[by:whisper.cpp]
[00:00.00]大家好,歡迎大家來到「Lit and Space Pockest」
[00:02.50]我是Alessio,會員,和CTO在職業的職業會員
[00:05.74]我是Makojo Swicks, founder of SmallAI
[00:08.84]今天我們有David Luan, co-founder of ADEPT,在工作室,歡迎
[00:12.98]謝謝你
[00:14.10]一段時間在工作,我遇到你在VC的社交平台上
[00:17.98]你也說了,你很興奮,我們終於能夠做到這件事了
[00:21.88]對,很高興認識你
[00:23.88]我們想介紹你的職業,然後再說一下你剛才說了什麼,在你的連結,什麼人應該知道你
[00:32.02]你開始了一間公司,是第一次在實際視頻的視頻研究,例如DEXTRO,那是你的路,在你導致的AI,你開始了XON,然後你開始了30年,你開始了OpenAI?
[00:47.06]對,30、35年,或是在那裡,或是在那裡,VP Avenge,兩年半,兩年半後,
[00:53.48]我們在2022年開始了一個大型模式的創新
[00:57.08]然後在2022年開始了一個大型模式的創新
[01:00.32]所以那是一個短暫的CV
[01:02.98]是否有其他東西?
[01:03.98]對,是否有其他東西?
[01:04.98]你覺得要做什麼?
[01:05.98]或是人們應該知道更多?
[01:07.98]我猜是一個比較大的故事
[01:09.48]是加入OpenAI比較早期的
[01:11.98]然後就做了兩、三個月的研究
[01:15.48]那是很有趣的
[01:16.48]第二或第三天的我的時間在OpenAI
[01:18.98]Gregg and Ilya 找我住在房間,我們說要拿到我們的創新,我們會去…
[01:23.98]我看過很多創新的工作
[01:25.98]所以那是很有趣的
[01:26.98]就在結合了一堆團隊
[01:28.98]有幾個早期的領導人已經有了
[01:30.98]公司的資料項目是很努力的
[01:32.98]然後再多次地在大型研究中放大型的圖案
[01:35.98]我們在做基本研究
[01:36.98]所以我花了很多時間在做這個
[01:37.98]然後我再加上Google的LM項目
[01:39.98]但也加上Google的Brain
[01:41.98]是一個Brain的領導人,更多次地
[01:42.98]你知道,有幾個不同的領導人在AI的研究
[01:46.98]我們在2012 before prehistory
[01:48.98]很多人很討厭我
[01:50.98]我跟你們三個最好的朋友
[01:51.98]寫了一個研究的文件
[01:53.98]從2012到2017
[01:56.98]我覺得遊戲的改善在2017
[01:58.98]然後很多學生都沒有發現
[01:59.98]但是我們在OpenAI上真的做了
[02:01.98]我想大部分的幫助是
[02:02.98]Ilya的 constant beating of the drum
[02:04.98]讓世界被遮蓋在data centers
[02:06.98]還有其他人需要…
[02:07.98]對,我覺得我們有確定在那裡
[02:10.98]但沒有到我們開始看到
[02:11.98]結果的結果,那是我們要去的
[02:14.98]但也有一個部分
[02:15.98]是在OpenAI上
[02:16.98]我第一次加入
[02:17.98]我認為一件事我必須要做
[02:19.98]是如何告訴我們
[02:20.98]我們是否有不同的觀點
[02:22.98]比起我們是更小的GoogleBrain
[02:25.98]或是我們在OpenAI上
[02:26.98]只要生活在SF
[02:27.98]然後不想接受Mountain View
[02:28.98]或不想要生活在London
[02:29.98]那是不足夠的
[02:31.98]利用你的技術活動
[02:33.98]所以我們真的…
[02:34.98]我花了很多時間在推廣這個
[02:36.98]就是我們要怎麼
[02:37.98]要專注在
[02:38.98]一個大學生的大學生
[02:41.98]你從最底下的研究
[02:44.98]變成了
[02:45.98]如何讓你放棄這個環境
[02:47.98]而讓你覺得
[02:48.98]什麼是大學生的大學生
[02:50.98]想要展現
[02:51.98]然後你把他們解決
[02:52.98]所有的財困
[02:53.98]不管是否要在創意
[02:54.98]創作什麼
[02:55.98]這就變成了
[02:56.98]大學生的大學生
[02:57.98]對嗎
[02:58.98]然後現在的改變
[02:59.98]是我認為
[03:00.98]第一次加入AiPrice
[03:01.98]在下一幾年
[03:02.98]會是最深的
[03:03.98] co-design
[03:04.98]和 co-evolution
[03:05.98]產品和資料
[03:07.98]和實際技術
[03:08.98]而我認為
[03:09.98]每個技術的技術
[03:10.98]都會做得很好
[03:11.98]那是一大部分
[03:12.98]為何我開始深入
[03:13.98]你提及Dota
[03:14.98]哪些記憶在想
[03:16.98]從RL 和 Transformers
[03:18.98]在時間中
[03:19.98]然後我認為
[03:20.98]製造的工具
[03:21.98]更加在LM 上
[03:23.98]然後離開
[03:24.98]更多的Agent Simulation
[03:25.98]工作
[03:26.98]像在移動的道路
[03:27.98]我覺得Agent
[03:28.98]是一個
[03:29.98]完全正確的長途
[03:30.98]你只要去找
[03:31.98]AGI 是吧
[03:32.98]你會說
[03:33.98]首先
[03:34.98]我其實不喜歡AGI
[03:35.98]用人的改變
[03:36.98]因為我真的不想
[03:37.98]這樣會發生
[03:38.98]我認為這個改變
[03:39.98]AGI 是一些
[03:40.98]人們表現的
[03:41.98]非常值得的技術
[03:43.98]是一個
[03:44.98]極端的看法
[03:45.98]和人的改變
[03:46.98]我認為
[03:47.98]我比較有興趣
[03:48.98]AGI 的改變
[03:49.98]就是
[03:50.98]一個模式
[03:51.98]可以做任何的
[03:52.98]人能做的
[03:53.98]如果你想到
[03:54.98]超級有趣
[03:55.98]Agent
[03:56.98]是一種
[03:57.98]自然的
[03:58.98]改變
[03:59.98]所以
[04:00.98]所有的工作
[04:01.98]我們在RL
[04:02.98]這些技術
[04:03.98]導致我們
[04:04.98]有很清楚的
[04:05.98]形容
[04:06.98]你需要增加
[04:07.98]你需要增加
[04:08.98]對
[04:09.98]而自然的LM
[04:10.98]形容
[04:11.98]沒有出現
[04:12.98]我認為
[04:13.98]我們
[04:14.98]在這個場地
[04:15.98]有很多想法
[04:16.98]想想
[04:17.98]我們如何解決
[04:18.98]問題的問題
[04:19.98]然後
[04:20.98]我們忘記
[04:21.98]我們在RL
[04:22.98]是一個
[04:23.98]很不容易的
[04:24.98]方式
[04:25.98]我們為何
[04:26.98]我們在世界
[04:27.98]找到所有的
[04:28.98]知識
[04:29.98]我們在一年
[04:30.98]和一位
[04:31.98]伯克里斯教授
[04:32.98]教授
[04:33.98]我們會拿到
[04:34.98]AGI
[04:35.98]他的觀點
[04:36.98]對
[04:37.98]他的理想
[04:38.98]對
[04:39.98]所以
[04:40.98]我們都在
[04:41.98]記錄
[04:42.98]我們會
[04:43.98]解決
[04:44.98]我們已經解決
[04:45.98]LM
[04:46.98]我們已經解決
[04:47.98]我們已經解決
[04:48.98]我們已經解決
[04:49.98]我們已經解決
[04:50.98]我們已經解決
[04:51.98]我們已經解決
[04:52.98]我們已經解決
[04:53.98]我們已經解決
[04:54.98]我們已經解決
[04:55.98]我們已經解決
[04:56.98]我們已經解決
[04:57.98]我們已經解決
[04:58.98]我們已經解決
[04:59.98]我們已經解決
[05:00.98]我們已經解決
[05:01.98]我們已經解決
[05:02.98]每一句
[05:03.98]文字
[05:04.98]然後所有的圖案都會學習到模式
[05:07.94]然後你能夠合作任何的組織
[05:10.14]例如寫進、聲音、畫面、其他畫面、影片等等
[05:14.42]這些都是圖案的圖案,可以學習到這類的動作
[05:18.50]所以我希望我們能夠解決這件事
[05:20.10]然後我們回到當時的歷史
[05:22.74]我們如何跟我們一起學習這些圖案的學習
[05:27.06]這就是我們要去進行的進步
[05:28.62]我還要向大家提醒你多多的明年開放的故事
[05:31.30]我們再回到大陸的故事
[05:32.90]在你的個人網站,我愛的,因為是一個很好的個人的故事
[05:37.38]故事的內容,像你的歷史
[05:39.38]我需要更新,因為太老了
[05:42.38]但是你提及GPC2,你忘記了GPC1嗎?我認為你忘記了,對吧?
[05:46.18]我其實不太記得,我記得在那邊,我記得在那邊
[05:50.70]對,《Canonical Story》是阿力的故事,他很擔心傳播者和傳播者
[05:58.74]傳播者和傳播者和傳播者的訊息
[06:01.38]對,你帶我們去… 拿我們傳播者和傳播者和傳播者的訊息
[06:03.66]GPC的歷史,你也知道,對你來說
[06:07.46]對我來說,歷史和GPC的歷史是一個很好的問題
[06:10.02]所以我認為《Canonical Story》的故事,GPC的歷史是在谷歌上,對吧?
[06:14.30]因為那是關於傳播者的故事
[06:17.30]而我認為最驚訝的一件事,是…
[06:21.26]這是一個成績,例如在谷歌設立,你跟你的最好的朋友寫文章,對吧?
[06:26.26]好,所以在調查,我認為我的工作,當我當了學校的學長,是一個領導的領導人,對吧?
[06:33.02]所以我真的有很好的朋友,我的工作是把人們的小數目和好幾個好意義,然後向他們進行完結的工作
[06:41.10]我的工作不是在提供一百萬個意義,然後沒有任何股份的資料
[06:45.54]然後當我的想法開始合作,然後我開始工作,我的工作是向他們扭動資料,向他們做好工作
[06:52.50]然後開始將一些不正確的工作拆除,對吧?
[06:56.06]那股股份並沒有存在在我的時間在谷歌上
[06:59.34]如果他們有做好工作,他們會說:
[07:02.06]"喂,你真棒,你懂這些東西的效果嗎?"
[07:05.98]"這裡是所有的我們的TPUs,然後我認為他們會殺掉我們"
[07:09.94]他肯定是想要的,他在2017年也說了一百萬公升的計劃
[07:13.18]對,所以我認為這回合是在關於GPT的故事,對嗎?
[07:15.98]就是我正在跳舞歷史,對嗎?
[07:18.38]但在GPT2之後,我們都很期待GPT2,我可以告訴你更多的故事
[07:22.50]這是我最後的一篇文章,我甚至真的受到觸碍了,所以我變成了研究研究研究員
[07:27.70]每天每天我們進行GPT3,我會醒來,然後感到緊張
[07:32.38]我感到緊張,因為...你只要看看Fax,對嗎?
[07:35.54]Google有所有的帖子,Google有所有的人 who invented all of these underlying technologies
[07:40.74]有一個人叫Noam,他很聰明,他已經做了這個討論,他想要一百萬的計劃模式
[07:46.54]我認為我們可能只是在做一些複雜的研究,對嗎?他有這個扣子,只有轉換模式,他可能會在我們之前進行的
[07:54.66]我心想,拜託,讓這個模式結束,對嗎?
[07:57.90]然後,整個時間都變成了他們沒有得到股票的資金
[08:01.62]所以,我年紀中,我帶了Google的LM的活動,我當時是一名手機的,我變得很清楚為什麼,對嗎?
[08:06.98]那時候,有一個東西叫做"Brain Credit Marketplace"
[08:11.06]你記得Brain Credit Marketplace嗎?
[08:13.26]沒有,我沒聽過這說法
[08:14.30]其實,你會問任何Google,就像一件事,對嗎?
[08:18.58]對,有限定資訊,你必須有一個市場的市場,對嗎?
[08:23.06]你可能,有些時候是貧富,有些時候是政治欺負
[08:27.34]你可能,所以,基本上,每個人都要給錢,對嗎?
[08:30.10]如果你有錢,你必須買N-CHIPS,按照貿易和責任的方式
[08:33.74]如果你想做一個大職業,你可能有19、20個朋友不願意去工作
[08:38.86]如果這就是它們的效果
[08:40.74]它們很難得獲得
[08:42.14]當中的肺炎
[08:43.86]去學習這些東西
[08:44.98]而 Google 的團隊
[08:45.86]正在打架
[08:47.02]但我們只能打擊它們
[08:48.22]因為我們拿了大大的肺炎
[08:50.62]然後我們注射
[08:51.42]然後我認為
[08:52.30]這就像是一部分的故事
[08:53.54]像是一部分的歷史
[08:54.34]像是一部分的歷史
[08:55.62]像是一部分的歷史
[08:57.58]像是一部分的歷史
[08:58.90]我認為同樣的
[09:00.22]我認為一部分的
[09:01.02]三部分會成為
[09:01.90]一部分的歷史
[09:03.22]因為是一部分的
[09:04.18]一部分的成績
[09:05.62]對
[09:06.30]我覺得這部分的內容是如何的
[09:07.70]和影片也有關的
[09:09.02]在前一天的情況下
[09:10.06]我認為可能
[09:11.10]我認為是Jensen
[09:11.90]不確定是誰
[09:12.86]把最近的照片
[09:13.90]給大家看過的
[09:15.26]他在第一張DGX的照片中
[09:17.66]我覺得Jensen 已經是
[09:19.06]一個完美的
[09:21.22]技術
[09:21.94]和精神的一切
[09:24.10]我對NVIDIA的尊敬有多大關注
[09:26.22]是不實際的
[09:26.94]但我會打開
[09:27.74]我給他們的需要
[09:29.46]讓他們構思一下
[09:30.30]或者
[09:31.34]你只要用任何NVIDIA給他們的東西
[09:33.70]所以我們很接近他們的工作
[09:35.38]我不確定能分享所有的故事
[09:37.62]但例子是我找到的
[09:39.42]特別有趣的
[09:40.14]所以 Scott Gray 是很棒的
[09:41.54]我很喜歡他
[09:42.22]他在我的隊伍中
[09:43.30]是一名超級電腦隊伍
[09:45.62]就是Chris Burner 做的
[09:46.74]Chris Burner 還做了很多東西
[09:48.82]結果
[09:49.70]我們有很接近NVIDIA的 ties
[09:52.62]其實我的 co-founder
[09:53.70]在Adept Eric Elson
[09:54.74]是一位以前的GPGPU人士
[09:56.78]所以他和Scott
[09:57.82]和Brian Kanzaro
[09:58.86]NVIDIA
[09:59.66] and Jonah
[10:00.26] and Ian at NVIDIA
[10:01.14]我覺得我們全都很接近
[10:02.54]我們是一部分的組織
[10:03.70]我們如何推動這些股票的限度
[10:05.82]我覺得那種組織
[10:07.42]幫助了我們
[10:08.38]我想有趣的部分
[10:09.50]是 knowing the A100 generation
[10:11.22]那個Quadsbar city
[10:12.26]會是一件事
[10:12.98]是我們想找到的
[10:14.50]來解決
[10:15.22]這是我們可以利用的
[10:16.50]模特兒訓練
[10:17.14] really what it boils down to
[10:18.50]是
[10:19.22]我認為更多人
[10:20.06]知道這件事
[10:21.26]6 年前
[10:22.34]甚至3 年前
[10:23.34]人們拒絕接受
[10:24.98]這個AI 是一件故事
[10:27.02]是一件故事
[10:27.62]如何讓你更能复入
[10:29.22]實際使用模特兒
[10:30.38]使用模特兒
[10:31.66]還有GPT 2 3 故事嗎
[10:35.78]你喜歡在外面
[10:37.78]我認為是
[10:38.78]很欣賞
[10:39.86]這個模特兒的作用
[10:41.66]有趣的GPT 2 故事
[10:43.66]我花了很長的時間
[10:45.86]幫Alex使用模特兒
[10:48.58]我記得
[10:49.82]最有趣的一刻
[10:52.22]是我們寫了模特兒
[10:54.70]我確定模特兒
[10:56.22]是一個最短的模特兒
[10:57.70]有任何ML
[10:58.70]像是最理想的
[10:59.90]ML 模特兒
[11:01.42]是三個模特兒
[11:03.18]這是一種模特兒
[11:04.54]Vanilla 模特兒
[11:05.58]只有轉換的模特兒
[11:06.38]這些特別的東西
[11:07.34]我記得是在《ParaGraph》裡
[11:08.58]我記得是在《ParaGraph》裡
[11:09.42]我們都在看這件事
[11:11.02]我認為是很難看的模特兒
[11:11.82]OGs 在廣場上
[11:13.02]會很討厭這個模特兒
[11:14.02]他們會說沒有創意
[11:15.50]為什麼你們要做這個作用
[11:16.94]現在是很有趣的
[11:18.02]在後期的看法是
[11:19.54]一件很刺激的作用
[11:20.82]但我覺得是一件很早的事
[11:22.54]我們完全遲到
[11:24.42]我們都要關心的問題是 AI 和不關的
[11:27.58]是否有四種不同的想法
[11:29.34]是否有一個很簡單的想法
[11:30.34]是否有一個很簡單的想法
[11:31.34]是否有一個很簡單的想法
[11:32.34]是否有一個很簡單的想法
[11:33.34]是否有一個很簡單的想法
[11:34.34]是否有一個很簡單的想法
[11:35.34]是否有一個很簡單的想法
[11:36.34]是否有一個很簡單的想法
[11:37.34]是否有一個很簡單的想法
[11:38.34]是否有一個很簡單的想法
[11:39.34]是否有一個很簡單的想法
[11:40.34]是否有一個很簡單的想法
[11:41.34]是否有一個很簡單的想法
[11:42.34]是否有一個很簡單的想法
[11:43.34]是否有一個很簡單的想法
[11:44.34]是否有一個很簡單的想法
[11:45.34]是否有一個很簡單的想法
[11:46.34]是否有一個很簡單的想法
[11:47.34]是否有一個很簡單的想法
[11:48.34]是否有一個很簡單的想法
[11:49.34]是否有一個很簡單的想法
[11:50.34]是否有一個很簡單的想法
[11:51.34]是否有一個很簡單的想法
[11:52.34]是否有一個很簡單的想法
[11:53.34]是否有一個很簡單的想法
[11:54.34]是否有一個很簡單的想法
[11:55.34]是否有一個很簡單的想法
[11:56.34]是否有一個很簡單的想法
[11:57.34]是否有一個很簡單的想法
[11:58.34]是否有一個很簡單的想法
[11:59.34]是否有一個很簡單的想法
[12:00.34]是否有一個很簡單的想法
[12:01.34]是否有一個很簡單的想法
[12:02.34]是否有一個很簡單的想法
[12:03.34]是否有一個很簡單的想法
[12:04.34]是否有一個很簡單的想法
[12:05.34]是否有一個很簡單的想法
[12:06.34]是否有一個很簡單的想法
[12:07.34]是否有一個很簡單的想法
[12:08.34]是否有一個很簡單的想法
[12:09.34]之前 Microsoft invested in OpenAI
[12:11.34]Sam Altman, myself, and our CFO
[12:13.34] flew up to Seattle
[12:14.34] to do the final pitch meeting
[12:16.34] and I'd been a founder before
[12:17.34] so I always had a tremendous amount of anxiety
[12:19.34] about partner meetings
[12:21.34] which this basis is what it was
[12:22.34] it was like Kevin Scott
[12:23.34] and Satya and Amy Hood
[12:25.34] and it was my job to give the technical slides
[12:27.34] about what's the path to AGI
[12:29.34] what's our research portfolio
[12:30.34] all of this stuff
[12:31.34] but it was also my job to give the GPT-2 demo
[12:34.34] we had a slightly bigger version of GPT-2
[12:36.34] that we had just cut
[12:38.34] maybe a day or two before this flight up
[12:40.34] and as we all know now
[12:42.34]Model behaviors you find predictable
[12:44.34] at one checkpoint
[12:45.34] are not predictable in another checkpoint
[12:46.34] and so like I spent all this time
[12:48.34] trying to figure out how to keep this thing on rails
[12:50.34] I had my canned demos
[12:51.34] but I knew I had to go
[12:52.34] turn it around over to Satya and Kevin
[12:54.34] and let them type anything in
[12:56.34] and that just that really kept me up all night
[12:58.34]Nice, yeah
[13:00.34]I mean that must have helped you
[13:01.34] talking about partners meeting
[13:03.34]You raised 420 million for ADAPT
[13:06.34]The last round was a $350 million series B
[13:09.34]So I'm sure you do great
[13:10.34]Pitching and painting
[13:12.34]Nice
[13:13.34]No, that's a high compliment coming from a VC
[13:15.34]Yeah, I mean you're doing great
[13:17.34]Let's talk about ADAPT
[13:19.34]and we were doing pre prep
[13:21.34]and you mentioned that maybe a lot of people
[13:22.34]don't understand what ADAPT is
[13:23.34]So usually we try and introduce the product
[13:26.34]and then have the founders fill in the blanks
[13:27.34]but maybe let's do the reverse
[13:28.34]Like what is ADAPT?
[13:30.34]Yeah, so I think ADAPT
[13:31.34]is the least understood company
[13:34.34]in the broader space of foundation models
[13:36.34]plus agents
[13:37.34]So I'll give some color
[13:39.34]and I'll explain what it is
[13:40.34]and I'll explain also
[13:41.34]why it's actually pretty different
[13:43.34]from what people would have guessed
[13:44.34]So the goal for ADAPT
[13:46.34]is we basically want to build an AI agent
[13:48.34]that can do
[13:49.34]that can basically help humans
[13:50.34]do anything a human does on a computer
[13:51.34]and so what that really means is
[13:53.34]we want this thing to be super good
[13:55.34]at turning natural language
[13:56.34]like goal specifications
[13:58.34]right into the correct set of end steps
[14:00.34]and then also have all the correct sensors
[14:02.34]and actuators
[14:03.34]to go get that thing done for you
[14:04.34]across any software tool
[14:05.34]that you already use
[14:06.34]and so the end vision of this
[14:07.34]is effectively like
[14:08.34]I think in a couple years
[14:09.34]everyone's going to have access
[14:10.34]to an AI teammate
[14:11.34]that they can delegate arbitrary tasks to
[14:14.34]and then also be able to use it
[14:16.34]to a sounding board
[14:17.34]and just be way, way, way more productive
[14:19.34]right and just changes the shape
[14:21.34]of every job
[14:22.34]from something where you're mostly
[14:23.34]doing execution
[14:24.34]to something where you're mostly
[14:25.34]actually doing these core liberal arts skills
[14:26.34]of what should I be doing and why
[14:28.34]right and
[14:29.34]I find this like really exciting
[14:31.34]motivating because
[14:32.34]I think it's actually
[14:33.34]pretty different vision
[14:34.34]for how AI will play out
[14:36.34]I think systems like ADAPT
[14:37.34]are the most likely systems
[14:38.34]to be proto-AGI's
[14:40.34]but I think the ways in which
[14:41.34]we are really counterintuitive
[14:42.34]to everybody
[14:43.34]is that
[14:44.34]we've actually been really quiet
[14:45.34]because we are
[14:46.34]not a developer company
[14:47.34]we don't sell APIs
[14:48.34]we don't sell open source models
[14:50.34]we also don't sell bottom-up products
[14:52.34]we're not a thing
[14:53.34]that you go and click
[14:54.34]and download the extension
[14:55.34]and like we want more users
[14:56.34]signing up for that thing
[14:57.34]we're actually an enterprise company
[14:58.34]so what we do is
[14:59.34]we work with a range
[15:00.34]of different companies
[15:01.34]some like late-stage
[15:02.34]multi-thousand people start-ups
[15:04.34]some Fortune 500s etc
[15:06.34]and what we do for them
[15:07.34]is we basically give them
[15:09.34]an out-of-the-box solution
[15:11.34]where big complex workflows
[15:12.34]that their employees
[15:13.34]do every day
[15:14.34]could be delegated to the model
[15:15.34]and so we look a little
[15:16.34]different from other companies
[15:17.34]in that in order
[15:18.34]to go build this
[15:19.34]full agent thing
[15:20.34]the most important thing
[15:21.34]you gotta get right
[15:22.34]is reliability
[15:23.34]so initially zooming
[15:24.34]way back when
[15:25.34]one of the first things
[15:26.34]debt did was we released
[15:27.34]this demo called Act 1
[15:28.34]act 1 was like pretty cool
[15:30.34]it's kind of become
[15:31.34]a hello world thing
[15:32.34]for people to show
[15:33.34]agent demos
[15:34.34]by going to redfin
[15:35.34]and asking to buy a house
[15:36.34]somewhere
[15:37.34]because like we did that
[15:38.34]in the original Act 1 demo
[15:39.34]and like showed that
[15:40.34]showed like Google Sheets
[15:41.34]all this other stuff
[15:42.34]over the last like year
[15:44.34]since that has come out
[15:45.34]there's been a lot
[15:46.34]of really cool demos
[15:47.34]and you go play with them
[15:48.34]and you realize
[15:49.34]they work 60% of the time
[15:50.34]but since we've always
[15:51.34]been focused on
[15:52.34]how do we build
[15:53.34]an amazing enterprise product
[15:54.34]enterprises can't use
[15:55.34]anything
[15:56.34]the reliability
[15:57.34]and so we've
[15:58.34]actually had to go down
[15:59.34]a slightly different
[16:00.34]tech tree than what you
[16:01.34]might find in the
[16:02.34]prompt engineering
[16:03.34]sort of plays in
[16:04.34]the agent space
[16:05.34]to get that reliability
[16:06.34]and we've decided
[16:07.34]to prioritize reliability
[16:08.34]over all else
[16:09.34]so like one of our use
[16:10.34]cases is crazy enough
[16:11.34]that it actually ends
[16:12.34]with a physical truck
[16:13.34]being sentto a place
[16:15.34]as the result
[16:16.34]of the agent workflow
[16:17.34]and if you're like
[16:18.34]if that works like 60%
[16:19.34]of the time
[16:20.34]you're just blowing money
[16:21.34]and poor truck drivers
[16:22.34]going places
[16:23.34]interesting
[16:24.34]one of the
[16:25.34]common teams
[16:26.34]has this idea of services
[16:27.34]as software
[16:28.34]I'm actually giving a talk
[16:29.34]at nvidia gtc
[16:30.34]about this
[16:31.34]but basically
[16:32.34]software as a service
[16:33.34]you're wrapping
[16:34.34]user productivity
[16:35.34]in software
[16:36.34]with agents
[16:37.34]and services as software
[16:38.34]is replacing things
[16:39.34]that you know
[16:40.34]you would ask somebody
[16:41.34]to do
[16:42.34]and the software
[16:43.34]just does it for you
[16:44.34]when you think
[16:45.34]about these usecases
[16:46.34]do the users
[16:47.34]still go in
[16:48.34]and look at the agent
[16:49.34]kindof like
[16:50.34]doing the things
[16:51.34]and can intervene
[16:52.34]or likeare they slowly
[16:53.34]remove from them
[16:54.34]are there people
[16:55.34]in the middle
[16:56.34]checking in
[16:57.34]I think there's two current flaws
[16:58.34]in the framing
[16:59.34]for services
[17:00.34]as software
[17:01.34]or I think what you just said
[17:02.34]I think that one of them
[17:03.34]is likein our experience
[17:04.34]as we've been rolling
[17:05.34]out adept
[17:06.34]the people who actually
[17:07.34]do the jobs
[17:08.34]are the most excited
[17:09.34]about it
[17:10.34]because they don't go from
[17:11.34]I do this job
[17:12.34]to I don't do this job
[17:13.34]they go from
[17:14.34]I do this job
[17:15.34]for everything
[17:16.34]including the shitty
[17:17.34]wrote stuff
[17:18.34]to I'm a supervisor
[17:19.34]and I literally
[17:20.34]likeit's pretty magical
[17:21.34]when you watch the thing
[17:22.34]being used
[17:23.34]sequentially by hand
[17:24.34]as a human
[17:25.34]and you can just click
[17:26.34]in any one of them
[17:27.34]be like hey I want to watch
[17:28.34]the trajectory
[17:29.34]the agent went through
[17:30.34]to go solve this
[17:31.34]and the nice thing
[17:32.34]about agent execution
[17:33.34]as opposed to
[17:34.34]like LLM generations
[17:35.34]is that
[17:36.34]a good chunk of the time
[17:37.34]when the agent
[17:38.34]fails to execute
[17:39.34]it doesn't give you
[17:40.34]the wrong result
[17:41.34]it just fails to execute
[17:42.34]and the whole trajectory
[17:43.34]is just broken and dead
[17:44.34]and the agent knows it
[17:45.34]right so then
[17:46.34]those are the ones
[17:47.34]that the human
[17:48.34]then goes and solves
[17:49.34]and so then they become
[17:50.34]a troubleshooter
[17:51.34]they work on the more
[17:52.34]present piece
[17:53.34]of it
[17:54.34]that we found
[17:55.34]is our strategy
[17:56.34]as a company
[17:57.34]is to always be
[17:58.34]an augmentation company
[17:59.34]and I think
[18:01.34]one out of principle
[18:02.34]that's something
[18:03.34]we really care about
[18:04.34]but two
[18:05.34]actually if you're
[18:06.34]framing yourself
[18:07.34]as an augmentation
[18:08.34]company
[18:09.34]you're always going to
[18:10.34]live in the world
[18:11.34]where you're solving
[18:12.34]tasks that are a little
[18:13.34]too hard for what
[18:14.34]the model can do today
[18:15.34]and still needs a human
[18:16.34]to provide oversight
[18:17.34]provide clarifications
[18:18.34]provide human feedback
[18:19.34]and that's how you
[18:20.34]build a data flywheel
[18:21.34]smart as humans
[18:22.34]how to solve
[18:23.34]things models
[18:24.34]can't do today
[18:25.34]and so I actually
[18:26.34]think that
[18:27.34]being an augmentation
[18:28.34]company
[18:29.34]forces you to go
[18:30.34]develop your core
[18:31.34]AI capabilities
[18:32.34]faster than someone
[18:33.34]who's saying
[18:34.34]ah okay
[18:35.34]my job's like
[18:36.34]deliver you
[18:37.34]a lights off
[18:38.34]solution for X
[18:39.34]it's interesting
[18:40.34]because we've seen
[18:41.34]two parts
[18:42.34]of the market
[18:43.34]one is
[18:44.34]we have one company
[18:45.34]that does
[18:46.34]agents for
[18:47.34]sock analysts
[18:48.34]people just
[18:49.34]don't have them
[18:50.34]which is
[18:51.34]the augmentation product
[18:52.34]and then you have
[18:53.34]sweep.dev
[18:54.34]any of these products
[18:55.34]which they just
[18:56.34]do the whole thing
[18:57.34]I'm really curious
[18:58.34]to see how that evolves
[18:59.34]I agree that today
[19:00.34]the reliability is
[19:01.34]so important
[19:02.34]in the enterprise
[19:03.34]that they just
[19:04.34]don't use
[19:05.34]most of them
[19:06.34]that's cool
[19:07.34]but it's great
[19:08.34]to hear the story
[19:09.34]because I think
[19:10.34]from the outside
[19:11.34]people are like
[19:12.34]oh that
[19:13.34]they do act one
[19:14.34]they do person on
[19:15.34]they do foo you
[19:16.34]they do all these
[19:17.34]it's just the public stuff
[19:18.34]it's just the public stuff
[19:19.34]我們想要更多的客人來領導
[19:22.20]所以我們想要更多的客人來領導
[19:26.08]但我們希望我們會更多的客人來領導
[19:29.32]我們想要更多的客人來領導
[19:31.48]我們想要更多的客人來領導
[19:33.68]所以這次我們想要更多的客人來領導
[19:36.70]為什麼你變得更多的客人?
[19:38.78]如果整個推動...
[19:40.12]你已經領導了你的公司
[19:41.82]但是你也會更加努力去領導更多的客人來領導
[19:46.20]我覺得我們剛剛領導過那一步
[19:48.14]因為我最近還沒有領導過那一步
[19:49.14]這是一個好問題
[19:50.14]我認為這兩件事其實是很重要的
[19:51.14]一件事我認為是...
[19:53.14]坦白說,大部分是公共的歷史
[19:56.14]在公司中的公司中的歷史是最重要的
[19:58.14]我非常高興這件事發生
[20:00.14]因為當我們開始公司在2022年代
[20:03.14]大家都在社會中知道歷史的歷史
[20:06.14]但公司中的歷史沒有任何意義
[20:08.14]他們還會把所有的歷史都放在桌上
[20:11.14]所以我認為現在
[20:13.14]我真的要注意的是
[20:15.14]當人們認為歷史
[20:16.14]他們會認為是對的
[20:17.14]對,所有各種各樣的東西都會被引起
[20:19.14]會被引起的電話電話電話電話
[20:20.14]會被引起的東西都會被引起的東西
[20:21.14]或是被引起的電話電話電話
[20:22.14]我認為電話電話電話
[20:23.14]是一個可以給你一個目標
[20:25.14]再次進行的工作
[20:27.14]並且在最少數個步驟中
[20:28.14]所以這就是一個大部分的原因
[20:30.14]我認為其中一個部分
[20:31.14]是因為我認為更好讓人們
[20:33.14]更加 aware of the depth
[20:34.14]他們想要做的事情
[20:35.14]他們的生意
[20:36.14]這塊地是在世界中
[20:38.14]在於在更多的利益
[20:40.14]我認為大量的利益
[20:43.14]會發生從
[20:44.14]你使用的研究模式
[20:46.14]作為大量學童的學童
[20:49.14]去解決這些事
[20:50.14]我認為那些人
[20:51.14]想要做的研究
[20:52.14]應該有所改善
[20:53.14]當你提到
[20:54.14]研究已經變成
[20:55.14]更多的一部分
[20:56.14]有什麼特別的東西
[20:57.14]你會問我嗎
[20:58.14]我會給你一個名字
[20:59.14] Bill Gates 在 his blog post
[21:00.14]提及「Agent of the Future」
[21:02.14]我是那個人 who made OSs
[21:04.14]我認為「Agent of the Next Thing」
[21:05.14]所以 Bill Gates
[21:07.14]我會叫他出來
[21:08.14]然後 Sam Altman 也會說
[21:09.14]「Agent of the Future for Open AI」
[21:10.14]我認為之前
[21:11.14]我認為
[21:12.14]有些人在《紐約 Times》
[21:13.14]Kade Metz 也在《紐約 Times》
[21:15.14]對於現在
[21:16.14]在一些不同的
[21:17.14]我看過 AI 開始的
[21:18.14]使用的研究模式
[21:19.14]是 AI 公司
[21:20.14]現在的 AI 公司
[21:21.14]是 AI 公司
[21:22.14]只是我認為
[21:23.14]是一段時間
[21:24.14]從 VC 開始
[21:25.14]是有點混合
[21:26.14]是嗎
[21:27.14]我認為有很多 VC
[21:28.14]會說我不會
[21:29.14]觸碰 any agent start-ups
[21:30.14]因為
[21:31.14]為什麼
[21:32.14]你告訴我
[21:33.14]我認為有很多 VC
[21:35.14]比較少技術
[21:37.14]不懂得
[21:38.14]限制的東西
[21:39.14]不不不
[21:40.14]你會這樣嗎
[21:41.14]不不
[21:42.14]我認為
[21:43.14]今天的可能性
[21:44.14]是否適用
[21:46.14]我認為
[21:47.14]人們會看你
[21:48.14]然後說
[21:49.14]這傢伙
[21:50.14]需要 400 億元
[21:51.14]去做
[21:52.14]所以有很多 VC
[21:53.14]都會說
[21:54.14]我會再加上
[21:55.14]有些東西
[21:56.14]協助 AI
[21:57.14]有些東西
[21:58.14]是比較容易
[21:59.14]進行
[22:00.14]進行的
[22:01.14]但我還驚訝
[22:02.14]有些 funders
[22:03.14]不想做 agent
[22:04.14]不只是 funding
[22:05.14]有時候
[22:06.14]我們在看
[22:07.14]為什麼沒有人
[22:08.14]做 agent for acts
[22:09.14]那是好
[22:10.14]其實
[22:11.14]我從沒知道
[22:12.14]我的觀點
[22:13.14]是
[22:14.14]有新的 agent company
[22:16.14]在進行
[22:17.14]所以可能
[22:18.14]他們也有
[22:19.14]但我提供人員
[22:20.14]去取消 agent
[22:21.14]他們的名字
[22:22.14]是因為
[22:23.14]他們的名字
[22:24.14]他們的名字
[22:25.14]所以
[22:26.14]他們不等待
[22:27.14]對
[22:28.14]那是好處
[22:29.14]你的 portfolio allocator
[22:31.14]有些人
[22:32.14]知道 about persimmon
[22:33.14]一些人知道
[22:34.14]for you and for you heavy
[22:35.14]你覺得
[22:36.14]怎麼想
[22:37.14]那個 evolution of that
[22:38.14]什麼人
[22:39.14]想想
[22:40.14]那是
[22:41.14]a depth
[22:42.14]搜尋個案
[22:43.14] kind of take us
[22:44.14]through the stuff
[22:45.14]you should recently
[22:46.14]and how people
[22:47.14]should think about
[22:48.14]the trajectory
[22:49.14]what you're doing
[22:50.14]the critical path
[22:51.14]for adept
[22:52.14]is we want to build
[22:53.14]agents that can do
[22:54.14]a higher and higher
[22:55.14]level of abstraction
[22:56.14]things over time
[22:57.14]all while keeping
[22:58.14]insanely
[22:59.14]high reliability standard
[23:00.14]because that's
[23:01.14]what turns this from
[23:02.14]research into something
[23:03.14]that customers want
[23:04.14]and if you build
[23:05.14]agents with really
[23:06.14]high reliability standard
[23:07.14]your users
[23:08.14]how to get that
[23:09.14]next level of
[23:10.14]straction faster
[23:11.14]so that's how
[23:12.14]you actually build
[23:13.14]the data level
[23:14.14]that's the critical path
[23:15.14]for the company
[23:16.14]everything we do
[23:17.14]is in service of that
[23:18.14]so you go zoom
[23:19.14]way way back to
[23:20.14]act one days right
[23:21.14]like the core thing
[23:22.14]behind act one
[23:23.14]is can we teach
[23:24.14]large model basically
[23:25.14]how to even
[23:26.14]actuate your computer
[23:27.14]and I think we're
[23:28.14]one of the first places
[23:29.14]to have solved that
[23:30.14]and shown it
[23:31.14]and shown the generalization
[23:32.14]that you get when you
[23:33.14]give it various different
[23:34.14]workflows and texts
[23:35.14]but I think from
[23:36.14]these models
[23:37.14]to be able to
[23:38.14]get a lot better
[23:39.14]at having some
[23:40.14]specificationof some
[23:41.14]guardrails for what it
[23:42.14]actually should be doing
[23:43.14]and I think in conjunction
[23:44.14]with that a giant thing
[23:45.14]that was really
[23:46.14]necessaryis really
[23:47.14]fast multimodal models
[23:48.14]that are really good
[23:49.14]at understanding
[23:50.14]knowledge work
[23:51.14]and really good
[23:52.14]at understanding screens
[23:53.14]and that needs to
[23:54.14]kind of be the base
[23:55.14]for some of these
[23:56.14]agentsback then
[23:57.14]we had to do a ton
[23:58.14]ofresearchbasically
[23:59.14]on how do we
[24:00.14]actually make that
[24:01.14]possiblewell first off
[24:02.14]back in
[24:03.14]free at exact
[24:04.14]one month of 23
[24:05.14]and then
[24:06.14]we had to
[24:07.14]get a lot better
[24:08.14]at the first place
[24:09.14]and then
[24:10.14]we had to
[24:11.14]get a lot better
[24:12.14]at the first place
[24:13.14]and then
[24:14.14]we had to
[24:15.14]get a lot better
[24:16.14]at the first place
[24:17.14]and then
[24:18.14]we had to
[24:19.14]get a lot better
[24:20.14]at the first place
[24:21.14]and then
[24:22.14]we had to
[24:23.14]get a lot better
[24:24.14]at the first place
[24:25.14]and then
[24:26.14]we had to
[24:27.14]get a lot better
[24:28.14]at the first place
[24:29.14]and then
[24:30.14]we had to
[24:31.14]get a lot better
[24:32.14]at the first place
[24:33.14]and then
[24:34.14]we had to
[24:35.14]get a lot better
[24:36.14]at the first place
[24:37.14]and then
[24:38.14]we had to
[24:39.14]get a lot better
[24:40.14]at the first place
[24:41.14]and then
[24:42.14]we had to
[24:43.14]get a lot better
[24:44.14]at the first place
[24:45.14]and then
[24:46.14]we had to
[24:47.14]get a lot better
[24:48.14]at the first place
[24:49.14]and then
[24:50.14]we had to
[24:51.14]get a lot better
[24:52.14]at the first place
[24:53.14]and then
[24:54.14]we had to
[24:55.14]get a lot better
[24:56.14]at the first place
[24:57.14]and then
[24:58.14]we had to
[24:59.14]get a lot better
[25:00.14]at the first place
[25:01.14]and then
[25:02.14]we had to
[25:03.14]get a lot better
[25:04.12]at the first place
[25:05.12]and then
[25:06.12]we had to
[25:07.12]get a lot better
[25:08.12]at the first place
[25:09.12]and then
[25:10.12]we had to
[25:11.12]get a lot better
[25:12.12]at the first place
[25:13.12]and then
[25:14.12]we had to
[25:15.12]get a lot better
[25:16.12]at the first place
[25:17.12]and then
[25:18.12]we had to
[25:19.12]get a lot better
[25:20.12]at the first place
[25:21.12]and then
[25:22.12]we had to
[25:23.12]get a lot better
[25:24.12]at the first place
[25:25.12]and then
[25:26.12]we had to
[25:27.12]get a lot better
[25:28.12]at the first place
[25:29.12]and then
[25:30.12]we had to
[25:31.12]get a lot better
[25:32.12]at the first place
[25:33.12]and then
[25:34.12]we had to
[25:35.12]get a lot better
[25:36.12]at the first place
[25:37.12]and then
[25:38.12]we had to
[25:39.12]get a lot better
[25:40.12]at the first place
[25:41.12]and then
[25:42.12]we had to
[25:43.12]get a lot better
[25:44.12]at the first place
[25:45.12]and then
[25:46.12]we had to
[25:47.12]get a lot better
[25:48.12]at the first place
[25:49.12]and then
[25:50.12]we had to
[25:51.12]get a lot better
[25:52.12]at the first place
[25:53.12]and then
[25:54.12]we had to
[25:55.12]get a lot better
[25:56.12]at the first place
[25:57.12]and then
[25:58.12]we had to
[25:59.12]get a lot better
[26:00.12]at the first place
[26:01.12]and then
[26:02.12]we had to
[26:03.12]get a lot better
[26:04.12]at the first place
[26:05.12]and then
[26:06.12]we had to
[26:07.12]get a lot better
[26:08.12]at the first place
[26:09.12]and then
[26:10.12]we had to
[26:11.12]get a lot better
[26:12.12]at the first place
[26:13.12]and then
[26:14.12]we had to
[26:15.12]get a lot better
[26:16.12]at the first place
[26:17.12]and then
[26:18.12]we had to
[26:19.12]get a lot better
[26:20.12]at the first place
[26:21.12]and then
[26:22.12]we had to
[26:23.12]get a lot better
[26:24.12]at the first place
[26:25.12]and then
[26:26.12]we had to
[26:27.12]get a lot better
[26:28.12]at the first place
[26:29.12]and then
[26:30.12]we had to
[26:31.12]get a lot better
[26:32.12]at the browser level
[26:33.12]I really want
[26:34.12]at your papers
[26:35.12]you have like a different representation
[26:36.12]kind of like
[26:37.12]you don't just take the dome
[26:38.12]and act on it
[26:39.12]you do a lot more stuff
[26:40.12]how do you think about
[26:41.12]the best way
[26:42.12]the models will interact
[26:43.12]with the software
[26:44.12]and like how
[26:45.12]the development of products
[26:46.12]is going to change
[26:47.12]with that in mind
[26:48.12]as more and more
[26:49.12]the work is done by agents
[26:50.12]instead of people
[26:51.12]this is
[26:52.12]there's so much surface area here
[26:53.12]and it's actually one of the things
[26:54.12]I'm really excited about
[26:55.12]and it's funny because
[26:56.12]I've spent most of my time
[26:57.12]doing research stuff
[26:58.12]but this is like a whole
[26:59.12]new ball game that I've been
[27:00.12]doing about
[27:01.12]and I find it
[27:02.12]really cool
[27:03.12]so I would say
[27:04.12]the best analogy
[27:05.12]I have to
[27:06.12]why ADAPT
[27:07.12]is pursuing a path
[27:08.12]of being able to
[27:09.12]use your computer
[27:10.12]like a human
[27:11.12]plus of course
[27:12.12]being able to call
[27:13.12]APIs
[27:14.12]being able to call
[27:15.12]APIs is the easy part
[27:16.12]like being able to
[27:17.12]use your gear like humans
[27:18.12]is a hard part
[27:19.12]it's in the same way
[27:20.12]why people are excited
[27:21.12]about humanoid robotics
[27:22.12]right
[27:23.12]in a world where
[27:24.12]you had t=infinity
[27:25.12]right you're probably
[27:26.12]gonna have various
[27:27.12]different form factors
[27:28.12]that robots
[27:29.12]do
[27:30.12]without changing
[27:31.12]everything along the way
[27:32.12]it's the same thing
[27:33.12]for software
[27:34.12]right
[27:35.12]if you go itemize out
[27:36.12]the number of things
[27:37.12]you wanna do on your computer
[27:38.12]for which every step
[27:39.12]has an api
[27:40.12]those numbers
[27:41.12]will workflows add up
[27:42.12]pretty close to zero
[27:43.12]and so then many
[27:44.12]points along the way
[27:45.12]you need the ability
[27:46.12]to actually control
[27:47.12]your computer like a human
[27:48.12]it also lets you learn
[27:49.12]from human usage
[27:50.12]of computers
[27:51.12]as a source of training
[27:52.12]data that you don't get
[27:53.12]if you have to somehow
[27:54.12]figure out how every
[27:55.12]particular step needs to be
[27:56.12]some particular custom
[27:57.12]private api thing
[27:58.12]it's the most practical path
[27:59.12]i think a lot of
[28:00.12]success will come
[28:01.12]from going down
[28:02.12]this path
[28:03.12]i kinda think about this
[28:04.12]early days of the agent
[28:05.12]interaction layer
[28:06.12]level is a little bit
[28:07.12]like do y'all remember
[28:08.12]windows 3.1
[28:10.12]like those days
[28:11.12]this might be
[28:12.12]i might be too old
[28:13.12]for you guys on this
[28:14.12]but back in the day
[28:15.12]windows 3.1
[28:16.12]we had this transition period
[28:17.12]between pure command line
[28:18.12]right
[28:19.12]being the default
[28:20.12]into this new world
[28:21.12]with the gui is the default
[28:22.12]and then you drop into the
[28:23.12]command line for like
[28:24.12]programmer things
[28:25.12]the old way was
[28:26.12]you booted your computer up
[28:27.12]and then it would
[28:28.12]give you the c colon
[28:29.12]slash thing
[28:30.12]and you typed windows
[28:31.12]and you hit enter
[28:32.12]and then you got
[28:33.12]put into windows
[28:34.12]and then the gui
[28:35.12]kind of became a layer
[28:36.12]above the command line
[28:37.12]the same thing
[28:38.12]is gonna happen
[28:39.12]with agent interfaces
[28:40.12]is like today
[28:41.12]what we have in the gui
[28:42.12]is like the base layer
[28:44.12]and then the agent
[28:45.12]just controls
[28:46.12]the current gui
[28:47.12]layer plus apis
[28:48.12]and in the future
[28:50.12]as more and more
[28:51.12]trust is built towards
[28:52.12]agents and more and more
[28:53.12]things can be done by
[28:54.12]agents and more UIs
[28:55.12]for agents are actually
[28:56.12]users
[28:57.12]then that just becomes
[28:58.12]a standard
[28:59.12]interaction layer
[29:00.12]and if that becomes
[29:01.12]a standard
[29:02.12]interaction layer
[29:03.12]what changes for
[29:04.12]software is that
[29:05.12]a lot of software
[29:06.12]is gonna be
[29:07.12]either systems
[29:08.12]or record
[29:09.12]or like certain
[29:10.12]customized
[29:11.12]workflow
[29:12.12]execution engines
[29:13.12]and a lot of
[29:14.12]how you actually
[29:15.12]do stuff will be
[29:16.12]controlled at the
[29:17.12]agent layer
[29:18.12]and you think the
[29:19.12]rabbit interface
[29:20.12]is more like
[29:21.12]it would like
[29:22.12]you're not actually
[29:23.12]seeing the app
[29:24.12]that the model
[29:25.12]I can see that
[29:26.12]being a model
[29:27.12]I think
[29:28.12]I don't know
[29:29.12]enough about
[29:30.12]what using
[29:31.12]rabbit in real life
[29:32.12]will actually be like
[29:33.12]to comment on
[29:34.12]that particular
[29:35.12]thing but I think
[29:36.12]the broader idea
[29:37.12]that you know
[29:38.12]you have a goal
[29:39.12]the agent knows
[29:40.12]how to break
[29:41.12]your goal down into steps
[29:42.12]the agent knows
[29:43.12]how to use
[29:44.12]the underlying
[29:45.12]software
[29:46.12]and systems
[29:47.12]or record
[29:48.12]to achieve
[29:49.12]that goal for you
[29:50.12]the agent may presents
[29:51.12]you information
[29:52.12]in a custom way
[29:53.12]that's only
[29:54.12]you're a power
[29:55.12]user
[29:56.12]for some niche thing
[29:57.12]general question
[29:58.12]so first of all
[29:59.12]I think like
[30:00.12]the sort of input
[30:01.12]mode conversation
[30:02.12]I wonder if you have
[30:03.12]any analogies
[30:04.12]that you like
[30:05.12]with self-driving
[30:06.12]because I do think
[30:07.12]there's a little bit
[30:08.12]of how the model
[30:09.12]should perceive the world
[30:10.12]and you know
[30:11.12]the primary split
[30:12.12]in self-driving
[30:13.12]is LiDAR
[30:14.12]versus camera
[30:15.12]and I feel like
[30:16.12]most agent companies
[30:17.12]that I'm tracking
[30:18.12]are all moving towards
[30:19.12]camera approach
[30:20.12]which is like
[30:21.12]the multimodal approach
[30:22.12]that we're doing
[30:23.12]you're
[30:24.12]focusing on that
[30:25.12]including charts
[30:26.12]and tables
[30:27.12]and do you find
[30:28.12]inspiration there
[30:29.12]from the self-driving
[30:30.12]world?
[30:31.12]that's a good question
[30:32.12]I think sometimes
[30:33.12]the most useful
[30:34.12]inspiration I've found
[30:35.12]from self-driving
[30:36.12]is the levels analogy
[30:37.12]I think that's awesome
[30:38.12]but I think that
[30:39.12]our number one
[30:40.12]goals for agents
[30:41.12]not to look like
[30:42.12]self-driving
[30:43.12]we want to minimize
[30:44.12]the chances
[30:45.12]that agents are sort
[30:46.12]of a thing
[30:47.12]that you just
[30:48.12]have to bang
[30:49.12]your head at
[30:50.12]for a long time
[30:51.12]to get to like
[30:52.12]completely
[30:53.12]and that takes you
[30:54.12]all the way
[30:55.12]up to the top
[30:56.12]but similarly
[30:57.12]I mean
[30:58.12]compared to self-driving
[30:59.12]like two things
[31:00.12]that people really
[31:01.12]undervalue
[31:02.12]that's like really
[31:03.12]easy to driving
[31:04.12]a car down
[31:05.12]highway 101
[31:06.12]in a sunny day
[31:07.12]demo
[31:08.12]that actually
[31:09.12]doesn't prove anything
[31:10.12]anymore
[31:11.12]and I think
[31:12.12]the second thing
[31:13.12]is that
[31:14.12]as a non-self-driving
[31:15.12]expert
[31:16.12]I think one of the things
[31:17.12]that we believe
[31:18.12]really strongly
[31:19.12]is that
[31:20.12]everyone under
[31:21.12]get a lot
[31:22.12]of reliability
[31:23.12]is a really
[31:24.12]strong focus on
[31:25.12]actually why
[31:26.12]does the model
[31:27.12]not do this thing
[31:28.12]and the non-trivial amount
[31:29.12]of time
[31:30.12]the time the model
[31:31.12]doesn't actually
[31:32.12]do the thing
[31:33.12]is because if
[31:34.12]you're a wizard
[31:35.12]of ozing it yourself
[31:36.12]or if you have
[31:37.12]unreliable actuators
[31:38.12]you can't do the thing
[31:39.12]and so we've
[31:40.12]had to fix
[31:41.12]a lot of those problems
[31:42.12]I was slightly
[31:43.12]surprised just because
[31:44.12]I do generally
[31:45.12]consider the way
[31:46.12]most that we see
[31:47.12]all around San Francisco
[31:48.12]as the most
[31:49.12]I guess real case
[31:50.12]it's a big
[31:51.12]job but it has taken
[31:52.12]a long time
[31:53.12]for self-driving
[31:54.12]temperature from
[31:55.12]when it entered
[31:56.12]the consciousness
[31:57.12]and the driving down
[31:58.12]when it went on a sunny
[31:59.12]day moment
[32:00.12]happened to now.
[32:01.12]so I want to see
[32:02.12]the more compressed
[32:03.12]cruise, you know,
[32:04.12]R.I.P.
[32:05.12]recently.
[32:06.12]and then one more thing
[32:07.12]on just like
[32:08.12]just going back on
[32:09.12]this reliability
[32:10.12]thing, something
[32:11.12]I have been holding
[32:12.12]in my head
[32:13.12]that I'm curious
[32:14.12]to get your commentary on
[32:15.12]is I think there's a
[32:16.12]treatup between
[32:17.12]reliability and generality
[32:18.12]or I want to broaden
[32:19.12]because you have
[32:20.12]reliability also have
[32:21.12]cost of speed
[32:22.12]speed is a huge emphasis
[32:23.12]for a debt
[32:24.12]the tendency or the
[32:25.12]attemptation is to reduce
[32:26.12]generalityto improve
[32:27.12]reliability
[32:28.12]and to improve
[32:29.12]cost improve speed
[32:30.12]do you perceive a tradeoff
[32:31.12]do you have any
[32:32.12]insights that
[32:33.12]solve those tradeoffs
[32:34.12]for you guys
[32:35.12]there's definitely a tradeoff
[32:36.12]if you're at
[32:37.12]the predo frontier
[32:38.12]I think a lot of folks
[32:39.12]aren't actually
[32:40.12]at the predo frontier
[32:41.12]I think the way you get
[32:42.12]there is basically
[32:43.12]how do you frame
[32:44.12]the fundamental
[32:45.12]agent problem in a way
[32:46.12]that just continues
[32:47.12]to benefit from data
[32:48.12]I think one of
[32:49.12]the main ways
[32:50.12]of being able to solve
[32:51.12]that particular tradeoff
[32:52.12]is you basically
[32:53.12]just want to formulate
[32:54.12]the problem such that
[32:55.12]every particular use
[32:56.12]case just looks like
[32:57.12]you collecting more
[32:58.12]data to go make
[32:59.12]that use case possible
[33:00.12]I think that's how
[33:01.12]you really solve it
[33:02.12]then you get into the
[33:03.12]other problems like
[33:04.12]are you overfitting
[33:05.12]on these end use cases
[33:06.12]right but like you're
[33:07.12]not doing a thing
[33:08.12]where you're like
[33:09.12]being super prescriptive
[33:10.12]for the end steps
[33:11.12]that the model can
[33:12.12]only do for example
[33:13.12]then the question becomes
[33:14.12]kind of do you have
[33:15.12]one sort of house model
[33:16.12]they customize
[33:17.12]the customer's
[33:18.12]specific use case
[33:19.12]we're not sharing
[33:20.12]we're not sharing
[33:21.12]it's tempting
[33:22.12]but that doesn't
[33:23.12]look like AGI to me
[33:24.12]you know what I mean
[33:25.12]that is just
[33:26.12]you have a good
[33:27.12]base model
[33:28.12]and then
[33:29.12]you fine tune it
[33:30.12]for what it's worth
[33:31.12]I think there's
[33:32.12]two paths
[33:33.12]to a lot more
[33:34.12]capability coming out
[33:35.12]of the models
[33:36.12]that we
[33:37.12]all are training
[33:38.12]these days
[33:39.12]one path
[33:40.12]is you figure out
[33:41.12]how to spend
[33:42.12]compute and turn
[33:43.12]into data
[33:44.12]and so in that
[33:45.12]path I consider
[33:46.12]off play
[33:47.12]all that stuff
[33:48.12]the second path
[33:49.12]is how do you
[33:50.12]get super
[33:52.12]competent
[33:53.12]high intelligence
[33:54.12]demonstrations
[33:55.12]from humans
[33:56.12]and I think
[33:57.12]the right way
[33:58.12]to move forward
[33:59.12]is you kind of
[34:00.12]want to combine the two
[34:01.12]the first one
[34:02.12]gives you maximum
[34:03.12]sample efficiency
[34:04.12]for the second
[34:05.12]but I think
[34:06.12]that is going to be
[34:07.12]hard to be running
[34:08.12]at max speed
[34:09.12]towards AGI
[34:10.12]without actually
[34:11.12]solving a bit of both
[34:12.12]you haven't talked
[34:13.12]much about synthetic
[34:14.12]data as far as I can
[34:15.12]any insights
[34:16.12]on using synthetic
[34:17.12]data to augment
[34:18.12]the expensive
[34:19.12]human data
[34:20.12]the best part
[34:21.12]about framing AGI
[34:22.12]is being able
[34:23.12]to help people do
[34:24.12]things on computers
[34:25.12]is you have an environment
[34:26.12]yes
[34:27.12]so you can
[34:28.12]simulate all of it
[34:29.12]you can do a lot
[34:30.12]of stuff
[34:31.12]when you have an environment
[34:32.12]we were having dinner
[34:33.12]for our one year
[34:34.12]anniversary
[34:35.12]the other round
[34:36.12]thank you
[34:37.12]Raza from human
[34:38.12]loop was there
[34:39.12]and we mentioned
[34:40.12]you were coming on
[34:41.12]the pod
[34:42.12]this is our first
[34:43.12]so he submitted a question
[34:44.12]now you had
[34:45.12]gbd4 vision
[34:46.12]and help you
[34:47.12]building a lot
[34:48.12]of those things
[34:49.12]how do you think
[34:50.12]about the things
[34:51.12]that are unique to you
[34:52.12]as a depth
[34:53.12]and like going back
[34:54.12]to like the maybe
[34:55.12]research direction
[34:56.12]that you want to take
[34:57.12]the team and what you
[34:58.12]want people to come
[34:59.12]work on at a depth
[35:00.12]versus what is maybe
[35:01.12]not become commoditized
[35:02.12]that you didn't expect
[35:03.12]everybody would
[35:04.12]have access to
[35:05.12]yeah that's
[35:06.12]a really good question
[35:07.12]I think implicit
[35:08.12]in that question
[35:09.12]and I wish he were
[35:10.12]tier two so he can
[35:11.12]push back on my
[35:12.12]assumption about his
[35:13.12]questionbut I think
[35:14.04]is calculus of where
[35:16.04]does advantage a crew
[35:18.04]in the overall
[35:19.04]ML stack
[35:20.04]and maybe part
[35:21.04]of the assumption
[35:22.04]is that advantage
[35:23.04]a crew is solely
[35:24.04]to base model scaling
[35:25.04]but I actually
[35:26.04]believe pretty strongly
[35:27.04]that the way
[35:28.04]that you really
[35:29.04]win is that you
[35:30.04]have to go build
[35:31.04]an agent stack
[35:32.04]that is much more
[35:33.04]than that
[35:34.04]of the base model itself
[35:35.04]and so I think
[35:36.04]like that is
[35:37.04]always going to be
[35:38.04]a giant advantage
[35:39.04]of vertical integration
[35:40.04]I think like
[35:41.04]it lets us do things
[35:42.04]like have a really
[35:43.04]bad cat and dog
[35:44.04]photo
[35:45.04]it's pretty good
[35:46.04]at cat and dog
[35:47.04]photo
[35:48.04]it's not like
[35:49.04]soda at cat
[35:50.04]and dogphoto
[35:51.04]so like we're allocating
[35:52.04]our capacity wisely
[35:53.04]is like one thing
[35:54.04]that you
[35:55.04]really get to do
[35:56.04]I also think that
[35:57.04]the other thing
[35:58.04]that is pretty
[35:59.04]important now
[36:00.04]in the broader
[36:01.04]foundation modeling
[36:02.04]space is
[36:03.04]I feel despite any
[36:04.04]potential concerns
[36:05.04]about how good
[36:06.04]is agents as
[36:07.04]like a startup area
[36:08.04]like we were talking
[36:09.04]about earlier
[36:10.04]I feel super good
[36:11.04]that we're
[36:12.04]cap just flowing
[36:13.04]from can we make
[36:14.04]a better agent
[36:15.04]because right now
[36:16.04]I think we all see
[36:17.04]that you know
[36:18.04]if you're training
[36:19.04]on publicly available
[36:20.04]web data
[36:21.04]you put in the
[36:22.04]flops and you do
[36:23.04]reasonable things
[36:24.04]then you get
[36:25.04]decent results
[36:26.04]and if you just
[36:27.04]double the amount
[36:28.04]of compute
[36:29.04]then you get
[36:30.04]predictably
[36:31.04]better results
[36:32.04]and so I think
[36:33.04]pure play foundation
[36:34.04]model companies
[36:35.04]are just going to be
[36:36.04]pinched by how
[36:37.04]good the next couple
[36:38.04]lamas are going to be
[36:39.04]and the next
[36:40.04]what good open source
[36:41.04]on these base foundation
[36:42.04]models I think it's
[36:43.04]gonna commoditize a lot
[36:44.04]of the regular llms
[36:45.04]and soon regular
[36:46.04]multimodal models
[36:47.04]so I feel really good
[36:48.04]that we're just focused
[36:49.04]on agents so you
[36:50.04]don't consider yourself
[36:51.04]a pure play foundation
[36:52.04]model company no
[36:53.04]because if we were pure
[36:54.04]play foundation model
[36:55.04]company we would be
[36:56.04]traininggeneral foundation
[36:57.04]models that do
[36:58.04]summarization and
[36:59.04]all this dedicated
[37:00.04]towards the agent
[37:01.04]yeah and our business
[37:02.04]is an agent business
[37:03.04]we're not here to
[37:04.04]sell you tokens right
[37:05.04]and I think like
[37:06.04]selling tokens unless
[37:07.04]there's like yeah I
[37:08.04]love it there's like
[37:09.04]if you have a particular
[37:10.04]area of specialty
[37:11.04]right then you won't
[37:13.04]get caught in the fact
[37:14.04]that everyone's just
[37:15.04]scaling to ridiculous
[37:16.04]levels of compute
[37:17.04]but if you don't have a
[37:18.04]specialty I find that
[37:19.04]I think it's gonna be
[37:20.04]a little tougher
[37:21.04]interesting are you
[37:22.04]interested in robotics at
[37:23.04]all just a personally
[37:24.04]fascinated by robotics
[37:25.04]have always loved robotics
[37:26.04]embodied agents as a
[37:27.04]business you know figure
[37:28.04]is like a big also
[37:29.04]so the open ai
[37:30.04]affiliated company
[37:31.04]that raises a lot of
[37:32.04]money I think it's
[37:33.04]cool I think I mean
[37:34.04]I don't know exactly
[37:35.04]what they're exactly
[37:36.04]what they're doing but
[37:37.04]robots yeah yeah
[37:38.04]well I mean that's
[37:39.04]well Christian
[37:40.04]would you ask
[37:41.04]like if we
[37:42.04]had them on like
[37:43.04]what would you ask them
[37:44.04]oh I just want to
[37:45.04]understand what their
[37:46.04]overall strategy is
[37:47.04]gonna be between now
[37:48.04]and when there's reliable
[37:49.04]stuff to be deployed
[37:50.04]but honestly
[37:51.04]I just don't know
[37:52.04]enough about it
[37:53.04]and if I told you
[37:54.04]hey fire your entire
[37:55.04]warehouse workforce
[37:56.04]and you know
[37:57.04]put robots in there
[37:58.04]isn't that a strategy
[37:59.04]oh yeah yeah sorry
[38:00.04]I'm not questioning
[38:01.04]whether
[38:02.04]they're doing smart
[38:03.04]things I genuinely
[38:04.04]don't know what
[38:05.04]they're doing as much
[38:06.04]but I think there's
[38:07.04]two things one
[38:08.04]it's just
[38:09.04]I think it's
[38:10.04]just gonna work
[38:11.04]like I will die
[38:12.04]on this hill
[38:13.04]like I mean
[38:14.04]like again this whole
[38:15.04]this whole time
[38:16.04]like we've been on this
[38:17.04]podcast it's just
[38:18.04]gonna continually saying
[38:19.04]these models
[38:20.04]are basically behavioral
[38:21.04]cloners right
[38:22.04]so let's go behavioral
[38:23.04]clone all this like
[38:24.04]robot behavior right
[38:25.04]and then
[38:26.04]now you figure out
[38:27.04]everything else
[38:28.04]you have to do in order
[38:29.04]to teach you how to
[38:30.04]solve new problem
[38:31.04]that's gonna work
[38:32.04]I'm super stoked for that
[38:33.04]I think unlike
[38:34.04]what we're doing with
[38:35.04]helping humans with
[38:36.04]knowledge work
[38:37.04]and I'm personally
[38:38.04]less excited about that
[38:39.04]we had a
[38:40.04]canjun from imbu
[38:41.04]on the podcast
[38:42.04]we asked her
[38:43.04]why people should
[38:44.04]go work there
[38:45.04]and not at adept
[38:46.04]so I wanna
[38:47.04]well she said
[38:48.04]you know
[38:49.04]there's space for everybody
[38:50.04]in this market
[38:51.04]we're all doing
[38:52.04]interesting work
[38:53.04]and she said
[38:54.04]they're really excited
[38:55.04]about building
[38:56.04]an operating system
[38:57.04]for agent
[38:58.04]and for her
[38:59.04]the biggest research
[39:00.04]thing was like
[39:01.04]getting models
[39:02.04]better reasoning
[39:03.04]and planning
[39:04.04]for these agents
[39:05.04]the reverse question
[39:06.04]I'm excited to
[39:07.04]come work at adept
[39:08.04]instead of imbu
[39:09.04]and maybe
[39:10.04]what are like
[39:11.04]the core research
[39:12.04]questions
[39:13.04]that people should
[39:14.04]be passionate about
[39:15.04]to have fun at adept
[39:16.04]yeah first off
[39:17.04]I think that
[39:18.04]I'm sure you guys
[39:19.04]believe this too
[39:20.04]the AI space
[39:21.04]to the center
[39:22.04]there's an AI space
[39:23.04]and the AI agent
[39:24.04]space are both
[39:25.04]exactly as
[39:26.04]she likely said
[39:27.04]I think colossal
[39:28.04]opportunities
[39:29.04]and people are just
[39:30.04]going to end up
[39:31.04]winning in different
[39:32.04]areas and a lot
[39:33.04]of companies are
[39:34.04]going to do well
[39:35.04]to be at
[39:36.04]adept
[39:37.04]I think there's
[39:38.04]two huge reasons
[39:39.04]to be at adept
[39:40.04]I think one of them
[39:41.04]is everything we do
[39:42.04]is in the service
[39:43.04]of like useful agents
[39:44.04]we're not a
[39:45.04]research lab
[39:46.04]we do a lot of research
[39:47.04]in service of that goal
[39:48.04]but we don't
[39:49.04]think about ourselves
[39:50.04]as like a classic
[39:51.04]research lab at all
[39:52.04]and I think the second
[39:53.04]reason at work at
[39:54.04]adeptis
[39:55.04]if you believe that
[39:56.04]actually having customers
[39:57.04]and a reward signal
[39:58.04]from customers
[39:59.04]lets you build
[40:00.04]AGI faster
[40:01.04]which we really believe
[40:02.04]then you should come here
[40:03.04]and I think the examples
[40:04.04]are evaluations
[40:05.04]they're not
[40:06.04]academic evals
[40:07.04]they're not simulator
[40:08.04]evals
[40:09.04]they're like
[40:10.04]okay we have a
[40:11.04]customer that
[40:12.04]really needs us to do
[40:13.04]these particular things
[40:14.04]we can do some
[40:15.04]of them
[40:16.04]these other ones
[40:17.04]they want us to
[40:18.04]we can't do them at
[40:19.04]all we've turned
[40:20.04]those into evals
[40:21.04]solve it
[40:22.04]I think that's
[40:23.04]really cool
[40:24.04]like everybody knows
[40:25.04]a lot of these evals
[40:26.04]are like
[40:27.04]pretty saturated
[40:28.04]and the new ones
[40:29.04]that even are
[40:30.04]not saturated you look
[40:31.04]at someone and you're
[40:32.04]like is this actually
[40:33.04]and all of this stuff
[40:34.04]but they're very grounded
[40:35.04]and actual needs
[40:36.04]right now
[40:37.04]which is really cool
[40:38.04]yeah this has been
[40:39.04]wonderful dive
[40:40.04]I wish we had more time
[40:41.04]but I'll just leave it
[40:42.04]kind of open to you
[40:43.04]I think you have broad thoughts
[40:44.04]you know just about
[40:45.04]the agent space
[40:46.04]but also just general AI
[40:47.04]space any sort of rants
[40:48.04]or things that
[40:49.04]they're just helping
[40:50.04]might for you right now
[40:51.04]any rants
[40:52.04]minding you
[40:53.04]for just general
[40:54.04]wow okay
[40:55.04]so Amelia's already
[40:56.04]made the rant better
[40:57.04]than I have
[40:58.04]but not just
[40:59.04]not just chatbots
[41:00.04]is like kind of rant one
[41:01.04]but the rant two
[41:02.04]is AI's really been
[41:03.04]the story of compute
[41:04.04]and compute plus data
[41:06.04]and ways in which
[41:07.04]you could change one
[41:08.04]for the other
[41:09.04]and I think as much as
[41:10.04]our research community
[41:11.04]is really smart
[41:12.04]we have made many
[41:13.04]many advancements
[41:14.04]and that's going to
[41:15.04]continue to be important
[41:16.04]but now I think
[41:17.04]the game is
[41:18.04]increasingly changing
[41:19.04]and the rapid
[41:20.04]industrialization
[41:21.04]error has begun
[41:22.04]and I think
[41:23.04]we unfortunately
[41:24.04]have to embrace it
[41:25.04]excellent awesome David
[41:26.04]thank you so much
[41:27.04]for your time
[41:28.04]cool yeah thanks guys
[41:29.04]this was fun
[41:30.04]thank you
[41:31.04]thank you
[41:32.04]thank you
[41:32.04]thank you
[41:33.04]thank you
[41:34.04]thank you
[41:35.04]thank you
[41:36.04]thank you
[41:37.04]thank you
[41:38.04]thank you
[41:39.04]thank you
[41:40.04]thank you
[41:41.04]thank you
[41:42.04]thank you
[41:43.04]thank you
[41:44.04]thank you
[41:45.04]thank you
[41:46.04]thank you
[41:47.04]字幕by索兰娅
[41:49.04]字幕:J Chong
[41:50.04]请不吝点赞 订阅 转发 打赏 打赏