Children learn by playing because not much is expected of the outcome in play. Improvement happens when you can play. When AI has a play environment to learn with reinforcement. When entrepreneurs are allowed to try and fail and do better. Doctors learn by practicing under supervision, or on corpses, until they can do it for real. No straight line goes up without a jiggle in the beginning.
I suppose another interesting thing about this observation is that this is true about the universe too! Einstein thought God doesn't play dice with the universe and Niels Bohr proved him wrong.
So either it's an interesting statement because it's infinitely generalizable or not interesting for the same reason.
I like the networking perspective, but the ML perspective is such a loose analogy that it's hard to even judge. I mean, we've known forever softening constraints allows you to reach solutions otherwise unreachable, for one? There's a gulf of difference between succeeding at something deterministic by allowing failure vs. good pattern matching by optimizing over a rough landscape of examples.
The LLM reasoning models behave strikingly similar to superscalar out-of-order execution processors with decomposition, verification, and error correction steps.
Moreover, the LLM reasoning models are reliably consistent solving the same task with the same prompt using different variables. This can be demonstrated.
Not everything has to be deterministic to be useful. Nonetheless, understand how LLM models can be applied and be useful will help a lot of people to be less frustrated and spend less tokens.
Children learn by playing because not much is expected of the outcome in play. Improvement happens when you can play. When AI has a play environment to learn with reinforcement. When entrepreneurs are allowed to try and fail and do better. Doctors learn by practicing under supervision, or on corpses, until they can do it for real. No straight line goes up without a jiggle in the beginning.
I suppose another interesting thing about this observation is that this is true about the universe too! Einstein thought God doesn't play dice with the universe and Niels Bohr proved him wrong.
So either it's an interesting statement because it's infinitely generalizable or not interesting for the same reason.
I like the networking perspective, but the ML perspective is such a loose analogy that it's hard to even judge. I mean, we've known forever softening constraints allows you to reach solutions otherwise unreachable, for one? There's a gulf of difference between succeeding at something deterministic by allowing failure vs. good pattern matching by optimizing over a rough landscape of examples.
Yeah, I didn't find his initial take very convincing, but he lost me at the followup:
> For most cases I don't think having explainability is worth the trade offs in capability. That'll be a good topic for a future post.
I'm not seeing how describing measures over possibility space as allowing for mistakes.
Seems like content reverse engineered from title.
The LLM reasoning models behave strikingly similar to superscalar out-of-order execution processors with decomposition, verification, and error correction steps.
Moreover, the LLM reasoning models are reliably consistent solving the same task with the same prompt using different variables. This can be demonstrated.
Not everything has to be deterministic to be useful. Nonetheless, understand how LLM models can be applied and be useful will help a lot of people to be less frustrated and spend less tokens.
Interesting. I could apply this to some people I've worked with. They work so well because they don't have to.