OpenAI's roon argues automated post-training benchmarks are key to unlocking personalized AI models

until UNSONG strikes of course

@tszzl @johncelery hold up... full fucking stop 🛑 , no.

@tszzl What artform isn’t accessible

@tszzl factsssss

@tszzl I don't know if I'm insane but it makes me a lot more excited than nervous

@tszzl text2model wen

imagine you can tell an AI "I need a hyperspecific AI for exactly x, y and z" and it then trains some open source 8b model to be perfect at these exact tasks

@johncelery you know how making a high graphics video game or movie or w/e requires tens of millions and writing a poem requires some paper?

@tszzl

@tszzl I have thought about this flywheel alot: then, when the models themselves have the know-how it will really come down to the data access and quality

@tszzl Posttrain API for current gen models?

@tszzl 9 out of 10 desirable explosions are Cambrian. That’s just science.

@tszzl can't wait for temperament engineering to become a phrase

@tszzl accessible artform. so etsy, but the listings have opinions about you

@tszzl I think this is much more fun question than capabilities post training which I would expect to be just cargo culting on open data sets and ends with a generic assistant.

@tszzl GLM 5.2 quietly eating everyone’s lunch on post-train benchmarks while being 5-11x cheaper is the real signal.

@tszzl Bodies have many brains - from your cerebellum to your spinal cord to ganglia in your joints, intelligence is distributed and specialized for optimal latency and performance. A billion minds in a billion niches

@tszzl A cambrian explosion of authoring minds powered by models? That's going to make creativity more accessible than ever, wild stuff.

@tszzl Yeah!

@tszzl everheard of (1/2)^n

@tszzl without a meaningful mechanism for exploring the combinatorial space of trajectories that may lead to new types of minds, wouldn't this just turn the space we explore into a Chihuly sculpture?