6/8/2026 - The Shift From Prompting to Looping and What It Means for the Future of Work

Transcript

1. 0:02[music] [music] [music]
2. 0:22[music] [music] Good evening. Good evening. Good
3. 0:37evening. Welcome to the AI learning lab. Happy Monday. It is Monday, June 8th, 2026. June 8th, 2026. What is the significance
4. 0:49of that? I don't know. 6826. Nothing. Everything's divisible by two in those numbers. You go play the lotto
5. 1:00or the numbers. Go play the numbers. The pick four. Um, welcome, welcome, welcome.
6. 1:18It's International Nerd Day. [laughter] What we're going to do tonight is we're going to read through some blog posts. [laughter]
7. 1:29So, if you're here, if [clears throat] you're here, welcome. Okay, a couple of things, couple little housekeeping
8. 1:34things. If you're new to the Hey, Mary Chester. Uh, Chester, if if you're new to the AI salon, welcome to the AI
9. 1:41Salon. If you're new to this streaming platform, up in this upper corner right here is the um is the uh is the chat.
10. 1:52So, if you click on that, you'll be able to chat um and put comments in and things like that. Did you watch the
11. 1:58Apple event? I didn't watch the Apple event, but I've seen a couple of posts on it. Um, so I don't have I don't have
12. 2:05a huge amount of opinion on it right now, but we can go look at some of that stuff, but there's other stuff going on. Um, Apple
13. 2:17Apple's going to be fine. I I think I think where a lot of all of what we're going to talk about tonight is going,
14. 2:26including Apple and including their entire ecosystem of apps. I think the concept of apps is is going to functionally go away and
15. 2:51We just We're just living in Say hi to Greg. Yo, Greg, thanks for having me. Thanks for being here. Welcome, welcome, welcome. Um,
16. 3:02thanks for having me. We're excited to have you here. Um, loops. We're going to talk about loops. We're going to talk about loops and and
17. 3:16prompting. You're still prompting. What are you thinking? We don't prompt anymore.
18. 3:30If me saying that is making you anxious and making you feel exhausted and making you feel like, "What are loops? I can't
19. 3:37keep up with this. What?" That's okay. It's gonna get more insane. And and like an absolute option is just
20. 3:52don't deal with it because I think there's actually going to be an advantage. There's going to be if you just ignore this AI [ __ ] for the
21. 4:03next year or two and then just sort of come in, you know, on the other side of the chasm when the early adopters are like, well,
22. 4:11this AI stuff supposed to deliver value in my life in two years or so. That's it'll just be that. It'll just do cool
23. 4:18[ __ ] for you. But if you want to be in early, you got you got to just deal with the [ __ ] You got to deal with [ __ ] like
24. 4:26prompting was a big deal and now it's loops. Just the deal. That's the deal. Um, so I want to there there's there's two
25. 4:38pieces that I want to read that I want to start with and then and then we'll see how far we get because they they might take a while. Um,
26. 4:52but there's an absolute shift happening and and one of the pieces we're going to read is a piece that OpenAI put out
27. 4:58today and they talk about there being three phases of this AI stuff and that we're entering the third phase. Phase two for them was AGI.
28. 5:14So they say we're entering phase three. So that implies something has shifted. Kelly Camp. Hey Greg. Oh, say hello to
29. 5:24my childhood friend Greg. Oh my god. Welcome. That's so cool. We we said hi to him earlier. I didn't know he was
30. 5:28your childhood friend. That's so cool. Welcome. Welcome. Welcome. [clears throat] So, we're going to go read those two pieces, and then
31. 5:40I'm going to do my best to put things in perspective as we read them. And then we'll go look at we'll go look at the Peter Steinberger post
32. 5:54on uh on you shouldn't be prompting, you should be looping. And then I'm going to show you I I I made a I made a looping thing today.
33. 6:13By the way, any of you who've been playing with any of the agentic things like Open Claw or Codeex or Claude
34. 6:18Co-work or any of that [ __ ] you're already doing this. So, so it's not that big a deal. Okay, that said, let me go find some posts.
35. 6:35Matt Van Horn, you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents. What the [ __ ]
36. 6:43is a loop? So, we're going to go look at that one. We'll do that later. I think I think we'll do that later because
37. 6:53I think everybody wants to know what the loops are all about. [laughter] We're going to do that later. So, that's a that's a cliffhanger.
38. 7:07Brandon, here's this. There's that. Okay, so this was a post from June 4th. This was four days ago. I'm not sharing my screen.
39. 7:24Brandon's probably like, "Why is he not sharing his screen?" And I don't know why Brandon talks like that. Okay.
40. 7:39So from four days ago, Anthropic. So if you don't know the history, Anthropic was founded by Daario Amade and his
41. 7:55sister Daniela, I think her name is Daniela Amade. And they both worked at OpenAI and or maybe just Daario did. I don't
42. 8:06know. One or both of them worked at Open AI. They didn't like how OpenAI was dealing with safety and so they started Anthropic.
43. 8:16And so right now you could argue that Enthropic and Open AAI are the two kings or queens of the hill.
44. 8:27Google's a close second. Grock is a kind of or a close third, right? Grock is kind of a distant fourth. And then Meta
45. 8:35is not even really in the game right now. And then Apple sort of threw its hat in the ring today, but I think
46. 8:41Apple's going to play on a different level. But but the for the core frontier model companies, it's really OpenAI and Anthropic. Um
47. 8:52SpaceX filed its S1 for its IPO. Open AAI filed theirs today. I don't remember if Anthropic has, but I think
48. 9:01they've announced that they're going to. So all three of the the the top three are going to go public. So, they're
49. 9:07going to be incredibly wellunded. This is I saw I saw a rumor that there was a big dip in the stock stock market. It's
50. 9:15like everyone's, you know, getting their getting their uh their cash liquid so that they can they can jump on these
51. 9:22IPOs. I have a feeling these are going to be incredible IPOs. So, as they're preparing to go public, these two pieces were written within
52. 9:30four days of each other. So, this is the one from four days ago [clears throat] from Anthropic For most of AI's history, humans drove
53. 9:40every step of its development cycle. But at Anthropic, we are delegating a growing share of AI development to AI
54. 9:48systems themselves, which is speeding up our work. Taken far enough and given enough compute, that trend points to an
55. 9:56AI system capable of fully autonomously designing and developing its own successor. This is called recursive
56. 10:05self-improvement. We are not there yet, and recursive self-improvement is not inevitable, but it could come sooner
57. 10:12than most institutions are prepared for. Using public benchmarks and previously unreported data from within Anthropic,
58. 10:21the Anthropic Institute is showing that AI is already accelerating the development of AI systems. To take just
59. 10:28one example, today anthropic engineers on average ship eight eight times as much code per quarter as they did from
60. 10:362021 to 2025. Right? So, so 2021 chat GPT comes out chat or GPT3 comes out 2022 chat GPT comes out on on
61. 10:50GPT3.5. And so we've been, you know, for all the time we've been using it, they're shipping eight times as much code this
62. 10:59year per quarter as they were, you know, for the four years previous. The technical trends discussed in this
63. 11:06piece suggest that AI systems are going to be much more capable in coming years. And I think I I think I read this piece um
64. 11:18Oh, am I not sharing this? Oh, unbelievable. Sorry about that. [laughter] Um, I talked about this piece on
65. 11:28was it last Monday? No, it couldn't have been last Monday. I think I talked about this on Friday. I went through this a
66. 11:33little bit on Friday, but I'm going to go through it again because I want to compare this to what OpenAI just wrote.
67. 11:38The technical trends discussed in this piece suggest AI systems are going to become much more capable in coming years. The trends have huge
68. 11:46implications. AI that can build itself would be a major development in the history of technology. One that can
69. 11:51bring enormous good for the world in science, healthcare, and beyond. But full recursive self-improvement also
70. 11:57might increase risks of human losing control of AI. Um if system if systems are capable of full fully building their
71. 12:05own successor, um the ways we secure them, monitor them, and shape their behavior, um all grow much more
72. 12:12important. And so so you know we used to live in a world where people used computers. Then they used computers
73. 12:18amplified by chat bots. Then those chatbots became agents. Then the agents now have these subworkers that do all
74. 12:26the work for them. And then we're moving to this age where the agent spins up a bunch of agents, builds a new model, and
75. 12:34then keeps improving it and keeps improving it and and likely is launching it, right? building it, launching it,
76. 12:41improving it, launching it. Um, major model releases for the first three years we did this were happening approximately every six
77. 12:55months. They're now happening approximately every six weeks, and we haven't hit recursive self-improvement yet.
78. 13:03Um, so there's some history they put in there. Um, the rate at which AI models improve is accelerating. the length of
79. 13:10tasks they can reliably complete on their own has been doubling roughly every four months up to an early trend
80. 13:16of doubling every seven months. Um Claude Opus 3 could comp complete software tasks that took humans four
81. 13:25minutes to complete. A year later sonnet 3.7 could do an hour and a half. A year after that claopus 4.6 managed a 12
82. 13:3412-hour task. Um, in 2027, AI systems could be capable of tasks that take a person weeks, right? It just goes and builds and
83. 13:46builds. Um, the same pattern appears on coding and research benchmarks. Benchmarks measure the performance of
84. 13:52models in a given domain and they're saturated when models achieve close to 100% performance. The SWEBench is a
85. 14:00standard test of real world software engineering. It hands a model an open-source codebase, a real bug report,
86. 14:07and asks it to write code, a code change that fixes the issue, and passes the project's own tests. Models have gone
87. 14:16from scoring in the low single digits to saturating the benchmarks in two years. Right? So, fixing broken software, it's
88. 14:24getting, you know, really, really good, at least in the benchmarks. The core bench tests whether a model can reproduce existing research. a
89. 14:32prerequisite for them to conduct original research. It gives AI it gives an AI model the code and data behind a published paper
90. 14:44and asks it to rerun everything and confir confirm it can replicate the paper's results. AI systems went from
91. 14:51succeeding at reproducing the results roughly 20% of the time in 2024 to saturating the benchmark 15 months
92. 14:58later. meter, which which runs the benchmark benchmark measuring how well the models can complete long durations
93. 15:05tasks, found that Claude Mythos preview could work for at least 16 hours and was in the upper end of what meter can
94. 15:13measure without new tasks, right? So all of the stuff that says, hey, if if you do this, AI will be really good. All of
95. 15:21those benchmarks are just getting saturated. Public benchmarks say a lot about the capabilities of these systems,
96. 15:27but they can't reveal the impact AI systems are having on speeding up AI development itself. For that, we need
97. 15:34direct evidence from within AI companies like Anthropic. Evidence from within anthropic. Building a frontier model takes two broad
98. 15:43categories of work. There is engineering, writing the code, standing up the infrastructure, and overseeing the model training. And then
99. 15:52there's research. Desi deciding what experiments to run, interpreting what comes back, and figuring out which ideas
100. 15:58to try next. Across both engineering and research, the picture is consistent. In engineering, Claude can be handed an
101. 16:07unspecified problem and figure out how to solve it. Humans supply the goal, but they no longer need to supply the
102. 16:13method. Right? That's huge, right? Humans supply the goal. This is it. And this is directly related to
103. 16:23loops. We're going to be talking about loops. Our job is going to be to supply the goal. But we no longer need to think about the
104. 16:33method. In research, Claude can already match or outperform skilled humans at executing a well specified experiment.
105. 16:41However, large performance gaps persist when it comes to Claude exercising judgment in choosing goals in both
106. 16:48engineering and research. That's the gap between AI today and a future system that could autonomously design its own
107. 16:54successor. So basically what he's saying is kind of what we've all felt on a much lower level of of cognitive capability where even if
108. 17:07you're just doing simple writing or or you know code you know code development um we've experienced AI going from you
109. 17:18know barely usable to that's pretty good to okay it's mostly good if you know how you if you know what you're doing With
110. 17:25prompting, you can generally get these things to give you pretty decent results. We're not that far from us not
111. 17:33being required to be there, right? That happened in in coding in December 2026 where a lot of coders were like, "Oh, it
112. 17:41can just code now. I'm now doing something different." [snorts] So with the research, human supplying the
113. 17:49goal, this is the thing that AI isn't currently good at. But it sounds like we're pretty [ __ ] close, right? The gap between AI today
114. 18:00and a future system that could autonomously design its own successor. If if they're putting this here, it
115. 18:06means it means we're close. It's common for employees at Anthropic to receive more open-ended and important tasks as
116. 18:13they gain more experience. Early on, they execute a task someone else specified, like the export button isn't
117. 18:19working. Please fix it. With experience, they're handed a goal and design the approach themselves, such as investigate
118. 18:26why the network slows down under heavy load. At the most senior levels, they're deciding which problems are worth
119. 18:33working on at all. What should the team build next quarter? We can all use internal anthropic data to see how far
120. 18:40Claude has come in being able to handle these different kinds of tasks. Claude writes, "A significant proportion of
121. 18:47Anthropics code. As of May 2026, so last month, more than 80% of the code we merge into anthropics codebase was authored by Claude." 80%.
122. 18:58Before Claude Code launched in a research preview in in February 2025, this number was the low single digits.
123. 19:04So in six months, oh no, in in a year and three months, in 15 months, they went from low single digits to 80% of
124. 19:14code being written by Claude. So again, things are accelerating. So even that amount of time is is going to start shrinking.
125. 19:26That shift also shows up in the amount of output per engineer. lines of code merged per engineer per day, stayed
126. 19:33constant through Enthropic's first four years, 2021 to 2024, then began to climb upwards in 2025 when Claude began to
127. 19:42write run code rather than just suggesting it um for an engineer to copy paste. Right? So this was so clouded
128. 19:49code where it could get into the terminal and start running its own code started to up the game. The slope
129. 19:56steepened again in 2026 when models began to work autonomously over long periods of time. These two inflection points are shown in the chart
130. 20:06below. In the second quarter of 2026, the typical engineer was merging eight times as much code as they were in 2024.
131. 20:15So here here you go. Look how flat this is. So this is actually a really good example. David Shapiro talked about this
132. 20:22a lot with AI. Oh, you know what I want to do? I'm going to pop out my comments. So, if you guys are having comments, I'm
133. 20:27not looking at them right now. All right. Beautiful.
134. 20:41Beautiful. Beautiful. Beautiful. Beautiful. Where is my sharing? What did I lose? Did I lose it?
135. 20:57Oh, here it is. We're just stuck in loops. You're good. Loops in the comments. Oh, loops. Loops. Loops. [laughter]
136. 21:16I'm not feeling like I'm much of a human in the loop these days. Well, this is I want a cinnamon roll other kinds of
137. 21:25loops. Um, but look at this. I I want you to look at this chart. So, David Shapiro talked about um working on an exponential and and
138. 21:35working in uh machine learning that the history of machine learning is like nothing happens, nothing happens,
139. 21:42nothing happens and then all of a sudden all of a sudden something starts happening. So this is start starting to
140. 21:48be what's happening with code right. So Q4 right here um where it's two and a half times the amount of code internally
141. 21:58at um at anthropic they're they're kind of it's it's increasing but it's kind of this linear increase right it's it's
142. 22:06it's a linear increase um December of 2025 right when they're they're internally doing two and a half times code they drop Opus 4.6.
143. 22:22And that was the model that all of the developers all of a sudden said, "Oh [ __ ] this thing can really code." And
144. 22:28then look at the next quarter. They jump from 2.5% to 5.8%. Almost almost a 2x gain in one quarter. Um and and you
145. 22:38know, this thing's starting to increase. So if these models are getting better exponentially, you know, this this
146. 22:45basically just goes through the roof. Okay. Um, a caveat, lines of code is an imperfect measure as it measures
147. 22:54quantity over quality. So, eight times the lines of code per engineer per day in the second quarter of 2026
148. 23:01is almost certainly an overstatement of the true productivity gain. Nonetheless, it indicates an acceleration and
149. 23:07enthropic. We don't report reward people for how many lines of code they write. Rather, team members are producing more
150. 23:14code simply because they're using AI systems to write more code. The increase in lines of code written lines up with
151. 23:20the subjective impression of large productivity increases. In March 2026 poll of 130 employees from across
152. 23:27anthropic research teams, the median respondent estimated that they produced around four times as much output with
153. 23:34mythos preview as they would have without access to any AI models on the kinds of projects they would have been
154. 23:42working on. Regardless, we expect that the true degree of uplift in March was somewhat lower. Nonetheless, we find
155. 23:49overall claim plausible and in line with our other observations. A significant fraction of Anthropics technical staff
156. 23:57is accomplishing their core work multiple times faster than they could without AI assistance. And listen,
157. 24:04there's they're selling anthropic here. So they're of course they're going to say this it's it's increasing productivity but
158. 24:14all of these tools are being optimized for code because these frontier model companies want to beat the other
159. 24:19frontier model companies. So these engineers are going to see the the the results sooner because they're all in on
160. 24:26this stuff and they're using the tools that they're building. We also see evidence that people at Enthropic are
161. 24:32using Claude to do work that they simply wouldn't have hap that simply wouldn't have happened otherwise like building
162. 24:37exploratory tooling and addressing long deferred cleanup. For example, in 20 April 2026, Claude shipped over 800
163. 24:44fixes that reduced a class of API errors by a factor of a thousand. Right? 800 fixes in April in in one in
164. 24:55one month. 800 fixes reduced a class of API errors by a factor of a thousand. The engineer overseeing Claude estimated
165. 25:03that a human would have taken four years to complete this work. Solving people's bugs is slow and painstaking and humans
166. 25:10struggle to hold that much unfamiliar context in their head at once. I've experienced this with our tech team at
167. 25:15Story Mine. It's like once you get to a certain level of complexity of the codebase or bugs that tap into multiple
168. 25:25pieces of the codebase, the the progress in squashing those bugs drops dramatically. I started leaning hard
169. 25:32into clottifying about a year ago. That's been a crazy adventure and now it's been approximately five months
170. 25:38since I lost wrote any code myself. If you're working at Anthropic, you probably loved writing code. The code that Claude writes is good and
171. 25:54improving. Good code means two things. It works and it's written in a manner that allows another engineer to
172. 25:59understand it and build upon it, which I don't think that particular criteria is all that important. Too much math. Damn it. No.
173. 26:11What are you guys talking about? Okay. Um, if if agents start writing all the code,
174. 26:26there's going to be a finite amount of time where it's actually valuable for a human to be able to read and interpret
175. 26:32it because where it ultimately goes, this is this is wild. So, where computer coding started out was in binary. people
176. 26:43literally writing ones and zeros, right? And then they had very low-level languages and then they had higher level
177. 26:49languages and then high level languages and then scripting. They had, right, we we have code that got increasingly more
178. 26:56friendly for humans to to be able to read it and write it and share it. And I think what's going to happen is
179. 27:03we're going to basically go run right back down that because the more human readable code is, the farther it it is away from the ones
180. 27:13and zeros of how computers actually work. So the most efficient code is no code at all. It's just ones and zeros,
181. 27:21right? So that's probably where we're heading. Elon Musk just said something about that. So that's a really interesting one. Okay. On first
182. 27:29criterion, the evidence is clear. The rate at which anthropic staff is staff correct, redirect, or take over midtask
183. 27:37from Claude has been falling steadily for a year, including the most complex and open-ended tasks. This means that
184. 27:45with no clear specification where the engineer isn't sure what the answer should look like, this is evident in
185. 27:52Claude's success rate over time. And then so so basically Claude's success rate I see open-ended problems is going up, right? They're all
186. 28:02sort of trending toward 100%. So with these open-ended problems, if you're in the whatever this is 11% range, you're going to have lots of
187. 28:13human involvement. As you get up toward 80% and higher, there's a lot less human involvement in those problems.
188. 28:22Um, on the most open-ended tasks, Claude's success rate jumped reached 76% in May of 2026, up 50 percentage points in six months.
189. 28:32To give an example of tasks in this difficulty tier, a routine upgrade began crashing tens of thousands of training
190. 28:39jobs. An engineer pointed Claude at the live incident with little more than some text content and cluster access. Working
191. 28:49through the running jobs and testing one environment setting at a time, Claude isolated a single obscure debugging flag
192. 28:56that was triggering the crash, reproduced it reliably, and confirmed a fix. In about two hours, Claude
193. 29:03delivered what normally would have been two or three days of work. That's pretty cool. The second criterion was writing
194. 29:09is writing code that another engineer can understand and build on. Here the gaps between humans and AI persists but
195. 29:16is closing fast. There isn't full consensus among staff at Anthropic but many believe that claude written code
196. 29:23was still worse in human worse in quality than human written code at Enthropic in late 2025. It is roughly at
197. 29:31parody today. We expect it to be better within the year. And again, I think this is a short-term requirement that it
198. 29:40write code that humans can understand because I think what happens is right now we think about writing code as
199. 29:48building these systems that are going to be persistent and live forever. I think most code in the future is going to be
200. 29:54dynamically generated. All right. Oh my god. It's literally the matrix. It is literally the matrix is where is
201. 30:03where this is headed. It's ones and it's raining ones and zeros because that's ultimately where where if if if
202. 30:12this thing is going to get more and more and more efficient and it's going to start training itself, it's going to collapse the gap. It's
203. 30:21it's going to say, "Well, how do machines work?" Well, machines work with all this binary ones and zeros. Why
204. 30:28don't I just go down there and just write things directly into the into just pipe it on in. Yeah, exactly. I'm
205. 30:37telling you, this is why I'm reading these is because we're we're very very close to the I don't know to the beginning of all this.
206. 30:50All right. This has changed the way Enthropic now reviews its own code. proposed changes to our codebase are now
207. 30:56read by an automated cloud reviewer that looks for bugs, security flaws, and other defects before it can merge. Using
208. 31:04this tool, we ran a retrospective analysis and found that an automated claude review of every change to our
209. 31:12codebase would have caught roughly a third of the bugs. A third of the bugs behind past incidents on Claude AI before they
210. 31:23reached full production. The engineers who wrote that code are among the best in the world at building these systems.
211. 31:29Claude is now catching the mistakes that they missed. Um, Claude written code was somewhat worse than human written code in late
212. 31:362025. It's roughly at parody today. Okay. [clears throat] Claude is good at running experiments to hit a goal that
213. 31:43someone else has set. Every time Anthropic releases a model, we run the same test. We give Claude some code that
214. 31:50trains small AI model and we ask it to make the code run as fast as possible while still passing the same correctness
215. 31:58checks. The goal and the success metrics are fixed in advanced. So Claude's job is to find speed ups by rewriting the
216. 32:06code, running it, timing it, and repeating. It's a miniature version of an experimental research loop. A
217. 32:13research loop. Look at that. See the loops? You see what we're doing, people? We're heading to loops.
218. 32:24In May of 2025, Claude Opus 4 averaged a three times speed up over the starting code. By April 2026, so a year later,
219. 32:33Claude Mythos preview was achieving 52 times speed up. For calibration, a skilled human researcher would need 4 to eight
220. 32:43hours to reach 4x improvement. In this in this part of the research workflow, optimizing steps within a
221. 32:52clearly defined experiment, Claude has gone from super helpful to Okay, look at this sentence. Claude has gone from super helpful to
222. 33:05superhuman in under a year. The shape of stuff today is roughly. Humans have ideas and the models are able to implement, test, and evaluate
223. 33:16them. an order of magnitude faster than before. Okay. Claude is getting better at proposing its own experiments. In
224. 33:23April 2026, Anthropic published first demonstration of Claude running an open-ended research project end to end.
225. 33:32Claude powered agents were given an open problem in AIA safety. Roughly, can a weaker model reliably supervise a stronger one?
226. 33:44And we're left to solve it. This involved proposing hypothesis testing them, sharing findings with parallel
227. 33:50agents and iterating. The task has a clear performance floor and ceiling. The floor is how well the weak supervisor
228. 33:57would do on its own. The ceiling was how strong the model does when it's trained on the correct answers. Two human
229. 34:03researchers over a week recovered roughly 20% 23% of that gap. The agents recovered 97% over 800 cumulative hours. Wait, so over
230. 34:17about a week, two human researchers got 23%. The agents recovered 97% over 800 cumulative hours
231. 34:31and used roughly $18,000 in compute. There are some caveats to this work. The result didn't transfer cleanly to
232. 34:38production scale models and humans still chose the problem and created the scoring rubric. But within those bounds,
233. 34:45the agents designed every experiment themselves. Direction setting was the only meaningful s role a human played.
234. 34:53Meaningful role a human played. Claude did all of this with pretty minimal help from me over the course of one to two
235. 34:58days. I think if a junior colleague came back to me with the results like this in the same span of time, I would be mildly
236. 35:04impressed. The future is now geeks. Okay. Claude is getting better at steering research. We examined real
237. 35:11Claude code sessions where anthropic researchers were working with Claude in an open-ended investigative problem like
238. 35:18figuring out why training run keeps crashing. Um, a separate claude. Okay, let's see. Because we deliberately picked moments n
239. 35:26equals 129. So, wait, let's see. A separate claude was able to see how the session eventually turned out, then
240. 35:34judged whether the AI or the human suggested the better next step because we deliberately picked moments
241. 35:41129 where we know the human's choice had room for improvement. This isn't like like for-like comparison between model and human judgment.
242. 35:51What these moments give us is a set of realistic challenging situations where the right next step is not obvious and
243. 35:57where the human's choice serves as a useful yard stick to compare model performance over time. On the me on this
244. 36:04measure, our best model in November 2025, opus 4.5 beat the human choice 51% of the time. In April, so five months later, this
245. 36:16grew to 64%. The day-to-day work of research is largely a chain of these nextstep decisions, making this a relevant measure of the
246. 36:26model's ability to eventually run an investigation on its own. We view this result as an early signal that AI
247. 36:32systems are getting better at making the kinds of judgment calls that AI research depends on. All right, so basically shit's getting
248. 36:41better. The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking
249. 36:48behind the confines of the immediate task. So this is sort of the last [laughter] the the last highlevel human thing where
250. 36:58humans are clearly at an advantage is is getting thinner. What might be the future of work? What might the future of work at Anthropic
251. 37:09look like? Okay. So, let me just pause for a second. Let me look at comments here. Ones and zeros. Okay, we see the
252. 37:17loops. Okay, there's no comments. So, I put you all to sleep. Perfect. It's exactly what I was going for. Um,
253. 37:29so all this is to say like buried in here is is the speed of improvement and they're and they're giving some time frames, right? 2021
254. 37:42to 2024. So over a three-year span, roughly nothing happened. And then over the course of 2025, it started to get
255. 37:52pretty good. And then the first half of 2026, it's starting to get really good. Right? So this article is time is compressing and quality is
256. 38:03improving quickly. So what might the future of work at anthropic look like? What's valuable in looking at these labs
257. 38:18is that if you imagine every single job that involves a computer is going to be after what these engineers experience. So as these models
258. 38:31get better, it's going to have a ripple effect through all jobs. So what they're saying about the future of jobs here, we
259. 38:38should pay attention to the evidence suggests that the human role is narrowing at each step in the AI development process. Once human and AI
260. 38:46authored once human and AI authored code reach parody, humans will stop writing code entirely and shift to only reviewing it. But if
261. 38:58they can't write code as quickly as COD cla can generate it, human review will become the bottleneck. Right?
262. 39:10So if it starts writing all the code and humans just review it, but if they can't keep up, then humans become the bottleneck. And I
263. 39:18I think this is what we're going to increasingly see. Humans become the bottleneck. Similarly, once Claude can
264. 39:23run experiments, the question shifts toward which of these experiments is worth running. Put simply, the doing writing the code, running the
265. 39:32experiment, producing the result now costs almost nothing in human time, even if it still has costs and compute. An
266. 39:38area of human comparative advantage for now is research, taste, and judgment. Right? This is the Rick Rubin thing. We
267. 39:47need to have taste and judgment. So that's still valuable, but but getting less so, including choosing which
268. 39:53problems to master, which results to trust, and when an approach is a dead end. Work and life ran on a gift economy
269. 40:00of small favors between humans. Can you help me get this script running? Each one created a little debt, a little
270. 40:08mutual awareness. Claude is faster and it creates zero debt, but each of these is a lost bid for human collaboration.
271. 40:16So that's an important line right there. So human collaboration remains interesting. So they're saying if the
272. 40:23machine's just going to do everything, we lose this really interesting thing. On days where everything works well, I
273. 40:29can't help but think nothing I do matters. Everything is automated and better and faster than I will ever be.
274. 40:35But then there are those days. We've all had these days where everything breaks and I don't understand why and I realize
275. 40:41I have no idea what I've been up to anymore. What if we're wrong? Okay. What do we got here, Chef Kelly? Um, it will
276. 40:48become like hierogly like hieroglyphs or cursive. Oh, you mean code and reading code and all that sort of stuff? Yeah, because again,
277. 40:58I think these machines just they they will evolve to stop needing to include us. And and I think we will evolve to stop needing to care,
278. 41:09right? If you at the Google IO conference, they showed that example of the future of search where the search
279. 41:16realized that you've been searching for activities for your daughter's birthday or whatever it was and it said, "Hey,
280. 41:21would you like me to create a trip planner for you specifically for the thing you're researching and it
281. 41:27dynamically wrote an application for the for the searcher?" Right? So, like I think we go to just a a much more
282. 41:35dynamic world of code. Um, we're not anywhere close to that yet, but that's where this is pointing. Okay. What if
283. 41:43we're wrong? A natural objection to the evidence presented above is that the work is that wait is that the work that is
284. 41:52still in human hands, choosing which problems to work on is what matters most. Right? Without that judgment, Claude is a
285. 42:02capable assistant, but not a system that could drive AI progress on its own. It is genuinely unclear whether today's
286. 42:09training methods and architectures could unlock that capacity. So they're basically saying the transformer
287. 42:16might be might be a bottleneck here like we might not escape this this limit that we have right now about judgment and
288. 42:24taste with the current thing. But AI is rarely advanced by Eureka moments. There have been a few of those in AI's recent history like the
289. 42:33transformer architecture, right? That's what we just talked about. Or mixture of expert models or paradigm shifting
290. 42:38ideas. But but paradigm shifting ideas arrive years apart. In between [clears throat] most progress is
291. 42:45incremental. We scale something up, see what breaks, fix it, try it again. This is exactly the kind of workflow Claude
292. 42:52now excels at. Edison said that genius is 1% inspiration, 99% perspiration. But we see perspiration becoming increasingly automated.
293. 43:03It's becoming clear that much of what advances the frontier is automatable. Largecale research progress is mostly a
294. 43:10function of tools and resources which dictate how fast you can run experiments. Even if we suppose Claude
295. 43:17never achieves good good research taste, a conservative reading of our evidence still implies compounding acceleration.
296. 43:26So, if you haven't read um Andy's piece on Substack, go to the AI salon and go to community chat. I put a post
297. 43:38in there. I don't know where it is. It's the most recent post. Um and go read that article. She talks about compounding uh
298. 43:50compounding acceleration in there. Um where's my little tab? There we go. If humans spend most of their time on on
299. 43:58the singledigit fraction of work that is direction setting setting while cla handles the rest, this means that each
300. 44:06engineer or researcher is steering far more work than before. The evidence we see suggests that people at anthropic
301. 44:12are both moving faster and covering a broader surface. Right? So they're becoming more valuable. They're not as narrow. They're
302. 44:22moving faster. In practice, this means that AI already makes anthropic move much faster than it did before
303. 44:29um the advent of effective AI tools. The less conservative reading is that early evidence on clouds improving research
304. 44:36judgment, narrow as it is today, is an indicator that this capability is improving as well. Research taste might
305. 44:43just be another AI capability that AI systems fail at for a time, then get good at, right? So nothing, nothing,
306. 44:51nothing, nothing, and then all of a sudden it's great. We've seen a similar pattern with um other qualitative skills
307. 44:56like AI systems being able to explain why a joke is funny, demonstrate theory of mind, or solve linguistic riddles. Possible futures.
308. 45:06One, the trend stalls, but today's AI capabilities are widely diffused. I think that's decently likely. Two, AI
309. 45:14labs continue to see compounding efficiency gains. Okay. Three, AI systems themselves become capable of full recursive
310. 45:24self-improvement and begin building their successors. [clears throat] Is that the last one of these? Yeah. Let's read this one.
311. 45:36If technical trends and advancing capabilities continue and AI systems are able to develop capabilities inherent to
312. 45:44transformative human ingenuity, then it is plausible that AI systems could design and refine themselves in
313. 45:51this world. Okay, this is this is the paragraph that that I most want to sort of talk about. In this world, the pace
314. 46:01of progress in AI development becomes determined entirely by the availability of compute. Right? So, this this gets back to the
315. 46:11reason I've been pushing people to understand local large language models and and understand running your own um
316. 46:20AI compute is that I think that as these systems get more and more powerful, they're going to dedicate most
317. 46:28of the compute to these things improving themselves and they're going to ration compute to the rest of us, which means
318. 46:36running AI is going to get really expensive. ensive unless you're running it locally, unless you're running it on your own machines.
319. 46:43Um, humans [clears throat] play a sub substantially diminished role in the development of AI models like moving
320. 46:50most of our efforts toward oversight and validation and verification of an expanding virtual lab run by AI systems.
321. 46:58We expect we expect that systems capable of automated AI research and development would have skills that would transfer to
322. 47:05the rest of science allowing them to revolutionize other fields. So basically what this is saying is if we build a
323. 47:14machine if we build an AI model that can start building new AI models then you can basically take that method and just
324. 47:24transfer it across science. how the alignment problem gets solved or not in this future is something we are the least certain about. So the
325. 47:36alignment problem is articulates and this is one of the the reasons anthropic was created is how aligned is the AI's imperative to human
326. 47:51imperatives. And if they're not aligned and we have this thing that's way more capable than us then we're in trouble. Um,
327. 48:01okay. Models could improve could prove to be sufficiently aligned and capable enough of research taste
328. 48:08that they discover and implement novel solutions we've not yet re yet reached. They could also be sufficiently wise to
329. 48:15halt the development if not. Alternatively, the rare occurrences of misalignment present in today's models
330. 48:23could compound as models build their successors, growing more frequent and less understood until we lose control of
331. 48:30them. It's possible that we can't build, integrate, and verify the tools that we'd need to understand the trend line
332. 48:37that we're actually on. Yeah. like the the AI very much has the possibility of creating a version of itself that we
333. 48:44actually don't know what it's doing and and how to deal with it. Um we're not we do not have good institutions for what
334. 48:54this world would look like. Chef Kelly. In essence, it will become a thing like the internet was a thing
335. 49:12like the modem is because you will need to run it locally on your computer in order to have it at your will. Yeah. Yes. Yes. Yes. Okay.
336. 49:25Okay. You know what I'm Okay. Even if model development became fully automated and recursive, we can't predict
337. 49:34what that would mean for humans daily lives. Amdall's law applies here as well. Recursive intelligence could leave
338. 49:41lead to achieving many of the benefits outlined in machines of love and grace quickly in some domains. We expect that embodied intelligence
339. 49:52robots might quickly follow recursive intelligence and follow a similar path for increasing returns and decreasing cost.
340. 50:00More powerful intelligence blah blah blah. Okay. By achieving recursive self-improvement alone does not suggest
341. 50:06an immediate change in industrial production. Okay. So they're just doing some caveats here. What should we do? If
342. 50:13it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications,
343. 50:21we think that would likely be a good thing. [laughter] That's the [ __ ] weasleyest sentence I've ever read.
344. 50:33If it were possible to effectively slow down the development of this technology to give ourselves more time to deal with
345. 50:40its immense implications, we think that would be a good thing. But if a slowdown simply lets the least cautious actors catch up
346. 50:50technologically, it could leave everyone else less safe. So they're basically saying, "Let us keep building so that
347. 50:58you know, China or Elon Musk or whoever doesn't build something that just goes wild." Like, "Let us build something
348. 51:05that's going to combat that." [clears throat] Without a global coordination mechanism, that's an interesting statement.
349. 51:12Companies and governments will have to make difficult decisions about safety while under the competitive and while
350. 51:18while under competitive and geopolitical pressures. This We're [ __ ] living in science fiction. [laughter] [snorts]
351. 51:32This is a trip. We We believe it would be good for the world to have an option to slow our temporary slow or temporarily pause
352. 51:42frontier AI development. This is what Elon Musk asked for two years ago, three years ago, and then he said, "Fuck it."
353. 51:50And said, "I'm going to go build one, too." um to enable societal structures and alignment research to keep up with
354. 51:56the advance of technology. What's wild about this is [clears throat] you could slow AI development, but
355. 52:06enable societal structures and alignment research to keep up. The societal structures piece like what we found out with COVID was we
356. 52:18have to have an acute threat, an acute real time like you know if we don't stay inside people are going to die. If you
357. 52:26don't write checks for people, the economy is going to stop, right? There was a there was a distinct
358. 52:33like global acknowledgement that we have to do something. So if you slow AI research and it just sort of trickles job loss out,
359. 52:47you might not there might not be enough political cover to make big changes. >> [clears throat] >> So that's that's actually really
360. 52:54fascinating. The anthropic institute will conduct research in collaboration with many others, that'll be worse. Yeah, exactly. If it
361. 53:04if they slow it down and it just sort of trickles out and everyone thinks everything's okay, but it's really it's
362. 53:10really just building this pressure, then it's going to be really ugly. These systems would enable Frontier AI
363. 53:16developers to verify that others okay. A meaningful slowdown or pause would require multiple well-resourced labs at
364. 53:22or near the frontier in multiple countries agreeing to stop under the same conditions. So there's a there's a call to the to
365. 53:33the global technology community right there. It would also require that each can verify that other others have
366. 53:40actually stopped. Hey, look, we're in the Cold War again. Are are you sure you're not building nukes? Well, we're
367. 53:46not building nukes. Are are are you building nukes? Uh, no. I mean, these thousand that you found behind me, those Well, those were ones
368. 53:58we were building. I mean, we're building a little bit. Sure, because you probably are, too. [laughter]
369. 54:08I have to go turn my brain off tonight, y'all. Good night, Andy. Um, [clears throat] [laughter] okay. Let's see. Um, a meaningful
370. 54:18slowdown. It would also require that we can each verify each other. Due to the unique characteristics of AI systems,
371. 54:24the the detectability, a lower standard than verifiability element of this arms control problem, is much more challenging than with other
372. 54:32technologies. We can we can see your radiation from space. Training runs are far easier to conceal than missile silos, and their
373. 54:41inputs are general purpose. And the incentive to def defect quietly is enormous because whoever continues while
374. 54:50the other's pods will inherit the lead. Here you go. There's there's your capitalism slashgovernment, you know, um, [clears throat] safety
375. 55:06piece. A credible pause also has to specify what triggers it, what lifts it, and who adjudicates. None of this is
376. 55:12necessarily impossible in principle. In the coming months, we will organize conversations with policy makers,
377. 55:19researchers, civil society, and other AI companies to help address some of the questions this piece raises. Okay, so big long piece.
378. 55:30Let me just come back here for a second and just check in with you all. What are your thoughts? What are your thoughts?
379. 55:36If you're still here, if you outlaw AI, well, yeah, I mean, [snorts] like that's that's a place this goes is
380. 55:48that the government says, well, only the government can have AI, right? Then the then only the outlaws will have
381. 55:58AI, right? Well, but again, this is a reason why um you know, getting your hardware in place now, understanding
382. 56:06what local models are now um and getting them on a hard drive and on backup hard drives probably not a bad idea.
383. 56:14Um all right, I want to go look at another thing or rogue nations. Yeah. Yeah, that too. Okay.
384. 56:51All right, y'all can still see this, right? >> [clears throat] >> There we go. Okay. So, this was put out by this is shorter,
385. 57:05I think. Um, this was put out by OpenAI today. So, the anthropic piece says things are accelerating. The real advantage that humans have
386. 57:21right now is not the tactical [ __ ] It's taste. It's deciding which research we want to pursue. Humans are still better
387. 57:31at that, but that margin is shrinking. Open AI files their S1. They say we're going to go public. Dr. J comment Claire,
388. 57:43only a good guy with AI can stop a bad guy with AI. [laughter] Built to benefit everyone. Our plan. >> [clears throat]
389. 57:56>> Every few generations, a new technology changes everything. Imagine electricity reaching a rural American town in the 1920 20s. Before
390. 58:05the power lines arrived, daily life was shaped by physical limits, hauling water, washing clothes by hand,
391. 58:12preserving food with ice, and ending much of the day when the sun went down. Electricity did not transform every
392. 58:20household overnight, and many of its benefits reached people unevenly. Right. And so that's absolutely going to happen
393. 58:26with AI. My seven economies idea is very much about jagged adoption of AI. Um but as access spread, ordinary life changed.
394. 58:37Light at night extended the day. Electric electric pumps, appliances, and refrigeration reduced some of the
395. 58:43hardest daily work. Radios brought news, music, connection from hundreds of miles away into h homes and community spaces.
396. 58:52The first promise of electricity was practical, but its deeper impact came from new possibilities that opened up as
397. 58:58more people could use it. With time, a lot of new possibilities. I want to go back to this sentence. The first
398. 59:05possibility of electricity was practical. But its deeper impact came from new possibilities. In the AI salon, we talk about play first.
399. 59:17Why do we talk about that? Because what most people do with AI is they treat it as a practical improvement. The promise of electricity was
400. 59:28practical. Oh, we've got this powerful AI tool. Let me take this thing that I already do and make it more efficient,
401. 59:36right? This is how we've always used computers. But AI is way more than an efficiency machine, right? Its deeper impact like
402. 59:46electricity comes from the new possibilities it opened up as more and more people could use it. This is the
403. 59:52more and more people as more people could use it is people like the members of the AI salon who are not technologists who are
404. 1:00:02not process engineers right who are just people discovering new possibilities of how we can use it. With time, a lot of new
405. 1:00:13possibilities emerged with machines and computers greatly accelerating progress in medicine, engineering, and many other
406. 1:00:20fields. By the end of the 20th century, the average lifespan had increased by over 20 years, and the median inflation
407. 1:00:27adjusted income tripled or so. These gains were driven in no small part due to due to the advances in health care,
408. 1:00:34sanitation, living standards, many of which were enabled or accelerated by widespread electrification and related
409. 1:00:41technological progress. This is happening again with AI. AI will soon be capable of extraordinary things, but the point is not the
410. 1:00:50technology itself. The point is what people can do with it. Right? And again, this goes back to One of one of the questions that I've
411. 1:01:01been asked before is why do we need all these [ __ ] data centers? Well, if the capabilities are increasing and
412. 1:01:11people want to maximize those productivity gains, they're going to pay more for those tokens, for the for the
413. 1:01:19compute. If there's not enough compute, if there's not enough power, if there's not enough chips, then those high value
414. 1:01:29players that are going to demand that compute who who are also willing to pay for that compute means that there won't
415. 1:01:37be a lot of compute left for the rest of us. Right? So, this is the this is the weirdass place we are right now.
416. 1:01:44[clears throat] One of the weirdass places we are. The point is what people can do with it. Um, it can help people navigate a
417. 1:01:51medical bill. Oh, wait. This happening. Okay. AI will soon be capable of extraordinary things. But the point is
418. 1:01:56not the technology itself. The point is what people can do with it. It can help some people navigate
419. 1:02:04a medical bill, learn a new skill, start a small business, care for an aging parent, understand a legal or financial
420. 1:02:13decision, turn an idea into something real, or make a scientific discovery. Hang on a sec. While the wonder of light at night
421. 1:02:26probably wore off pretty fast, what people decided to do with it did not. And because technology has become a
422. 1:02:33reliable way to deliver prosperity over time, we think AI should be available to everyone to use as much as they need,
423. 1:02:41where and how they need it. We think AI should be available to everyone to use as much as they need it where and how
424. 1:02:50where and how they need it. I don't know how that's possible if you don't have enough compute. Does this mean that OpenAI is going to
425. 1:03:04be compressing their big models so that they can run on desktop machines? Maybe [cough and clears throat] that future will not happen
426. 1:03:12automatically. Transformative technologies can concentrate power or they can broaden it. So, it's interesting they're saying that out loud
427. 1:03:20as they start concentrating power. [laughter] Um, they can make life easier for a few or they can expand opportunity for many.
428. 1:03:29Our approach is rooted in the belief that AI should work for people, helping them pursue their own goals. increasing their capabilities and
429. 1:03:40distributing the benefits of this technology as widely as possible. Chef Kelly HJ Hines was the first to bring
430. 1:03:47electricity into his factory, which created a sauce called ketchup to mask the taste of spoiled meat. This was the
431. 1:03:55third attempt at entrepreneurship. He failed miserably prior to building the Hines Empire. That's amazing. Does that
432. 1:04:02mean we should give the teachers AI? Yes, it does. [clears throat] Okay. So, AI, our first commitment is to build AI in service of humanity.
433. 1:04:18That means we want to empower people broadly, not see power concentrated among a few companies, governments or
434. 1:04:25individuals. Well, this if I were Elon Musk, I'd read that and say, "Well, then why don't you open source it, you big fat fatties?
435. 1:04:33We believe that the safer future is one where power is broadly distributed so more of the world can participate in
436. 1:04:39building a relian a resilience ecosystem with an asterisk. I don't know what a resilience ecosystem is but maybe
437. 1:04:47there's something at the bottom. We're optimistic about AI because we believe it can expand human capability and prosperity.
438. 1:04:55But we are also cleareyed about the risks. Powerful systems must remain safe, aligned with human intent, and
439. 1:05:03subject to human control. Our mission at Open AI is to ensure that AGI benefits all of humanity. That means building
440. 1:05:11systems that help people do more of of what they choose, not systems that replace human judgment about what matters. Look at this. that
441. 1:05:22sentence. I wonder if that sentence was written in the past three days as as a response to Anthropic.
442. 1:05:31That's a that's a direct [ __ ] slap to Anthropic that we just read that Anthropic was like, "Listen, the only thing still
443. 1:05:40where we still got an edge here is our taste. Human judgment is is where it's at, but that that gap is closing
444. 1:05:46quickly, so we might not need humans anymore." And Sam Alman just said, [laughter] "We want to build systems that help
445. 1:05:54people do more of what they choose, not systems that replace the judgment about what matters." Well, there you go. All
446. 1:06:01right. [clears throat] So, they're so so the they're fighting. Okay. Entirely automating everything is not the future
447. 1:06:09we want. It would be unfulfilling and it would be dangerous. AI should help people pursue their goals, not become
448. 1:06:15untethered from them. There's the great repurpose right there. As AI systems become more capable, the human role
449. 1:06:22becomes more important. Setting direction, making trade-offs, applying judgment, and bringing values, taste,
450. 1:06:28care, and responsibility to the work. This, you could argue, is the is the mission of the AI salon, right?
451. 1:06:42Human centric. Let's use these tools to do all this stuff where we can bring our whole selves to using AI in ways that take the things
452. 1:06:54we want to do in the world and do them bigger, better, faster, cheaper. A key long-term role for people will be deciding what is
453. 1:07:03worth doing. There's your job. There's your Rick Rubin, right? That's our job in the future. I know this feels feels [ __ ] surreal
454. 1:07:14right now. You're like, "But our job, Isn't our job going to be to do the No, our job's not going to be to do the things."
455. 1:07:22You'll have things that will do the things. You can choose to do the things. You just don't have to do the things.
456. 1:07:32Well, I love Needle Point. Well, go do [ __ ] Needle Point. But if you want to start a needle point company to sell all
457. 1:07:39the pretty little needle point and you want to scale that, well [clears throat] then it's no longer a hobby, right? And
458. 1:07:46there would be machines for that. But your job is to decide, well, wait, why needle point? Why now? What's going to
459. 1:07:53be on it? What are you gonna How are you going to sell it? Is this pin cushions or pot holders? Mary Carol. So I don't really believe
460. 1:08:05that. Yes, companies pay millions for consultants to come in and give great advice, but in general, it's the work,
461. 1:08:12the output that has historically been paid across industries. I don't really believe that judgment is going to allow people to keep the jobs
462. 1:08:22across the vast majority of industries. I'm not buying it. I'm just saying that the outputs the outputs are going to be increasingly
463. 1:08:33valueless because the cost of doing them essentially drops to zero. But where the real value is, the judgment is in choosing what thing to
464. 1:08:49do. So, I don't know that we're we're disagreeing with that, but it's but the outputs are going to be increasingly commoditized.
465. 1:08:59How you do the outputs, how you choose to package them or frame them or which ones you choose in which order
466. 1:09:09because you have an intuition. That's going to become more and more valuable, I think. But we'll see. We believe that AI doing AI research
467. 1:09:20will become the determining factor of the pace of progress within the next few years. So again, this is this is right
468. 1:09:26in line with Ray Kerszswhile's singularity prediction of 2029. That matters because alignment itself is
469. 1:09:34a hard research problem. To make fast and deep progress, our researchers will need AI systems that can help test
470. 1:09:40ideas, find mistakes, explore alternatives, and iterate alongside us. the but faster technological progress makes human judgment and public
471. 1:09:50coordination more important not less. Right? So again this almost seems like a response to the anthropic piece.
472. 1:09:59The future should be shaped by people, institutions and societies, not by the companies building the most capable systems.
473. 1:10:08So that's that's a direct smack to to Daario. As frontier AI development continues, we expect national and global coordination
474. 1:10:18to become more important. We have long believed that there should ultimately be an international organization who helps
475. 1:10:24coordinatingly leading AI efforts to reduce catastrophic risk. This is what Sam and Daario seem to agree on.
476. 1:10:34So, so this is probably something that anthropic and open AI will come together on and and say something publicly is my
477. 1:10:42guess. Cooperation and shared safety standards are an important part of a new the path forward especially because the
478. 1:10:48incentives around commercial and national competition are hard to escape. One goal of such an organization should
479. 1:10:55be to make it possible for the world to take coordinated action including slowing frontier development when needed. Okay, again
480. 1:11:05that's Daario and Sam Alman both saying that on these um uh you know sort of battling point of views. So societal resilience, safety
481. 1:11:17and alignment can keep pace. So I guess if if Elon Musk, let's say Elon Musk, Sam Alman, Dario Amade,
482. 1:11:29Zuckerberg, even though he's not really in the game, um, uh, Sundar Pachai from from Google, you could throw Microsoft in there.
483. 1:11:42[clears throat] And then if they could get, I don't know, three or four of the Chinese, the major players in China
484. 1:11:51to say to the world, "We're going to voluntarily slow this [ __ ] down." It would at least make news. I don't
485. 1:12:01think anyone would slow down, though. [laughter] Okay. Currently in Open AI, we have three main goals.
486. 1:12:08One, build an automated AI researcher. This is what the anthropic piece was just about. An AI system that can
487. 1:12:16accelerate and increasingly automate the research proc process itself while remaining steerable, accountable, and
488. 1:12:24connected to people. So, it's saying Sam is saying we want it connected to people. We want human judgment to be a
489. 1:12:31big part of this. Daario is saying not sure if we're going to need them. Our internal belief is is that by March of
490. 1:12:382028, so basically a year and nine months, we may have significant fraction of our research being done by AI systems in
491. 1:12:47tandem with our own researchers in tandem with our own researchers. To make significant progress on alignment, we
492. 1:12:55believe we will need AIs to iterate alongside us. This will help us navigate the transition to a post AGI world
493. 1:13:03so that we can collectively decide the path forward toward the future. Okay. Accelerate the economy. Is that their Yeah. Okay. This is their
494. 1:13:15three things, three main goals. Build an automated researcher. Okay. Two, accelerate the economy by accelerating
495. 1:13:22scientific research, productivity, and economic growth while working to ensure the gains are widely shared. Everyone
496. 1:13:30should have an opportunity for a meaningful share in the prosperity AI creates. Okay, that's a fine statement, Sam, but
497. 1:13:39where's the [ __ ] meat in that? Does that look like you're you're putting 20% of open AI into a fund, a
498. 1:13:50capital fund for the American people? Are you going to challenge all the other frontier companies to to to match that
499. 1:13:59or are you just going someone should do this? Okay. Okay. Number three, give everyone on Earth a personal AGI,
500. 1:14:12empowering them to benefit from one of humanity's most trans transformative technologies in whatever way they
501. 1:14:18choose. But does that look like we've got to use your servers or does that mean you're going to make AGI small
502. 1:14:26enough to work on our machines? Who's who's developing that? Like this is thin to me to be able to deliver
503. 1:14:34this. We are entering the third phase of open AI. The first phase of open AI was about doing research toward AGI. The
504. 1:14:42second phase began when our research became relevant to the real world and we became a product company deploying our
505. 1:14:49systems, learning for how people use them and making continued progress toward AGI that is safe and aligned with our mission.
506. 1:14:56We are now entering the third phase. The economy is beginning to re reshape around AI. The central question now is
507. 1:15:05how to make advanced AI abundant, affordable, safe, useful, and easy enough for every person and organization
508. 1:15:13to benefit from it. Frontier capability is only part of the job. The bigger task is turning that capability in intools
509. 1:15:20that people can actually use to thrive. This feels This feels increasingly like he's talking about I wonder if this is the thing he's
510. 1:15:31building with Johnny Ies. Here's what this feels like to me. This feels like we're going to take our big models
511. 1:15:44and we're going to go run them on these ever increasingly powerful server farms and we're going to once once we get good
512. 1:15:52enough once we get the AI to help improve the model so that we can get full AGI and squash it onto Jensen Hang's 128 gigabyte chip
513. 1:16:04and then and then we'll give that out to people for cheap or free, but we're going to be playing up here in the
514. 1:16:10stratosphere. You guys go play with AGI. Kelly comment, Chef Kelly. But that concerns me because AI is
515. 1:16:18inherent is inherently biased and can be trained to deliver research results that could be skewed to serve an agenda. Well,
516. 1:16:31well, sure. True. I mean, you could argue that's why we have competition. That's why we have multiple companies taking different approaches.
517. 1:16:42But I could also see So I could also see there being a line of research
518. 1:16:58that says, "Hey, everyone's concerned about bias in the models. go do research that allows us to create a model where
519. 1:17:07each individual can can actually dial the bias on, you know, 50 different attributes and they can they can balance
520. 1:17:19the bias in either direction on all these attributes because the problem with like everything's got bias in it,
521. 1:17:26every human's got bias in it, every model's going to have bias in it. the bias that so if you tell me Chef Kelly
522. 1:17:33here's exactly what I want my model to do and we fix a model to do exactly what you want it to do that'll be [ __ ]
523. 1:17:42offensive and horrific to someone else like that's like it's just the inherent problem with it is that humans are all
524. 1:17:50biased and where's the sense of truth you know where who's who's who's the uh the central repository of truth I don't
525. 1:17:59know I don't No, I don't have a good answer for that one. But but just to just because the model is going to
526. 1:18:08improve itself, what Alman's saying is if you've got humans in the loop doing the judgment about here's the research
527. 1:18:16direction we want you to head. If someone like you puts it in one of those humans ears that we should be
528. 1:18:25developing models that deal with bias and someone hears that then that could potentially get in the in the mix. So
529. 1:18:32just because a computer is doing it doesn't mean it is it will be inherently bad food dye would be up in arms. Um above
530. 1:18:41all we believe broad distribution of power will lead to a better future. Human history shows that concentrated
531. 1:18:46power creates fragility while widely shared power makes societies more resilient, adaptable, and free. This is
532. 1:18:52why access matters. It's also why safety, privacy, affordability, and ecosystems and public oversight matter.
533. 1:18:58A good AI future cannot be one where a small number of institutions control most of the capability and most of the upside.
534. 1:19:07It should be a future where people, companies, communities, countries can build, benefit, and hold power. We believe this transformation
535. 1:19:15should belong to everyone. This feels to me a lot like Google's, you know, value from their early days.
536. 1:19:23first do no evil and then they presume proceeded to produce you know a company that that you know has done some um
537. 1:19:34okay if we get this right AI can be a foundation for greater productivity creativity scientific progress and
538. 1:19:40economic opportunity for many and we will achieve our mission to ensure AGI benefits all of humanity this is a this is like a f this just
539. 1:19:52honestly feels like a response to what Daario posted. Um, this section here about we have three main goals. These two are
540. 1:20:06mighty [ __ ] thin, right? Accelerate the economy by accelerating scientific progress, productivity, and economic [clears throat] growth
541. 1:20:18while working to ensure that gains are widely shared. working to ensure what's that mean you're going to tell other people that
542. 1:20:25they should share their profits or you're going to everyone should have an opportunity for meaningful share in the
543. 1:20:31prosperity a AI produ you know creates give everyone earth personal AGI empowering them to benefit like what does this mean
544. 1:20:41give everyone on earth give does that mean everyone's going to get a free device that has an AGI chip on it.
545. 1:20:52[laughter] Does this mean they will have access to free chat GPT? I don't know. All right. Well, shit's happening.
546. 1:21:05Um, we're at 10 minutes to the hour, so I'm going to stop sharing here. We're not going to get to loops tonight. We're
547. 1:21:11going to do loops on Wednesday. Now, I've decided. What are your thoughts? Who's still here? 14 people.
548. 1:21:22It was a small but appreciative crowd. [laughter] Is he really just going to read a blog post for two hours? Yeah, I am.
549. 1:21:37Thoughts? Anyone still awake? Um the any mention of the super app? Just just
550. 1:21:53that the rumors the rumors are that this week we're going to get um GPT 5.6 six and that maybe we're going to get a super
551. 1:22:04app that combines chat GPT codeex um computer use and the browser their browser but I don't know we'll [clears throat] see we'll see
552. 1:22:21Um,
553. 1:22:39so when we get to the loops, let me let me talk a little bit about the loops and then we'll we'll we'll go look at some
554. 1:22:45stuff and play with them on Wednesday. The basic idea with the loops, what they're talking about is
555. 1:23:03once you have a system that can go work for hours on your behalf, if you give it a prompt, go solve this problem,
556. 1:23:16it'll go solve it in like 10 minutes and then just sit there idle. And so loops are what are the things it
557. 1:23:24can do on an ongoing basis that are valuable for you. So do this and then go do that and then do that and then do
558. 1:23:31that and then repeat every hour, every day, every week. Loops, loops, right? And if you get all these loops running
559. 1:23:43at your behest and they those loops are doing the things that you want and need them to do, then the value of this thing
560. 1:23:50goes up dramatically. Um, but that also drives up demand for compute because all of your loops are burning tokens,
561. 1:24:04which comes back to are you running local models, right? Because tokens are going to get more expensive. Compute's
562. 1:24:11going to get more expensive. It just it just has to. Supply and demand. Who was I talking to today? I was talking to a guy I met at at AMI,
563. 1:24:21Mark, and a buddy of his uh one of one of the fathers in his like his kids group or whatever, um works at Google and and sells
564. 1:24:34cloud compute. The sales guy said he's got nothing. He's got no inventory he can sell this year and he's got no inventory he can sell
565. 1:24:46next year. This is a sales guy at Google Cloud is like we're out. like demand has [laughter] has exceeded supply to the
566. 1:24:57point that they're like they're finding land and saying if you can get this permitted and build [ __ ] here, we'll sell you compute on that.
567. 1:25:09That's today's demand. The cloud sales people at Google are like, "We ain't got nothing left." [laughter] Imagine
568. 1:25:19if we all start writing loops and our agents start using more tokens. I don't know what Sam Alman's talking
569. 1:25:32about. We should give everyone access to to AGI person. Everyone should have access to AGI. Okay. Running on what servers, dude?
570. 1:25:48We're in surreal times, people. All right. Any thoughts before I go? Thoughts or questions before I go?
571. 1:25:57Or is Claire Jacobs big food die would be up in arms? Is that going to be the last comment? Any other comments? Anyone awake? Ber.
572. 1:26:11All right, cool. All right, I'm gonna get out of here. Um, I'll see you Wednesday. We'll talk about loops. Um,
573. 1:26:19I'll demo them. I'm going to play around with the the loops I built today. Um, I put them into codeex. It did a bunch of
574. 1:26:25[ __ ] and now I need to figure out um I need to figure out what to do next. How can I test my agent?
575. 1:26:37I have so many tabs to crawl through as it should be. Um, I don't know, Mr. It depends what you built your agent to do. Like
576. 1:26:49ask it to do [ __ ] and see if it goes and does it. Um like do you have computer use enabled? Like can it go surf the web? Um
577. 1:27:03can it summarize YouTube videos? Like there's all sorts of So all of the agents, they need to have these little
578. 1:27:09skills. So, you need to understand all of what's possible and then test those individual skills. I made a cartoon in
579. 1:27:15the salon from your earlier comment, Kyle. Okay. Where'd you put it, Claire? Is it in uh is it in uh in this channel?
580. 1:27:30I can see clearly now. The rain is gone. Rick McCabiana,
581. 1:27:44thank you for doing this. AI learning lab. AI learning lab. We're in chat. [laughter]
582. 1:28:01There it is. This is this is the thing. This is the thing. AI has the very real possibility of giving us everything we've ever wanted
583. 1:28:16only to discover. Nope, that wasn't it. [laughter] There's something deeper we're going for. I built four agents. I need to test
584. 1:28:25them tonight. But Claude, syntax and commands are killing me. Yeah, I know. I know. Oh, it's a it's all this agent
585. 1:28:34stuff is an absolute nightmare right now. If you're early, it's meaning you're having to geek out. That's so good. It really is.
586. 1:28:43It's so good. Okay, cool. Um, beautiful people, let me get on out of here. Let you guys get on out of here. Um, have yourself a
587. 1:28:53fantastic night. I will see you on Wednesday. And I don't think there's anything tomorrow. Is there producer Brandon? You
588. 1:29:02are you still up? Did I put you to sleep, too? He's still going. I still see the I'm here, but nothing tomorrow. Okay, cool. So,
589. 1:29:12Wednesday, same time, same place, 7:30. Bring your friends. Kelly Camp brought a friend tonight. Thank you, Greg, if
590. 1:29:19you're still here. Um, Fabiana, great to see you. Um, all right, everybody, have a fantastic evening. Good night.