Many people had the idea.
Planning is a centerpiece of control engineering, robotics, and "classical" AI.
The question is how to plan with a *learned* world model.
Many people have worked on this since the late 1980s (e.g. Widrow's "truck backer-upper").
There have been many
@ylecun
-
Planning with Learned World Models in Robotics
By
–
-
LLM Reliability: Replacing Auto-Regressive Prediction with Planning
By
–
Please ignore the deluge of complete nonsense about Q*.
One of the main challenges to improve LLM reliability is to replace Auto-Regressive token prediction with planning. Pretty much every top lab (FAIR, DeepMind, OpenAI etc) is working on that and some have already published -
Foundation Models Regulation Debate: Open Source vs Big Tech
By
–
Max, how many times do we have to repeat that the main disagreement is about *foundation models* particularly *open source* ones. The ones who do want broad regulations are Google, OpenAI, Anthropic, and a few EA-funded doomer institutes like your FLI. The ones who don't want
-
Product-Level AI Regulation: EU AI Act Critique
By
–
Oh, come on, Max.
The kind of regulation @cedric_o was saying we need, and the kind of regulation he objects to in the EU AI Act, are completely different.
Of course, there is a need for regulations at the product level.
Such regulations already exist in many areas -
LLM Information Content Cannot Fit in 800MB Storage
By
–
You simply cannot squeeze the information content of a trained LLM (even a tiny one) in 800MB.
-
Minimal Code Architecture Enables Massive Learning
By
–
The code that specifies the architecture is tiny.
That's precisely my point.
There is a tiny amount of prior structure and a lot of learning. -
Language Hardwiring Constraints Within 8MB Memory Limits
By
–
Whatever "hardwiring" is required for language has to fit in 8MB.
-
Brain structure and learning in artificial intelligence
By
–
A tiny amount of brain structure and lots of learning.
-
Sensory Data and Multimodal Learning in AI Development
By
–
Yes, that's one of my points.
Text is insufficient.
We need sensory inputs to learn how the world works.
We can estimate the total amount of visual data seen by a 2 year-old: 2 years = 2x365x12x3600 or roughly 32 million seconds.
We have 2 million optical nerve fibers, carrying -
Genome Storage vs LLM Size: Evolution’s Compression Challenge
By
–
Whatever it is that we learned through evolution has to be squeezed in 800MB (the size of the genome, uncompressed).
most of it is just low-level biochemical machinery. Even a tiny LLM requires 14GB.