You aren't sampling from a learned manifold of the game space, you are sampling from the *actual* game space. You aren't in any way limited by what you've previously learned. That's the difference between search and interpolation.
@fchollet
-
LLM-generated training data differs from actual game data like AlphaGo
By
–
No, I'm talking about using a LLM to generate LLM training data, ungrounded (with no selection by a non-DL system). AlphaGo game data is actual game data, it isn't generated by a LLM.
-
Consciousness Requires Permanent and Instant Context Models
By
–
As for what drives this emergence — my view is that consciousness requires context, both permanent context (a "world model") and instant context (your sensorimotor data feeding into an instant "situation model"). Newborns have little permanent context (they've learned very
-
Timeline expectations for AI progress shift from previous predictions
By
–
Seems shockingly reasonable. The median view in SF seemed to be "two years away"… last year.
-
Testing scaling hypothesis with internet-scale data and distribution shift
By
–
If we had billions of dollars to spend we could test this hypothesis empirically — try to fit a big curve on *all the data* — literally everything that's on the Internet, and then test if it can adapt to things that are slightly out of distribution, like ARC tasks. Oh, wait a
-
Human Intelligence: Dynamic Adaptation Beyond Static Pattern Matching
By
–
Humans aren't parametric curves fitted to a data distribution and performing static test-time inference. The entire point of human intelligence is to let us operate in the present, which is always unlike the past.
-
AI Math Olympiad Competition: Gemma 2B KerasNLP Starter Notebook
By
–
Are you doing the AI Math Olympiad competition on Kaggle? Here's a notebook to get you started with Gemma 2B with KerasNLP:
-
NVDA PE Ratio Returns to November 2022 Levels Despite Higher Stock Price
By
–
NVDA now has a PE ratio comparable to what it was in November 2022, when it was trading around $150-160 per share.
-
François Chollet discusses AGI and Keras 3 on People of AI
By
–
I went on the People of AI podcast to talk about AGI, Keras 3, and a bunch of other things. Check it out!
-
LLMs Unnecessary When Cheaper Models Perform Better
By
–
You absolutely don't need LLMs for this. Using LLMs here where a model that's 0.1% as expensive would work 5x better would be a show of gross incompetence.