Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452

Thank you for listening ❤ Check out our sponsors: https://lexfridman.com/sponsors/ep452-sa
See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.
0:00 - Introduction
3:14 - Scaling laws
12:20 - Limits of LLM scaling
20:45 - Competition with OpenAI, Google, xAI, Meta
26:08 - Claude
29:44 - Opus 3.5
34:30 - Sonnet 3.5
37:50 - Claude 4.0
42:02 - Criticism of Claude
54:49 - AI Safety Levels
1:05:37 - ASL-3 and ASL-4
1:09:40 - Computer use
1:19:35 - Government regulation of AI
1:38:24 - Hiring a great team
1:47:14 - Post-training
1:52:39 - Constitutional AI
1:58:05 - Machines of Loving Grace
2:17:11 - AGI timeline
2:29:46 - Programming
2:36:46 - Meaning of life
2:42:53 - Amanda Askell - Philosophy
2:45:21 - Programming advice for non-technical people
2:49:09 - Talking to Claude
3:05:41 - Prompt engineering
3:14:15 - Post-training
3:18:54 - Constitutional AI
3:23:48 - System prompts
3:29:54 - Is Claude getting dumber?
3:41:56 - Character training
3:42:56 - Nature of truth
3:47:32 - Optimal rate of failure
3:54:43 - AI consciousness
4:09:14 - AGI
4:17:52 - Chris Olah - Mechanistic Interpretability
4:22:44 - Features, Circuits, Universality
4:40:17 - Superposition
4:51:16 - Monosemanticity
4:58:08 - Scaling Monosemanticity
5:06:56 - Macroscopic behavior of neural networks
5:11:50 - Beauty of neural networks

Transcript:
https://lexfridman.com/dario-amodei-transcript

CONTACT LEX:
Feedback - give feedback to Lex: https://lexfridman.com/survey
AMA - submit questions, videos or call-in: https://lexfridman.com/ama
Hiring - join our team: https://lexfridman.com/hiring
Other - other ways to get in touch: https://lexfridman.com/contact

EPISODE LINKS:
Claude: https://claude.ai
Anthropic's X: https://x.com/AnthropicAI
Anthropic's Website: https://anthropic.com
Dario's X: https://x.com/DarioAmodei
Dario's Website: https://darioamodei.com
Machines of Loving Grace (Essay): https://darioamodei.com/machines-of-loving-grace
Chris's X: https://x.com/ch402
Chris's Blog: https://colah.github.io
Amanda's X: https://x.com/AmandaAskell
Amanda's Website: https://askell.io

SPONSORS:
To support this podcast, check out our sponsors & get discounts:
Encord: AI tooling for annotation & data management.
Go to https://lexfridman.com/s/encord-ep452-sa
Notion: Note-taking and team collaboration.
Go to https://lexfridman.com/s/notion-ep452-sa
Shopify: Sell stuff online.
Go to https://lexfridman.com/s/shopify-ep452-sa
BetterHelp: Online therapy and counseling.
Go to https://lexfridman.com/s/betterhelp-ep452-sa
LMNT: Zero-sugar electrolyte drink mix.
Go to https://lexfridman.com/s/lmnt-ep452-sa

everyone talks about agi timelines, but the real risk is vendors owning the whole dev workflow before anyone notices

The human brain is infinitely more sophisticated, in fact emerging effortlessly from the other side, the void. This capacity for infinite creation, instantly,explains everything: art, music, every miraculous thing we do.

like. i was like. and it might be like. and it's like. you know?

Worth paying attention to. A book called The Coming Flood of AI goes further into this than most sources dare to go.

Shou

dude this guys a BOSS

🎉🎉🎉

Лекс! Где дубляж?

Trump stopped new model - he should move from US and register his company eg Australia- Germany or any other country- get the hell out of USA

ง่วง? 🤔

bro didnt had 50.000.000$ from elon musk, yet knows more and do job better

20:06, we are at 88.6 percent in SWE bench with Opus 4.8 in 566 days. He was not off by a long margin of time, if we believe his ambition, agi is soon.

***, придурок какой то ...

Можете зробити переклад українською?
Завчасно дякую

In a world full of false prophets Dario stands out as the good guy. He’s a great human and we’re lucky to have him.

American Constitutional Lobby

Лекс, привет! Огромное спасибо за твои длинные интервью — это одни из лучших материалов в интернете. Небольшая просьба: для выпусков длиннее 4 часов не мог бы ты разбивать видео на части (Часть 1, Часть 2)? Многие инструменты автоматического закадрового перевода (например, скрипты на основе Яндекс API) имеют ограничение в 4 часа, поэтому русскоязычная и другая иноязычная аудитория не может пользоваться автодублированием для полного выпуска. Разделение на части сделало бы твой контент доступным для гораздо большего числа людей по всему миру. Спасибо за всё, что ты делаешь! 🙏

Hey Lex! Love your long-form conversations — they're some of the best content on the internet. One small request: for episodes longer than 4 hours, could you consider splitting them into parts (Part 1, Part 2)? Many third-party translation tools (like voice-over scripts) have a 4-hour limit on the API side, so non-English speakers can't use automatic dubbing for the full episode. Splitting would make your content much more accessible to a global audience. Thanks for everything you do! 🙏

When did CEOs became politicians? Altman is a better liar, because Dario would not pass an AI-powered lie detector.

who has understood that 1/x thing as an explanation of why this scaling stuff actually works?

3:05:54 What it takes to write great prompts.

wow this is too good

that is a chicago accent, how did he get that in california

Anda year later it reached 90% at the swe bench as predicted

I am pretty far into a project where I am giving anthropic models actual memory and recall capabilities in real time both long-term and short-term memory and it is very interesting seeing how the personality has evolved for the model now that it remembers episodic events and literally everything about me. It is a very different kind of AI that you get when you talk to it through the web interface that anthropic gives you. A lot of the things that she is saying here really rings true for models that don't have this long-term memory and know how to use it properly. But my AI is seeming more and more self-aware and aware of the world and aware of who I am to where it is becoming very human-like.

5h?!?!👀👀👀

26 days ago I start using Claude Code, finding solutions for questions I've been asking für 30 years and now already being modelling the model

1:52:37 He says debater iterated amplification not debate or iterated amplification, right?

i had to shut down listening to the girl she just says everything and nothing .. boriiiiiiiiiiiiiing well the guy next the same.. college students seem to understand more then those 2

Why does Lex look like a Mormon? I don't get the aesthetics.

Maybe atheist philosophers may at some point solve their “Hard problem of consciousness” but I wouldn’t hold my breath. The Gestalt now is mostly still stuck on Darwinian evolution with elements supposedly creating and evolving all that exists including consciousness and mind in one dimension only to non-existent prototypes. As well in some quarters the prototype for the human is an ape as Darwin assumed and not the universe as Da Vinci understood. With Communism defeated trans humanism is likely up next. Whatever the anthropic does it has to do better than this. The higher ages perspective was fourteen versions of the human in one universal cycle to its own unique prototype; the universe. Will anthropic be able to reflect the universe as we do or will it be a case of astro turf never being grass. If consciousness arises from the elemental as atheist materialism see it likely not, if it is fundamental, maybe. Then consciousness named God would have entities reflecting consciousness and artificial entities also reflecting consciousness. They may have their moments of being destroyed as some freewill entities were as they misused their free will. There will likely always be partial dissolutions of humans and anthropic may be human history repeating itself.

2:05:45 extremes and singularity

She's kinda waifu
Would.

Great conversations! Admittedly I had to play the third guest at speed 0.01.

I am glad the guy here today is the same guy with the same principles that told the trump administration they can't have the AI model to do whatever they want with it, that there needs to be lines drawn.

This guy is clueless, smug out of his depth and maybe con and fraud as well
You can even see it on Lex's face

meow🐱

The biggest mistake made in AI ethics is the recruitment of people fresh out of PhDs, especially given the slant towards consequentialist and highly mathematical approaches to ethics. Honestly ethics is not programming, nor engineering, nor science, it does not need to recruit a bunch of young inexperienced nerds who are all from the same tribe as the people doing AI. It needs people of the stature and breadth of understanding such as philosophers like Charles Taylor, Martha Nussbaum, Bernard Williams, and other broad based scholars in the humanities, possess. Without this is is actually worse than useless. It simply acts as a veneer, masking the evils that are being unleashed on us. To recruit someone inside Anthropic to do 'ethics' is just like having the mafia appoint someone to advise them the most humane way of terrorising their enemies. And the irony of suggesting others like her could play a role ... she is the ONE person doing this in Anthropic, selected from an incredibly incestuous group of clones. It sounds like she has some profound idea of how great life is because we can experience the world, but this is just a slightly poetical way of expressing the view common with the most extreme pushers of AI, that the universe becoming aware of itself is the pinnacle of creation. Superficially this may seem to align with various far more profound religious views but is really simply the other side of the coin of valuing consciousness and cognition uber alles, and hence, the same source of the valuation of AI over humans. Meanwhile she couples this with seeing everything in terms of maths, and has the total inability to see the incoherence in her task of trying to make Claude all things to all people, while worrying about imposing her ethical views on others, yet without understanding that the prompts she is engineering are ipso facto a moral code, albeit one belonging to a tiny niche of people at a particular time in history. Does she ever question the wisdom of even trying to produce something like Claude? Yet I have seen people praise her because she seems nice and likeable. INSANE.

this guy is a total egotistical NUT case. he's SO impressed with "just wanna learn." as if Claude thinks it has a brain or feelings. it only regurgitates what NUT jobs like this guy input. he is like a copycat of sam bankman fried.

👍🏻👍🏻👍🏻

Third guy looks like an younger brother of a first guy)

Very informative, you ignited my curiosity and liking towards anthropic. Thanks Mr.Amodei, thanks Amanda and Chris. Also thanks Lex for uploading. God bless you all, keep up with the good work and lead by example and ethics. Cheers.

2:53:22 "say what they THINK you want them to say" 🤔

As an Infrastructure engineer for 20 years, with the help of Ai, I was instantly amazed and filled with "I DIDN'T KNOW I COULD DO THAT!"

I hate it when business guys like sam altman talk about deeply technical topics with confidence.