[152:26.00]The same
[152:27.00]Like Dubuque
[152:28.00]And then
[152:29.00]The same
[152:30.00]Like Dubuque
[152:31.00]And then
[152:32.00]The same
[152:33.00]Like Dubuque
[152:34.00]And then
[152:35.00]The same
[152:36.00]Like Dubuque
[152:37.00]And then
[152:38.00]The same
[152:39.00]Like Dubuque
[152:40.00]And then
[152:41.00]The same
[152:42.00]Like Dubuque
[152:43.00]And then
[152:44.00]The same
[152:45.00]Like Dubuque
[152:46.00]And then
[152:47.00]The same
[152:48.00]Like Dubuque
[152:49.00]And then
[152:50.00]The same
[152:51.00]Like Dubuque
[152:52.00]And then
[152:53.00]The same
[152:54.00]Like Dubuque
[152:55.00]And then
[152:56.00]The same
[152:57.00]Like Dubuque
[152:58.00]And then
[152:59.00]The same
[153:00.00]Like Dubuque
[153:01.00]And then
[153:02.00]The same
[153:03.00]Like Dubuque
[153:04.00]And then
[153:05.00]The same
[153:06.00]Like Dubuque
[153:07.00]And then
[153:08.00]The same
[153:09.00]Like Dubuque
[153:10.00]And then
[153:11.00]The same
[153:12.00]Like Dubuque
[153:13.00]And then
[153:14.00]The same
[153:15.00]Like Dubuque
[153:16.00]And then
[153:17.00]The same
[153:18.00]Like Dubuque
[153:19.00]And then
[153:20.00]The same
[153:21.00]Like Dubuque
[153:22.00]And then
[153:23.00]The same
[153:24.00]Like Dubuque
[153:25.00]And then
[153:26.00]The same
[153:27.00]Like Dubuque
[153:28.00]And then
[153:29.00]The same
[153:30.00]Like Dubuque
[153:31.00]And then
[153:32.00]The same
[153:33.00]Like Dubuque
[153:34.00]And then
[153:35.00]The same
[153:36.00]Like Dubuque
[153:37.00]And then
[153:38.00]The same
[153:39.00]Like Dubuque
[153:40.00]And then
[153:41.00]The same
[153:42.00]Like Dubuque
[153:43.00]And then
[153:44.00]The same
[153:45.00]Like Dubuque
[153:46.00]And then
[153:47.00]The same
[153:48.00]Like Dubuque
[153:49.00]And then
[153:50.00]The same
[153:51.00]Like Dubuque
[153:52.00]And then
[153:53.00]The same
[153:54.00]Like Dubuque
[153:55.00]And then
[153:56.00]The same
[153:57.00]Like Dubuque
[153:58.00]And then
[153:59.00]The same
[154:00.00]Like Dubuque
[154:01.00]And then
[154:02.00]The same
[154:03.00]Like Dubuque
[154:04.00]And then
[154:05.00]The same
[154:06.00]Like Dubuque
[154:07.00]And then
[154:08.00]The same
[154:09.00]Like Dubuque
[154:10.00]And then
[154:11.00]The same
[154:12.00]Like Dubuque
[154:13.00]And then
[154:14.00]The same
[154:15.00]Like Dubuque
[154:16.00]And then
[154:17.00]The same
[154:18.00]Like Dubuque
[154:19.00]And then
[154:20.00]The same
[154:21.00]Like Dubuque
[154:22.00]And then
[154:23.00]The same
[154:24.00]Like Dubuque
[154:25.00]And then
[154:26.00]The same
[154:27.00]Like Dubuque
[154:28.00]And then
[154:29.00]The same
[154:30.00]Like Dubuque
[154:31.00]And then
[154:32.00]The same
[154:33.00]Like Dubuque
[154:34.00]And then
[154:35.00]The same
[154:36.00]Like Dubuque
[154:37.00]And then
[154:38.00]The same
[154:39.00]Like Dubuque
[154:40.00]And then
[154:41.00]The same
[154:42.00]Like Dubuque
[154:43.00]And then
[154:44.00]The same
[154:45.00]Like Dubuque
[154:46.00]And then
[154:47.00]The same
[154:48.00]Like Dubuque
[154:49.00]And then
[154:50.00]The same
[154:51.00]Like Dubuque
[154:52.00]And then
[154:53.00]The same
[154:54.00]Like Dubuque
[154:55.00]And then
[154:56.00]The same
[154:57.00]Like Dubuque
[154:58.00]And then
[154:59.00]The same
[155:00.00]Like Dubuque
[155:01.00]And then
[155:02.00]The same
[155:03.00]Like Dubuque
[155:04.00]And then
[155:05.00]The same
[155:06.00]Like Dubuque
[155:07.00]And then
[155:08.00]The same
[155:09.00]Like Dubuque
[155:10.00]And then
[155:11.00]The same
[155:12.00]Like Dubuque
[155:13.00]And then
[155:14.00]The same
[155:15.00]Like Dubuque
[155:16.00]And then
[155:17.00]The same
[155:18.00]Like Dubuque
[155:19.00]And then
[155:20.00]The same
[155:21.00]Like Dubuque
[155:22.00]And then
[155:23.00]The same
[155:24.00]Like Dubuque
[155:25.00]And then
[155:26.00]The same
[155:27.00]Like Dubuque
[155:28.00]And then
[155:29.00]The same
[155:30.00]Like Dubuque
[155:31.00]There has been a tremendous explosion in terms of the number of people that are interested, the number of experiment that are being tried out.
[155:38.00]And I would say we are very fortunate that there is this kind of interest right now.
[155:43.00]The thing that you notice is it does go from some of the seminal moment to seminal moment.
[155:50.00]Between AlexNet and the transformer paper in 2017, you could see that there was a huge amount of innovation on the visual side.
[156:00.00]You know, we had resonance, you know, all kinds of interesting architectures.
[156:05.00]And then after 2017, it became again significantly focused on sequence to sequence, right?
[156:11.00]So the ideal sequence here, you know, machine translation therefore was text.
[156:16.00]You know, we keep hearing about maybe there will be a breakthrough moment on the visual side again, right?
[156:22.00]In another five years.
[156:23.00]And then suddenly, you know, people will start spending more energy there.
[156:29.00]So I would like to actually go back to the modality, right?
[156:32.00]So it was the visual modality and now the text modality.
[156:35.00]And then maybe it's proteins, right?
[156:38.00]Maybe it's some modality in the healthcare space, right?
[156:42.00]Where sudden breakthroughs come about in the next several years.
[156:46.00]And then you will suddenly see a whole bunch of people looking at gene sequences and so on and so forth, right?
[156:52.00]And that interest will spike.
[156:54.00]So I actually find it fascinating that we are growing as a community from where we started.
[156:59.00]I don't really see it as a hype versus not hype.
[157:03.00]I really see it as more of the ability of this modality coupled with the neural architecture.
[157:10.00]You know, that will be the most effective and efficient at that point of time in attracting a lot of energy from the academic and industrial research community.
[157:19.00]So we can only be better off because of it.
[157:22.00]So I don't see it as a hype as much as an opportunity for learning more.
[157:27.00]I will add this.
[157:28.00]And I think it is a good segue for the next part of what you're going to talk about LSU, right?
[157:33.00]Which is I get asked this question often, right?
[157:36.00]Why did I move from NVIDIA?
[157:37.00]Like who leaves NVIDIA?
[157:39.00]And the fact of the matter is that really if you have that unfulfilled desire in you to solve the actual end problem,
[157:46.00]you have to go towards one of these verticals, right?
[157:49.00]With the whether it's health care, whether it is financial services and you have to be embedded in an organization that actually has a culture for fostering big tech like a fang like work ethic.
[158:02.00]Whether it's having the data, right?
[158:04.00]Already being in the cloud, having the roots in data driven and machine learning.
[158:09.00]So you want to be in a place which allows for that kind of creativity so that you can actually take Gen AI to its logical next step, right?
[158:19.00]We're just solving problems so that we change the financial services domain for the good, right?
[158:24.00]There's benefit to the customers.
[158:26.00]And so when you do that, though, you have to be mindful.
[158:30.00]If you are working in places where you are building recommendation systems for what to buy or what to watch.
[158:39.00]There are a lot of ways in which you may not have the best possible answer, the most accurate answer and you are still fine.
[158:46.00]But when you are in one of these domains that matter, health care, financial services, right?
[158:51.00]You really are solving a harder problem because the tolerance for error of your end customer is going to be much,
[158:59.00]much,much lower.And what that does is it forces the research and development to go in specific directions.
[159:06.00]You may not be able to just take what's out there in open source and use it as is, right?
[159:12.00]You may have to actually innovate beyond what's in the state of the art publication to make it work for this kind of domain, right?
[159:21.00]And so it takes a special class of applied researchers.
[159:25.00]It takes a special class of machine learning engineers, data scientists that have the will.
[159:31.00]The last mile is hard and you have to have that will to solve that problem.
[159:36.00]And so in some cases what ends up happening is that these roles and the work we will do may be harder than what you end up doing in a generic setting or at a platform level only or in a use case which actually doesn't have this kind of extremely low tolerance for error.
[159:54.00]So that becomes, you know, one of the main motivating functions for people to join us, right?
[160:01.00]Who really want to actually take these kind of really hard challenges.
[160:05.00]I was going through some of your team papers just from 2023, some of them in Europe, some at ICML.
[160:10.00]You've done a lot of work like class imbalance on data sets.
[160:13.00]So how do you make performing neural networks when like the data is kind of in balance in certain domains.
[160:18.00]You're doing some research on transformer graphs versus like graph neural networks.
[160:23.00]So there's just kind of like a lot in there.
[160:25.00]Any favorite project that you want to shut out any interesting paper that you saw come out of your team.
[160:30.00]You know, I could choose one, but then it would basically end up being, you know, not fair to the others.
[160:35.00]So the only on that, you know,
[160:39.00]It's like the what's your favorite child?
[160:42.00]There's no way to say one versus the other.
[160:46.00]But there are a lot of interesting trajectories of exploration here, right?
[160:51.00]We are trying to look at the superset of what all these things are trying to do.
[160:55.00]We are trying to look at data sets that are very unique to our domain.
[160:58.00]We are trying to look at attributes that are important to us like time series, tabular, right?
[161:04.00]These are things that are important, very important.
[161:06.00]We are looking at the imbalance problems, right?
[161:09.00]There are other things that we are looking at.
[161:11.00]And so I wouldn't want to just highlight one.
[161:15.00]There are a bunch of things that are of great interest to us from our perspective.
[161:20.00]Let me just not pick one.
[161:23.00]That makes sense.
[161:25.00]And yeah, just to wrap, we always like to ask our guests, who are you looking for?
[161:29.00]You know, we got a lot of AI engineers, researchers in the audience.
[161:33.00]Like, who are the type of people that are going to have a good time working with you?
[161:36.00]And what are some of the open roles that you have?
[161:38.00]So let me start with the roles.
[161:40.00]And then, you know, we can talk about the kind of people who are going to have a great time with us, right?
[161:44.00]We have a number of open roles and they span the entire spectrum from applied researchers to data scientists, to machine learning engineers, AI engineers, right?
[161:55.00]People who are listening to you right now that we really weren't interested in what we are doing.
[162:01.00]We have roles at various levels of seniority.
[162:03.00]We have individual contributor roles.
[162:05.00]That's one of the things that we have really, really double clicked on in the last several months since I've joined.
[162:10.00]And that was a thrust also before I joined, but especially in the applied research field, right?
[162:15.00]We are looking for, you know, what would be the equivalent of, you know, principal research scientists, right?
[162:20.00]Distinguished research scientists, individual contributors.
[162:23.00]We are looking for fresh graduates, right?
[162:26.00]Masters and PhD students, you know, just out of school, all the way to people who have, you know, a decade or more of experience.
[162:33.00]So, really there is a huge spectrum of talent that we want to onboard and we are looking for.
[162:41.00]Now, what is the characteristic of someone who will really come and enjoy here, right?
[162:46.00]I think I started giving that to you earlier when I said people who are interested in actually solving the problem.
[162:52.00]That last mile is hard, so we really want people to know that they have to have the stomach for that last mile.
[162:58.00]People who are good at understanding what the product requirements are, they sometimes make the best applied researchers, right?
[163:07.00]Because they can understand what our needs are and formulate problems, change architectures.
[163:15.00]We are looking for people who are very good at dealing with ambiguity.
[163:19.00]A lot of what we are doing, a lot of the advancements we are seeing, they are empirical data driven, right?
[163:25.00]And so, we need to have that as a skill.
[163:29.00]How do you actually build algorithms, systems and solutions that you can show improvement on our data?
[163:37.00]It's great if somebody is doing some architecture or some network is doing great on some of the benchmarks, right?
[163:45.00]That get published outside, but it's equally important.
[163:48.00]It's actually more important that we are able to show to ourselves that these algorithms, architectures do well on our own internal data sets and benchmarks and evaluation, right?
[164:00.00]And so, to that point, how do we actually come up with meaningful evaluation frameworks and methodologies, right?
[164:05.00]That itself is something of great interest to us.
[164:09.00]So, people in general who like to solve problems, problem solvers, people who like to go from theoretical to practical and people who are not afraid that the empirical data may not bear out their best idea and they have to go and rethink it, right?
[164:25.00]And redo it.Those are the kind of people that will really do well here.
[164:28.00]Yeah, it was great to hear your story and not a lot of people in the world, I think, that have the same depth of experience in AI, so this was awesome.
[164:35.00]Yeah, a lot of people that you're looking for are also like the kind of AI engineers that we want to encourage.
[164:40.00]So, yeah, thanks for sharing your thoughts.
[164:42.00]Thank you, Sean.
[164:43.00](音乐)
[165:07.00](音乐)
[165:09.00]字幕by索兰娅
[165:11.00]我看你不太好看
[165:12.32]我看你不太好看