A hallmark of human intelligence is the capacity for rapid adaptation, solving new problems quickly under novel and unfamiliar conditions. How can we build machines to do so? In our new preprint, we propose that any general intelligence system must have an adaptive world model, i.e. they must be able to rapidly construct or refine their internal representation through interaction and exploration — a process we call “world model induction”. We propose a roadmap for evaluating adaptive world models in machines based on a special class of games we call “novel games”.

Intelligence isn't a collection of skills. It's the efficiency with which you acquire and deploy new skills. It's an efficiency ratio. And that's why benchmark scores can be very misleading about the actual intelligence of AI systems.

Of the top 20 US tech companies by market cap, only 1 is headquartered in SF (Salesforce). That's less than Austin (which has 2). The bulk of US tech companies are headquartered in Santa Clara county (11 out of the top 20). The nearest major city (by a lot) is San Jose. SF is over 1 hour away (up to 2 if there's traffic).

My favorite use case for video generation so far: use it to turn your kid's stories into animated clips Kids find it completely natural that computers can make videos, and they love to see what they describe come to life -- it’s like seeing your imagination validated, made real

The goal of education should not be to achieve a certain population average on standardized tests, it should be to unlock the highest total individual potential out of students

Btw we're hosting a private event for the launch of the ARC-AGI 3 developer preview, next week in SF. July 17th, 6pm. Open to sponsors, donors, and researchers. Several big announcements coming. Also @dwarkesh_sp will be hosting our fireside chat. Link in next tweet.

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

In my personal experience with Duolingo, I found that it isn't an education app, it's a mobile game -- it's completely ineffective at teaching languages but it's quite entertaining and it's great at making you *feel* like you're learning

Some other interesting findings! We find that developers are slowed down less on tasks they are less familiar with. This is intuitive—if you really know what you’re doing, AI can be less marginally helpful.

I was pretty skeptical that this study was worth running, because I thought that *obviously* we would see significant speedup.