Why does that matter?
The only questions are:
1. Is it safe?
2. Is it cost effective?
If remote operators have to intervene once in a while, so be it, as long as it's safer and cheaper than a human driver in the car. Historically, most AI-assisted services have used a
@ylecun
-
Remote-Operated Autonomous Vehicles: Safety and Cost Effectiveness
By
–
-
Non-Sample-Contrastive Methods for JEPA Research Focus
By
–
Essentially *all* of my own research today is on non-sample-contrastive methods for JEPA.
That tells you what I think about them. -
US Government AI Regulation: Big Tech vs Open Source Advocates
By
–
Both are false.
The US government is primarily concerned with 2 things: public safety and economic development. The corporations advocating for regulations are both large (Google) and small (OpenAI, Anthropic).
So are those in favor of open source: large (Meta, IBM, Shopify), -
Billionaire Funding for AI Extinction Risk Research Causes Unintended Damage
By
–
There is some "academic money", but fairly small amounts.
Most of the money comes from a few billionaires worried about extinction risks.
Much of it goes to private non-profits, not universities.
They have good intentions but are unwittingly causing a lot of damage. -
AI Fear Drives Policymakers to Enact Symbolic Regulations
By
–
The only half-cogent argument is this: because people have become scared of AI (by fallacious arguments) policymakers have to look like they are doing something about it.
-
Open-source LLMs surge: Can AI R&D regulation work?
By
–
High-performance open-source LLMs popping up around the world.
Does anyone believe regulating AI R&D can be effective?
(Assuming it's useful. Which it is not). -
FSD Falls Short of Human Learning Abilities Despite Data Investment
By
–
FSD has required enormous amounts of data, years of engineering to take care of corner cases, and the reliability is nowhere near that of a human (though it's clearly useful. I use it).
Clearly, we are very far from the learning abilities of a 17 year old. -
How Do Humans Learn World Models?
By
–
Exactly.
That's my whole point.
How do humans learn world models? -
Joint Embedding Architectures and Non-Contrastive Learning Methods
By
–
Well, technically, my first paper on Joint Embedding Architectures (AKA Siamese nets) is from NIPS 1994.
but that paper and many subsequent works use sample-contrastive learning and no predictor.
The JEPA idea trained non sample-contrastive losses (Barlow Twins, VICReg, MCR2) or -
LLMs Retrieval vs Reasoning: Understanding Fundamental Limitations
By
–
Don't confuse the approximate retrieval abilities of LLMs for actual reasoning abilities.