Improving AI with Anthropic's Dario Amodei

Cofounder and CEO of Anthropic Dario Amodei unpacks how far can scaling laws take us and how can AI be used to improve AI, in conversation with a16z's Anjney Midha.

[0:00] Dario's background
[2:26] GPT-2's emergent capabilities
[5:56] Scaling laws
[8:36] Why are there so many physicists in AI?
[11:38] Constitutional AI
[18:10] Anthropic's Roadmap

[JustinHalford]

I admire Dario’s willingness to speak directly to the implications of scaling models, both the alluring and the nightmarish. Anthropic’s focus on safety will age well as calls for regulatory oversight and safety oriented benchmarking emerge.

[jaysonp]

Finally an interviewer who asks good questions

[geaca]

Great interview and honest informative answers, thank you
6:36 with further scaling up and improved compute power, huge improvements that we're prone to underestimate
13:15 if you train it, the hope is
15:43 AI working on AI safety

In another video Dario Amodei mentions Anthropic's release of a Responsible Scaling Plan - 21.44 , that's great 👍
(Anthropic's Dario Amodei on AI Finding its Best Self | TechCrunch Disrupt 2023)

[davidbutler]

I was surprised he didn't mention that with a larger context windows you can increase the number and change the kind of instructions given to the model to fine-tune it's output. Not only can you review documents but you can provide many writing sample. You can inform future interactions by telling the model how to act if certain conditions are met during the coming chat. You can even provide examples illustrating different ways of interacting with form data. It's fine-tuning the model on a small scale.

[Jim-ex]

this man is single handedly going to get me my ftx bag back

[beopen]

He just received a funding round of up to 4 BILLION USD by Amazon. Jet only gets 800 views on a YouTube video... Interesting social world. 🤔

[AfricanManikin]

9:30 really important CTA for generalists.

[kit]

Why physicists? Maybe their maths background.

[JQRNY-YDJKD]

2024年交付成果 = AI自动化AI研究 = Anthropic-10亿美元AI模型
2025年交付成果 = Anthropic彻底胜利 - 比赛结束 = Anthropic-50亿美元AI模型

[hpbecraft]

if all you have to offer as progression is more money, try again.