- Introduction
- Scaling AI Capabilities 📈
- Introduction to Dario Amodei 🌟
I've been listening for , when does it pick up?
- Scaling Laws Explained 🔍
- Scaling laws
@ "Love never fails."
i would like to finish the whole video in detail, minute by minute, but oh mine it's long. any tips?
- Understanding AI Structure 🧠
- Limits of LLM scaling
Does anyone know if Anthropic already generates symbolic synthetic data to train long context reasoning specifically with generated logic tasks?I think that would be something that comes to mind.
- Competition with OpenAI, Google, xAI, Meta
- Claude
@, most scientific chart I have seen in my life. Real (ai hype) science under our eyes 😂
I like how doing the taxes is the example for using the smaller, worse model
- Opus 3.5
- Claude Model Variants 🎨
- Development and Testing of AI Models 🔍
Something falls on his shirt at - (1)
Something falls on his shirt at - (2)
- Sonnet 3.5
- Improvements in Sonnet 3.5 Performance 🌟
- Claude 4.0
Lex: Why'd you pick the name Sonnet "3.5"?
@ : "Sonnet, can you give me a robust and scalable naming convention for the Anthropic's AI models that will be clear to users what the model capabilities and trade offs are and that will stand the test of time, and that would still by concise?" Would that give an good answer?? Probably not yet
"It's not like software where you can say, oh, this is like, you know, 3.7, this is 3.8." - releases 3.7
- User Experience and Feedback 👥
- Criticism of Claude
I have had the opposite experience, at least with some small models.I was able to make a Google's Gemma 2 model significantly more intelligent using a specific system prompt. At least it was able to solve a handful of logic puzzles much better than before and generally felt more intelligent.It would certainly be interesting to explore this in more detail. My impression is that a certain linguistic framing can also make a major difference to performance.
I love Dario and everything but I do genuinely believe the model companies will sometimes "over quantities" the model. That's when they lower the "numerical resolution" to reduce compute (eg: using ints instead of floats for model weights,etc) The biggest culprit of this by far is openAI. My theory is based around the fact that during peak times models aren't only slower they are dumber. indicating that some dynamic capacity vs capability tradeoff is happening behind the scenes.
Awesome responde for conspiracy theories in AI❤
bookmark
Really great deep dive informative interview, thank you! and going to read mr. Amodei's 'Machines of Loving Grace'
- Challenges with AI Control ⚙
@ The solution is to have the model personalized to the individual. If the model knows enough about the user, it will know to "trust" the user with the information about small pox (because it witll "know" the user is a student) as opposed to giving the information to a terrorist (because it will have suspicion on why does this user want this highly sensitive information). In other words, the models have to stop being so generalized and start becoming more personalized. They need to start learning more about ots users. (Privacy concerns aside)
"Everyone agrees the model shouldn't talk about'
- AI Safety Levels
- Responsible Scaling Policy and AI Safety 🕵
Regarding the responsible scaling policy:Wouldn't it be possible that the model learns about this policy from the training data and then basically pretends to be dumb when you're testing it for these things?Especially if you actually build a super intelligent model
- AI Safety Levels (ASL) Overview ⚠
- Security Measures for ASL Levels 🔒
- Regulation and AI Safety 📜
- ASL-3 and ASL-4
- Computer use
- Government regulation of AI
Anthropic dunking on OAI never gets old 😂😂😂
"that it would damage the open source eco system... I think those were mostly nonsense." This was all I needed to hear to know he's a self-serving liar. That bill was absolutely going to kill open-source AI models, and ensured AI only remained in the hands of big companies like Anthropic. I'm sure he would have loved that.
- History and Insights from OpenAI 📚
- Call for Collaborative Regulation 🤝
- Vision for Organizational Safety 🚀
- I feel his pain about how unproductive (and frustrating) it is to argue with someone else’s vision. Even if you win the argument, sometimes you spend more time and energy convincing others than what would have been required to just do what you were proposing. 🤦🏼♂️
- Race to the Top vs. Race to the Bottom 🎯
- Talent Density Over Mass 💡
- Hiring a great team
inspiring. thanks alot.
- Qualities of a Great AI Researcher 🎓
- Machines of Loving Grace
- Positive AI Futures and the Essays 🌟
- Definition of AGI 🤖
- Acceleration of AI Development 🚀
- Impact of AI on Human Systems 🌐
Does anyone know what is Tyler Cowen’s essay in response to Machines of Loving Grace?
- Timeline for Achieving AGI ⏳
- AGI timeline
All natural laws are just empirical regularities. What else would they be?
- Programming
The farther the skill is from the people who are building the AI, the longer it’s going to take to get disrupted by AI
- Changing Nature of Programming 💻
I just experienced that, as a .NET expert / architect. I tried one night to use a well-known LLM to produce a piece of code for a very critical and general framework (ORM basically). The AI instantly understood my need and produced something, let's say, 80% good in just 30 seconds of voice conversation. So I started to refine the remaining 20% ... it took me 2 hours. Where it would have taken me maybe 45 min if I coded it all by myself.Since then I have a wild intuition that yes, junior coders might be replaced, yes, experts will still be needed to write critical code, but on top of that we will need even *more* experts to supervise AI generated code (which will be more quantitative than human crafted).I can feel a huge human expertise supply crisis soon.
- Integration of AI in Development Tools 💻
Which comapany is he talking about at ? It sounds like 'Expo' - the React Native deployment company? But he says 'in the security space' - maybe he means Expel?
- Emerging Opportunities for AI Companies 🌱
- Meaning of life
I don't feel satisfied with Dario's ish comments on meaning. Both the stuff about "does that make it meaningless?" His examples don't resonate with me. Likewise there are pretty good cases to be made that while yeah, poverty matters, there are huge examples of people who are poor but joyous, or rich but barren, and meaning is a huge part of that. Although YES for sure sorting out economics & inequality are critical, probably not trying to directly go after stuff about meaning.
- The Search for Meaning in an AI-Driven World 🧠
- Ethics and Power Distribution Concerns ⚖----------------------------------------------------------------------------------------------------------------------------------------------Amanda Askell (AI researcher on Claude’s character and personality)
- Amanda Askell - Philosophy
- Conversations with AI: Goals and Challenges 🔍
- Programming advice for non-technical people
- Talking to Claude
This is profound, and such an interesting challenge: “these are models that are going to be talking to people from all over the world with lots of different political views, lots of different ages, and so you have to ask yourself, what is it to be a good person in those circumstances? Is there a kind of person who can travel the world, talk to many different people, and almost everyone will come away being like, “Wow, that’s a really good person. That person seems really genuine.””
”If you really embody intellectual humility, the desire to speak decreases quickly” 😊
Your podcast is a... HOBBY, Lex ??? 🙀😹
- The Role of Prompts in Creativity 🎨
- Prompt engineering
- Iterative Prompting 🔄
- Engaging with Claude Effectively 🤖
- Post-training
- Introduction to Constitutional AI 📜
). You're contradicting yourself.
- Constitutional AI
- System prompts
- Is Claude getting dumber?
- Emotional Weight of System Prompts 💼
On the „Claude getting Dumber it’s not but we’re getting smarter and know how it acts. A good salesman might impress you in the beginning, but after a while you just know his tricks.
- Character and Ethics in AI 🤖
- Balancing Politeness and Confidence ⚖
- Character training
- Nature of truth
Maybe, Amanda, if you don't want people to treat these models like simple programs as you describe at you should also stop laughing and comparing them to your bike and your car, and saying that they're just objects (
- Optimal rate of failure
- Optimal Rate of Failure in Experimentation 📉
When I was learning roller skating that was my main motto "If you're not falling, you're not trying hard enough to learn."
- AI consciousness
This is not to be mean or rude, however, all of the dancing starting at essentially means: yes.
@ Man, why didn't we ask Dario this question? Oh well.
@ everything has a varying degree of consciousness. Our problem with the definition is a desired deliniation which is non exsitent as it is on a fluid spectrum. The best way is to acknowledge the sacrifice of others which most in power ignore. One should be thankful for any energy consumed.
- Ethical Considerations of AI Emotion 🧠
this would be genius if claude could introspect the server status and load and just integrate it conversationally, like "Sorry i have a lot going on right now can we talk later?"
- Human-AI Relationships ❤
- AI and Human Relationships 💬
- AGI
- Developing Conversations with AGI 🤖
- Identifying AGI 🔍
- What Makes Humans Special? 🌌----------------------------------------------------------------------------------------------------------------------------------------------Chris Olah (AI researcher on mechanistic interpretability)
- Chris Olah - Mechanistic Interpretability
- Mechanistic Interpretability 🔬
- Features, Circuits, Universality
You're telling me he LITERALLY lived in its head rent free? Ain't no way.
- Discussion on Confidence and Success 🌟
But I think there's also a lot of value in just being like, you know, I'm going to essentially assume,I'm gonna condition on this problem being possible or this being broadly the right approach,and I'm just gonna go and assume that for a while and go and work within that and push really hard on it.And, you know, society has lots of people doing that for different things. That's actually really usefulin terms of going and getting to, you know, either really ruling things out, right?We can be like, well, you know, that didn't work and we know that somebody tried hard.Or going in and getting to something that it does teach us something about the world.
- Superposition
- Superposition Hypothesis Explained 🔍
- Mechanistic Interpretability Challenges 🔎
- Extracting Mono-Semantic Features 🎯
- Monosemanticity
- Scaling Monosemanticity
- Future Directions in Mechanistic Interpretability 🚀
- Macroscopic behavior of neural networks
- Neural Networks vs Neuroscience 🤖
- Beauty of neural networks
- Aesthetic of Neural Networks 🌌
- Curiosity About Creation ❓
