tonii141

I just see a lot of people who’ve put money in the LLM basket and get scared by any reasonable comment about why LLMs aren’t almighty AGIs and may never be. Or maybe they are just dumb, idk.

I’m not demanding anything, I’m pointing out that performance tends to degrade as context scales, which follows from current LLM architectures as autoregressive models.

In that sense, Yann was right.

a) That "no-tools" win depends on prompt orchestration which can still be categorized as tooling.

b) Next-token training doesn’t magically grant inner long-horizon planners..

c) Long context ≠ robust at any length. Degradation with scale remains.

Not moving goalposts, just keeping terms precise.

a) Still true: vanilla LLMs can’t do math, they pattern-match unless you bolt on tools.

b) Still true: next-token prediction isn’t planning.

c) Still true: error accumulation is mitigated, not eliminated. Long-context quality still relies on retrieval, checks, and verifiers.

Yann’s claims were about LLMs as LLMs. With tooling, you can work around limits, but the core point stands.

Not true. AI has been around far longer than modern LLMs and has performed well in non-generative areas, often with orders of magnitude fewer parameters.

What do you mean by "they add instability"?

I agree, but maybe there is no need for billions of neurons to be simulated right away. Artificial neural networks were pretty small at the time.

I have been involved in research focused on ML control for some time, and believe me, I would love to see an AI model capable of controlling arbitrary systems at different operating points or in different environments. However, it is simply not feasible yet. This AI drone is no different, especially because reinforcement learning was used to train the model, which is generally not practical for real-world systems due to disturbance variables and the continuous need for adaptability.

Let's not forget that this works solely for this particular racing setup. If you change a single gate, the AI they are using would not be able to adapt. Still fascinating, though.

If the model uses FP16 precision and has 7 billion active parameters, it would require approximately 14 GB of VRAM. I didn't read the paper.

AI is used in scene understanding for those applications, but there is no neural network that is steering the wheel.

"they just totally failed to live up to expectation"

Because the expectation was too high. If you are aiming for precision, neural networks might not be the best solution for you. That is why generative AI works so well, it doesn’t need to be extremely precise. On the other hand you don't see people use neural networks in system control for cricital processes.

This article addresses the misconception that arises when someone lacks a clear understanding of the underlying mathematics of neural networks and mistakenly believes they are a magical solution capable of solving every problem. While neural networks are powerful tools, using them effectively requires knowledge and experience to determine when they are appropriate and when alternative approaches are better suited.

Random generator of tokens can also solve any problem if you give it enough time and memory.