World models from video: since 2014.
Learned World Models for planning: since 2018.
Joint Embedding Architectures: since 1993, but a lot more since 2019.
Joint Embedding Predictive Architecture (JEPA) for images: since 2021.
JEPA trained from video: since earlier this year.
@ylecun
-
Joint Embedding Predictive Architecture Evolution Since 1993
By
–
-
Auto-Regressive LLMs: Dumb Yet Knowledgeable and Useful
By
–
My views on how useful Auto-Regressive LLMs can be has not changed.
I've said numerous times that they are dumb and unreliable, yet knowledgeable and useful. -
Goals and Planning Mechanisms Essential for Advanced Machine Intelligence
By
–
I'm not entirely sure what @sapinker means by "explicit knowedge", but goals combined with a mechanism to plan actions that fulfill those goals do seem necessary for Advanced Machine Intelligence (AMI).
One of several things @sapinker and I agree on. -
Psychologists Study LLM Capabilities and Cultural Impact
By
–
Awesome study led by @AlisonGopnik on the capabilities of LLMs from the psychologists' standpoint. Quote: "Large language models such as ChatGPT are valuable cultural technologies. They can imitate millions of human writers, summarize long texts, translate between languages,
-
Open Source AI Platforms: Efficiency and Customization Benefits
By
–
Open source AI platforms aren't just cheaper.
They are more efficient and more customizable. -
Proactive Leaders Drive AI Progress Forward
By
–
People like me and my colleagues don't "wait" for progress to be made.
We make the progress happen. -
Game Playing Systems and MCTS Planning Procedures in AI
By
–
Most game playing systems, including AlphaGo, are objective driven and do perform planning. MCTS is a heuristic planning procedure for discrete action spaces.
-
Multimodal AI: Architectural Challenges Beyond Text Generation
By
–
Everyone is working on multimodal systems.
The question is how to do it.
And the problem is that the kind of generative architecture that works for text does not work for images and video. -
Distinguishing Approximate Retrieval from Reasoning in LLMs
By
–
Do not confuse approximate retrieval with reasoning.
I used this scenario as a simple didactic example of physical understanding (or lack thereof).
Naturally, a pure text LLM will answer such questions correctly if the scenario, or a significantly similar one, is described in its -
Engineering Over Optimism: Building Tomorrow’s AI Solutions
By
–
It's not optimism. It's engineering.