One reason I focus on auditing of American AI companies first, not jumping straight to China

(another reason being that verification is harder in a Chinese context for many reasons - legal differences, cultural barriers, etc. - more prep work is needed)

@tszzl @robertwiblin @jrysana

@jrysana american models seem to be significantly better than and actually pulling away from the Chinese open source ones. I would say the American frontier models are just racing with each other mostly

@tszzl @jrysana yup (only https://deepswe.datacurve.ai/blog currently shows this clearly, but practitioners notice this intuitively)

@tszzl @jrysana sorry Kimi K2 and Dipsy R1 are civilization-complete benchmaxxing may pump bags but doesn't address reality any better than "solving" DotA 2 did

@tszzl @jrysana anyone whos using these knows that benchmaxxed to hell doesnt actually translate. Qwen 35B are great FT models but like... the big boys cannot hang

@tszzl agreed but if we pause?

@tszzl @jrysana Are you betting there will be no Mythos level oss in the next 6 months?

@jrysana @tszzl i think it would increase gap, since the gap turns unscrapeable

@tszzl @jrysana homographic-carry

@tszzl @jrysana There won’t be a pause. The labs may say they are pausing to each other, but there is no way they will let the other validate that. And there is no way they trust that the other will not be advancing while saying they are pausing. Onward with RSI.

@tszzl @jrysana "better" is multi variable and western models are certainly not better in many of the most important axis (price/action, speed, censorship)

@tszzl @Miles_Brundage @jrysana Question 🙋‍♂️, was distillation a bigger issue that’s been solved or slowed; ie the ability to steal and copy leading models and convert to open source or your own model?

It seemed like an issue to releasing a better model was making it easier to steal. Wondering if that’s done.

@tszzl @jrysana American AI is only superior in your experience, because you get to use it without all the extra guardrails. For the regular user, it can actually lose to open-source and chinese models.

@tszzl @jrysana Are your open source models better than theirs though? 🤔

@tszzl @Miles_Brundage @jrysana The singularity is coming.

@Miles_Brundage so ur saying the chinese models need translation layers for safety, not just architecture

@Miles_Brundage honestly sounds like the china gap is getting exaggerated

more prep = more excuses or real blockers though

@intellectronica @tszzl @jrysana Yeah but the pace is also non linear, I bet 6 months ago Openai was were kimi is

@JohnGal43951639 @tszzl @robertwiblin @jrysana Do you have a source for that chart?