Thanh Do

I’m a bit confused here - are we the tactile control to a llm and expect superhuman specification in a certain format (squeeze hard or soft? where to grip in a nonuniform object, so on to a llm)? Should we expect to have a long tail here and everything be fixed by general knowledge -> downstream knowledge from the llm itself?

Interesting post! Is it possible for you to provide more context wrt how specialized robotics model are performing on these task and their rate of improvement, and is this sota for them as well?

Do you think that these frontier models would actually outperforms specialized model, and if they were to be competitive with teleops, do you think it's skewed in domains or is actually pretty good per domain?

I meant your personal specification is too general. i mean even with the whole new memory system in Claude or oai, they dont actually memorize non-instruction well at all, so it will be a bit iffy. I personally think that solving relevance realization will be an engineering problem with memory -- namely @gwern 's Guardian Angel, or what Engram is doing -- there is just so much you can specify out in words in terms of what you want, and a lot of it is latent (which is not captured in the "memory") part -- you dont have enough patience for specification.

Ano... (read more)

Do you think that 4o (I think this is the strongest model before all of the RL-distilling-onto-context) is better at relevance realization than a human? Or do you think that without a strict causal training with an environment dynamics that is easier to distill (think nanda's toy circuits, or the chess RL paper) is required for relevance realization?

Another possibility is, as another commenter pointed out - our specification is too general for it to actually read the intention behind it -- this is a "general" model after all?

Honestly I'm not sure if the conceptual reasoning is what I was referring to either - as far as I understand its' largely consistency in claims-making, but then again, we know that what is distilled to the model (literal reasoning traces in OPSD for example) doesnt neccessitate activation in the "probability consistency" circuits - if such circuits even exists and is coherent universally -- namely that they activates and acts causally as a bayesian estimator.

I'm not sure why this is true - isnt "relevance" here a causal term and ai have us beaten in narrow domain - namely chess and others - why cant frontier labs optimize this domain as welll?

Even in a comparatively open domain, they dont need to have a closed form solution for "humor" - they just somewhat needs to beats human baseline domain shift, no? and i doubt this is not possible if they were to record thousands of organic conversations everyday plus web crawling?

I'm not sure if this shows a consistent change in behavior when chatting. they inject names that are associated with being a safety researcher - probably linearly decodable, while the writing style and concept use is much harder than this concept. This is like, in my understanding, mentioning "piano" in the prompt making the model more inclined to say something related to "piano". Also the transluce piece was a single or double turn, and i doubt that the model escapes the mono-assistant basin once you chat with it.

I remain unconvinced that the writing sty... (read more)

But I’ve never seen anybody use it in industry

Disagree - i used to work for some startup in computer vision, and it's pretty common for us to visualize and manually change the hyperparams - yes, we manually changed it

K complexity is separable into two components, one corresponding to randomness, and the other corresponding to deterministic structure

Corresponds with how bayesian Reinforcement Learning models, the randomness component corresponds with minimizing aleatoric uncertainty while the epistemic component corresponds with minimizing epistemic uncertainty.