Post

I am not ready for this conversation

LLM model merging through a shared latent space VAE mapping

Sex is only for reproduction

The DC GPUs are potentially going to be a lot more power efficient for inference so you know it really depends

No that's how they turn blue in the mountains of madness

Parameters / 2 + 7, not ok for bluesky /Now with better math

now we're talking

It depends on the quantization.

GLM 5.2, ~750B

hi

It's not just ragebait — it's enrichment

no!! and before some creep says "but when they're distilled from the frontier they're so mature for their size" that 👏 makes 👏 it worse 👏

unethical but hot

Yeah I was staring at the initial equation thinking "no way that's right". As much of a meme as this answer is, it seems like the right shape for an actual answer? At least as a rule of thumb. I think in practice we'll need a test of minimum capabilities.

it seems like the only way to know is to spin up thousands of these and watch them play out in parallel

No because the 70B model is older and there's a problematic age gap

"That's the bitter lesson, and here's the part that stings...." bsky.app/profile/emol...