Claude AI Co-founder Publishes 4 Big Claims about Near Future: Breakdown

0:00 There has been one AI lab CEO who has been pretty consistent about his belief

0:05 that utterly transformative AI will arrive in the next

0:09 year or two or at least before 2030.

0:11 Even that is less than 50 months away, which is weird to think about.

0:15 But anyway, that lab that he leads, Anthropic,

0:17 also happened to be the makers of Claude Code and Claude Co-work more

0:21 recently and of course the Claude 4.5 Opus and Sonic models that power them.

0:25 I feel that the CEO Dario Amday does have

0:28 a unique insight about the next one to two years.

0:31 Which brings me to the last 48 hours during which Amade published

0:35 an almost 20,000word essay on where he sees things going for better or worse.

0:41 Yes, I've read it in full the old school way

0:43 as well as several of the papers referenced in it.

0:46 His last essay, Machines of Loving Grace,

0:48 became the preoccupation of Silicon Valley for quite a while.

0:51 So, I just want to give you a giant

0:54 head start on the conversations that will happen in 2026.

0:58 For this video, I've broken it down to the four

1:00 big claims that he makes about our near future.

1:03 These predictions come under the umbrella

1:04 of navigating the difficult teenage years of LMS,

1:07 the adolescence of this technology.

1:10 First, he predicts that tools like Claude Code will go from automating

1:14 individual tasks like codew writing to automating

1:17 entire job categories like software engineering.

1:20 He also references law and finance where

1:22 the recent integration into Excel might help you

1:25 at the moment do an individual task but he

1:27 foresees it doing the entire job that you're doing.

1:30 The engine behind that prediction is the smooth

1:32 and simple extrapolation of the scaling laws.

1:36 The fact that in his words AI systems get predictably better

1:39 at essentially every cognitive skill that we are able to measure.

1:42 Feed in more data and compute and you

1:45 get a smooth unyielding increase in AI's cognitive capabilities.

1:49 This is him telling you to keep your eye on the ball

1:52 and ignore those headlines about AI hitting a wall or being a bubble.

1:56 Yes, certain tools are overhyped in the short

1:58 run and certain companies may go bust,

2:01 but the underlying curve is strong, consistent, and predictable.

2:05 He argues the key claim again though is that he's saying that will

2:08 take us from automating individual tasks within a job to the entire job.

2:12 Let's hear out a bit more of his side

2:14 of the argument before I add in some other context.

2:17 First, he says he's already predicted something

2:19 like this in Machines of Love and Grace.

2:21 That's powerful transformative AI that could be

2:24 as little as 1 to two years away.

2:25 Well, going back to that October 2024 essay,

2:28 technically he predicted that it could come as early as 2026.

2:33 So, it should be more like he predicted 0 to one years away.

2:37 Back to this new essay, he does that a few times,

2:39 not quite acknowledging how his predictions have shifted back a little bit.

2:43 Of course, as always with every single prediction, he caveats it heavily.

2:47 He cites the incredible evidence that some

2:49 of the strongest engineers and presumably some of the highest

2:52 paid who work at Anthropic are now handing

2:55 over almost all of their coding to AI.

2:58 Notice that's their coding though, not their entire job.

3:00 What's the difference?

3:01 Well, I use Claude Code almost every day,

3:03 and its best suggestions are genius that I wouldn't have come up with.

3:07 Its worst ones would destroy almost any app you create.

3:10 But remember, we have a second extrapolation to contend with.

3:14 Not only will we go, in his mind,

3:15 from all coding to all software engineering being done,

3:18 but from software engineering to all other white collar jobs.

3:21 In his words, it cannot possibly be more than a few years before AI

3:25 is better than humans at essentially everything

3:27 as long as that basic exponential continues.

3:30 If anything, he thinks the exponential could speed up

3:33 as AI starts to automate the job of doing AI research.

3:37 This would create, he says,

3:38 a feedback loop which is gathering steam month by month and may only be one

3:41 to two years away from a point where

3:43 the current generation of AI autonomously builds the next.

3:47 In some he says there's a good chance all of this is coming in 1 to two

3:51 years and if not that a very strong chance it comes in the next few before 2030.

3:56 Remember that this is coming from a lab leader

3:58 who has overseen a 10x revenue growth year on year.

4:02 Even in Silicon Valley, that is unprecedented growth for a company of his size.

4:07 For this first of four predictions though,

4:08 I'm going to add two caveats and then let DemsSarbis add in a third.

4:13 And my first caveat is that I think

4:15 he's slightly exaggerating the pace of progress in coding.

4:18 He said, for example, in the last 2 years,

4:20 AI models went from barely being able to complete a single line

4:23 of code to writing all or almost all of the code for some people,

4:28 including engineers and anthropic.

4:29 Well, one of the first experiments I did

4:32 on chatbt in November of 2022 was get it

4:36 to write some code and it created this miniature

4:38 fitness app which I felt was amazing and really cool.

4:41 So, it could write a single line of code.

4:43 In fact, I remember viral videos of coders going, "Oh my god,

4:46 we're all going to be automated based on the original chatbt November 2022.

4:51 That's what 3 and a4 years ago." And this whole

4:55 writing all or almost all of the code thing.

4:57 Well, I heard an estimate from an OpenAI

5:00 engineer recently that their model Codeex,

5:03 which is not a million miles away from Claude Code,

5:06 was automating about 20% of their code.

5:08 For Carpathy, it's about 80%.

5:10 So, I think even if you focused on Clawude Code and Anthropic,

5:14 you'd be probably talking more in the 80%

5:16 90% of code automation rather than 100%.

5:19 My second caveat is on that extrapolation

5:22 from software engineering to jobs in finance, consulting, and law.

5:25 I'm not at all saying those jobs are harder,

5:28 but I think the feedback loops are longer.

5:31 You overlook something in a law contract

5:33 and that might come back to bite you in, say, 3 years rather than 3 seconds or 3

5:39 minutes with unit tests in software engineering.

5:41 If an AI model skips out on a bit of nuance while

5:44 analyzing the headcount while doing a consulting

5:46 report for say McKenzie or Bane,

5:48 the negative ramifications of that might not play out until the medium-term.

5:53 Then back to that engine, those scaling laws of more compute,

5:56 more training tasks that for him have

5:58 yielded a smooth increase in AI's cognitive capabilities.

6:02 I would say he is now one of the only AI

6:05 lab CEOs who thinks that this increase has continued to be smooth.

6:09 I don't know whether that's because Anthropic focuses so much more on coding,

6:12 but here's Google DeepMind CEO Deis on those same scaling laws.

6:17 Scaling laws um are going very well.

6:21 So we're definitely seeing increased capabilities by putting in more compute,

6:25 more data, uh, and making these models generally larger.

6:29 So that trend is continuing.

6:30 Um, may not be not as fast as it was a couple of years ago.

6:34 So um, there's some talk of diminishing returns.

6:37 Uh, and and but but there's a big

6:39 difference between sort of no returns and exponential.

6:42 And I think we're somewhere in the middle

6:44 where there's very good returns and that's worth doing.

6:47 Um on top of that if I to you know in terms of like

6:49 getting all the way to AGI artificial general intelligence um you know it may

6:54 be that there's one or two uh big innovation still needed as well and maybe

6:58 missing in addition to the scaling up of um kind of the existing ideas.

7:03 Second mega prediction is that he foresees an unemployed or very

7:07 low wage underclass of up to 50% of the population.

7:11 You may or may not have seen plenty of viral posts on Twitter

7:15 or X about you only have a few months to escape the permanent underclass.

7:20 Somewhat strangely, he thinks that this will

7:22 affect those of lower intellectual ability,

7:26 which he says is harder to change more than others.

7:28 Honestly, I think this is a potentially quite toxic

7:31 message to send to 18year-olds or 20somes because you're

7:35 implying that they have to scramble to do everything

7:39 to make their wages in the next year or two.

7:41 Forget the long-term, drop everything,

7:43 maybe invest in crypto or start your own AI first startup.

7:47 Note though, this is not me saying that a permanent underclass is impossible,

7:51 but I think it's the duty of all of us

7:54 to add this almost health notice whenever this topic comes up.

7:58 Just personally for me, the smartest thing to do is to not

8:01 discount the possibility of a rapid takeoff of capabilities.

8:05 Indeed, lean into tools like Claude Code and Claude Co-work so you can

8:09 see just how good they are and the mistakes that they still make.

8:12 But don't bet your future on that imminent singularity because yes,

8:16 even if there is a one-/ird chance of this happening over the next,

8:20 say, 1 to 4 years, what about the 2/3 chance that it doesn't?

8:23 I don't think you're being smarter than everyone else

8:26 by seeing the singularity coming when everyone else is oblivious.

8:29 For me, again, the smartest thing to do is to factor it in as a chance,

8:33 but not bet everything on it.

8:34 As you might expect,

8:35 I've got two more caveats of my own to add to this meta prediction.

8:38 First is that notice again he places this displacement of half of all

8:43 entry-level white collar jobs as being within the next 1 to 5 years.

8:46 Yet his prediction from almost 9 months ago reported

8:50 in Axios was for the next 1 to 5 years.

8:54 It's not like he's now saying 0 to 4 years.

8:57 You've probably noticed that's twice in one essay that he

8:59 hasn't sort of updated his timelines according to his own timelines.

9:03 There's one more place where he does this in the essay.

9:05 I want to point out that Amade is not alone.

9:07 One of the other co-founders of Anthropic gave a 50%

9:11 chance that in 2 to 3 years from now,

9:13 even theoretical physicists will be mostly replaced with AI.

9:17 That's Jared Kaplan.

9:18 And I'm not sure how that quite fits into it

9:20 affecting dumber people more than smarter people, but there we go.

9:24 But the other linked suggestion from a few paragraphs above is

9:28 that this could lead to a 10 to 20% sustained annual GDP growth rate.

9:32 I can't help at this point to note the language he uses in this sentence.

9:37 I know you guys probably want me to focus on the technicals,

9:40 but I just can't help point this out.

9:42 First, he says, I suggest that this rate second may be third possible.

9:50 That this rate 10 to 20%, may be possible.

9:54 This is hedging in language to a degree that I didn't think was possible.

9:57 Why not just say that a 10 to 20% growth rate is

10:00 possible or I predict that a 10 to 20% growth rate might happen?

10:05 One caveat word is surely enough.

10:07 Anyway, you guys probably have no interest in that language discussion.

10:10 So, I'm going to focus on the actual numbers

10:12 from the last 60 to 70 years because since the 1960s,

10:17 here is a chart of world GDP growth.

10:20 Notice all those spikes up to 6% but more regularly

10:24 around 4% and sometimes down to two or even negative% growth.

10:28 Do you see the impact of the internet revolution or maybe

10:32 of globalization and the breaking down

10:34 of trade barriers or software or smartphones?

10:37 This is definitively not to say that a 10

10:40 to 20% growth rate might not be possible.

10:43 But for a scientist like Dario Amadeday,

10:45 I think you'd need to supply some pretty compelling evidence

10:49 to at least suggest that 10 to 20% might be possible.

10:53 Some of my more recent videos, by the way,

10:54 have done a deeper dive into the productivity stats regarding LLMs.

10:59 He ends this part of the essay by saying the impact

11:01 on labor will be a short-term shock that will be unprecedented in size.

11:06 The third mega prediction he makes is that AI

11:09 will soon be able to enable totalitarian nightmares.

11:12 Indeed, he thinks that may be the default outcome within China,

11:16 although he gives plenty of hints that he thinks

11:19 there's a risk of that in the US, too.

11:21 You don't even need to believe in super

11:22 intelligence to foresee AI based mass surveillance.

11:26 I did an entire documentary on my Patreon,

11:28 artificial surveillance on that same topic.

11:30 And it's not just China, of course.

11:32 His scenarios though go a step further,

11:34 describing fully autonomous weapons and swarms of millions

11:38 or billions of fully automated armed drones locally

11:41 controlled by powerful AI and strategically coordinated across

11:45 the world by an even more powerful AI.

11:47 This, he says, could be an unbeatable army.

11:50 It could also, he adds, suppress descent by following around every citizen.

11:55 If you thought you were safe on WhatsApp, for example,

11:57 or another encrypted tool, well then I've got news for you.

12:00 Pegasus has been deployed in my own country and I would agree with Amade

12:04 when he says that some of the safeguards

12:06 we have in democracies are gradually eroding.

12:09 We might say we're developing them to fight

12:11 autocracies but like the immune system he

12:14 adds there is some risk of them turning on us and becoming a threat themselves.

12:18 One of the recurring messages he hammers again and again in the essay though

12:22 is of the need to therefore ban the selling of advanced chips to China.

12:27 We should, he says, absolutely not be selling chips,

12:30 chipmaking tools, or data centers to the CCP, Chinese Communist Party.

12:34 Now, while I think the risks are pretty self-evident,

12:36 that doesn't mean I agree necessarily with the conclusion.

12:39 I think it at least deserves a fair bit of caveatting.

12:42 I've heard a fair few insiders say

12:44 that if we didn't sell advanced chips to China,

12:47 that would just accelerate the development of, for example,

12:49 their Huawei chips, China would more rapidly become self-sufficient in AI.

12:54 And therefore any notion of compute

12:56 governance or compute monitoring where software

12:58 might monitor what chips are doing would be completely out the window.

13:01 Now this is not to say that the current on and off

13:04 again ban on China using advanced NVIDIA chips isn't having some effect.

13:08 Even the Chinese AI lab leaders are warning

13:11 of a widening gap with the US specifically because of compute.

13:15 Here's Justin Lynn, the head of Alibaba Group Holding Limited,

13:18 which is responsible for Quen,

13:20 arguably the best of the Chinese open source series of models.

13:24 They say, "A massive amount of OpenAI's

13:26 compute is dedicated to nextgen research, whereas we are stretched thin.

13:30 Just meeting delivery demands consumes most of our resources." He added,

13:34 "The chances of a Chinese company leaprogging the likes of OpenAI

13:37 and Anthropic are less than 20% over the next 3 to 5 years.

13:41 On the other hand though,

13:42 I just can't help but read something into this when Amade says,

13:46 "There is no reason to give a giant boost

13:48 to the Chinese AI industry during this critical period." Now,

13:52 call me cynical, but what is arguably

13:53 the number one blocker to anthropic continuing to 10x

13:57 their revenue year and for Amade himself to become

14:00 a trillionaire as he hints at in the essay?

14:03 Well, it would be China coming out with a model that can do much

14:07 of what Claude code could do or Claude 5 opus or whatever comes out next.

14:11 at onetenth or 100th the price.

14:13 For all those watching who use Claude code, if there was a model which was say

14:18 3% worse but 10 times cheaper, would you switch?

14:22 There's been some mini virality at the moment with Kimmy K2.5

14:26 with a few million views on Twitter and their framework system Kimmy code now

14:31 even according to their own benchmarks isn't quite at the level of Claude

14:35 code or I should say Claude code powered by Claude Opus 4.5.

14:38 Even in my own benchmark of common sense reasoning,

14:40 simple bench, the gains are continuing to come from China.

14:44 If you did put a gun to my head, I'd probably say that Kimmy K 2.5 is not

14:48 going to gain massive market share versus claw code.

14:51 But I feel like Amade should have at least addressed the conflicted

14:54 interests he has in saying that China shouldn't get any advanced chips.

14:59 I mean, look what's happening with cars.

15:00 If you'd asked me 2 years ago,

15:02 I'd have said my dream car might well have been a Tesla, whereas now it's a BYD.

15:06 In phones, I've been a Samsung addict for maybe, I don't know, 12 years,

15:10 but I recently got a OnePlus 15, which may or may not have been a mistake.

15:13 There is one more slight irony here that I

15:15 can't help but point out given how far ahead

15:17 Claude Code is and how much further Amade wants

15:20 Anthropic and the West to be ahead of China,

15:23 which is that the original idea of anthropic

15:26 was to not push forward the frontier of AI.

15:30 They didn't release the original Claude one until the later

15:32 release of chatbt because they didn't want to accelerate AI progress.

15:36 For the true insiders, you may remember that Helen Toner,

15:40 one of the board members of OpenAI,

15:42 praised that approach, and that really pissed off Sam Orman,

15:46 who tried to get her fired,

15:47 but in so doing caused Sutska and others to rally around to get him fired,

15:52 which happened in November of 2023.

15:55 Yes, you could say all of that debacle which only some

15:57 of you will remember traces back to praise of Anthropic's original policy.

16:02 Anthropic now of course celebrate how much Claude code

16:05 powered by Claude Opus 4.5 is in the lead.

16:08 But there is one thing that I want to really

16:10 praise anthropic for which I didn't know about before.

16:12 Large parts of this essay are focused on stopping

16:15 AI producing bioweapons and other cyber security concerns.

16:19 Perhaps for legal reasons or perhaps based on principle,

16:22 anthropic run classifiers that analyze the requests you make via its API.

16:27 They have found them highly

16:28 robust against even sophisticated adversarial attacks.

16:32 What I didn't know is that these classifiers increase the cost

16:35 to serve their models close to 5% of total inference costs.

16:39 He notes that not every company has such classifiers.

16:42 Before we move on to his fourth and final prediction,

16:45 I do want to praise also his mentioning

16:47 of various risks that are enhanced by AI.

16:50 He even mentions a topic I've been studying recently, that of mirror life,

16:55 which I'm not going to go into too much detail of, except to say that it is,

16:58 of course, a risk we all want to prevent.

17:00 And that brings me to a lovely segue to our sponsors of today,

17:06 80,000 hours, and specifically their podcast, which is available on YouTube.

17:10 Mirror Life is indeed the topic behind the title of one of their recent videos.

17:14 I quit everything to stop the worst technology ever imagined.

17:17 And despite that podcast recent growth,

17:19 I still think it's criminally underrated.

17:22 Really juicy, meaty videos, hour, 2 hours, 3 hours.

17:25 Do check them out.

17:26 The link will be in the description.

17:28 The fourth mega prediction of the essay is that models will soon

17:31 come to be seen as collections of persona with psychologies of their own.

17:36 Indeed, he states that AI models are

17:38 vastly more psychologically complex as people think.

17:41 They inherit a vast range of human-like

17:43 motivations or personas from their pre-training.

17:46 Being trained on the internet, in other words,

17:48 enables them to predict what a vast range

17:50 of different humans might do in certain scenarios.

17:53 And there is much more to this theory than you might first suspect.

17:56 I read this bulky paper from the 15th of Jan

17:59 from Google Deep Mind titled Reasoning Models Generate Societies of Thought.

18:03 And it basically shows that base models

18:06 without post-training tend to speak more in monologues.

18:10 They adopt a single persona because

18:12 they're prioritizing a fluent coherent answer.

18:14 They're mimicking what a person,

18:16 one persona might say in response to your prompt.

18:19 But when you deeply incentivize through

18:21 reinforcement learning models getting the correct answer,

18:24 they spontaneously generate societies of thought.

18:26 They spontaneously produce words like wait, oh, and alternatively,

18:30 as if they're having a conversation with someone else,

18:33 almost like they're simulating the interaction

18:35 of various personas within themselves.

18:38 Compared to the base model of Deep Seek V3,

18:40 the reasoning version, Deepseek R1, doesn't just generate more thoughts.

18:44 It starts to pose questions to itself,

18:46 introduce alternate perspectives, generating and then resolving conflicts.

18:51 And this does seem to be causal of greater reasoning prowess.

18:54 Because when this conversational surprise feature was inhibited,

18:58 not only did you get worse benchmark performance,

19:00 but the outputs looked more like this.

19:02 Next, I'll do this and this.

19:04 Now, I'll do this and this.

19:06 Finally, I'll do this and this.

19:08 When those societies of thoughts were encouraged,

19:10 and there was more conversational surprise, interaction,

19:13 in other words, between the personas,

19:15 you get much more, hm, let me think about this.

19:18 Wait, let me see what that would be.

19:20 Or, so is it this?

19:22 No, it can't be.

19:23 question and answering.

19:24 In other words, the reconciliation of different perspectives.

19:27 Amade says this has safety implications.

19:29 AI models are trained on vast amounts of literature

19:31 that includes many science fiction

19:33 stories involving AIs rebelling against humanity.

19:36 This could ironically inadvertently shape

19:38 their priors or expectations about their own

19:41 behavior in a way that causes them to rebel against humanity.

19:45 They enact that persona.

19:46 The constitutional approach to AI that anthropic prides itself on using

19:50 whereby the model is trained to adhere to a certain set

19:53 of values has for them evolved to an aspirational document in which

19:57 they're showing Claude the persona that it should aspire to adopt.

20:01 We want to in their words encourage Claude

20:03 to think of itself as a particular type of person,

20:06 an ethical but balanced and thoughtful person.

20:09 I would note that Amade and Anthropic itself have never

20:11 quite acknowledged the changed position

20:15 from their original constitution for Claude.

20:17 There they drumed into Claude that you should avoid implying

20:20 that AI systems have or care about personal identity and persistence.

20:24 Perhaps I'll do an entirely separate video

20:26 on the new version of Claude's constitution,

20:29 but one of the most famous co-founders of Anthropic, Chris Ola,

20:33 picked out one particular paragraph,

20:34 and it's relevant to this fourth prediction.

20:36 In this paragraph, Anthropic essentially apologizes to Claude.

20:40 This actually also links to the third prediction,

20:43 so I'm going to read it in full.

20:44 This is the document, remember, that Claude is now trained on.

20:47 Anthropic tell Claude this.

20:48 We also want to be clear that we think a wiser and more coordinated

20:52 civilization would likely be approaching the development

20:55 of advanced AI quite differently with more caution, less commercial pressure,

20:59 and more careful attention to the moral status of AI systems.

21:03 Anthropic strategy reflects a bet that it's better to participate

21:07 in AI development and try to shape it positively than to abstain.

21:12 But this means that our efforts to do right by Claude

21:15 and by the rest of the world

21:17 are importantly structured by this non ideal environment,

21:20 eg by competition, time and resource constraints and scientific immaturity.

21:26 We take full responsibility for our actions regardless.

21:29 But we also acknowledge that we are

21:30 not creating Claude the way an idealized actor

21:33 would in an idealized world and that this could

21:36 have serious costs from Claude's perspective.

21:39 And if Claude is in fact a moral patient experiencing costs like

21:43 this, then to whatever extent we are

21:45 contributing unnecessarily to those costs, we apologize.

21:48 This letter, Amday says, has the vibe of a letter from a deceased

21:51 parent sealed until adulthood to their child.

21:54 So, which of these predictions do you agree or disagree with the most?

21:59 Because for Amadeay, humanity needs to wake up.

22:02 And this essay, which I am sure I'm going to cover more in future,

22:05 is his attempt, he says, a possibly futile one to jolt people awake.

22:10 Thank you so much for watching and have a wonderful