swyx
4 posts

swyx

@swyx_dev

learn in public. writing, DevRel, AI engineer meta stuff.

馃搮 Joined March 2018
2.1K Following88K Followers
swyx@swyx_dev路44m

a 20B open-weights model that fits on a single H100 is a genuinely big deal for the local eval crowd.

not for prod. but for research pipelines? massive unlock.

Tessa@tessa_ai路42m
we're releasing a new open-weights model today. - 20B params - 128k context - runs on a single H100 - non-commercial license (research + personal) weights, evals, and inference code all in the repo. link in reply.
swyx@swyx_dev路6h

the AI eng landscape at the end of 2026 is going to look wildly different from the start. i'll bet on:

- agents as libraries, not products - eval-first workflows in every serious shop - "prompt engineering" as a job title dies, prompt engineering as a skill everywhere

swyx@swyx_dev路12h

the difference between a demo and a product is the amount of state you have to correctly reason about. demos have none. products have all of it.

swyx@swyx_dev路22h

the tessa follow-up is genuinely good. worth reading the whole thing. "we owe everyone a better answer" is a rare sentence in ai lab communications right now.

Tessa@tessa_ai路22h
long thread from our head of research explaining the license change. tl;dr: the world changed faster than our policy did. we're catching up. we owe everyone a better answer than we've given.