AI-Toolbox: Who's building it?
Andrej Karpathy recently tweeted this:
The Data Wall, Agents, and Planning-Based Evals
I’ve been thinking a lot about the whole “data wall” thing with LLMs lately. It’s the idea that LLMs can’t or won’t improve because we’ve exhausted all the possible training data. I don’t buy it. The best models are appearing to plateau, but it’s not a lack of training data.
Internal Monologue Capture
I can’t stop thinking about a new concept that AI applications could benefit from. I’m calling it internal monologue capture. When Daniel Miessler and I were hanging out a few months ago, I told him a huge level-up that AI applications need is the internal monologue from experts. I’m pumped to finally write a blog about it.
Unleashing Claude 3.5 Sonnet As A Hacker
Claude 3.5 was recently released, and it’s a clear step up from any other model currently available. Not only is it more advanced, but it’s also incredibly fast and cost-effective. This combination of features makes it perfect for a wide range of applications.
Defining Real AI Risks
Yann LeCun is making the same mistake Marc Andreesen makes about AI risk. They aren’t considering how powerful a system can be which incorporates generative AI with other code, tools, and features. LLMs can’t cause massively bad outcomes, but it’s not absurd to think human-directed LLM applications with powerful tools could cause large-scale harm.
