←
Junyang Lin3,242 posts
Junyang Lin avatar

Junyang Lin

@JustinLin610

building p7k @pragmatik_labs | ❤️ 🍵 ☕️

◷ Joined December 2015▧ 168 Photos and videos
Junyang Lin avatar
Junyang Lin@JustinLin610·Aug 11

a life update: i started a new company called Pragmatik (p7k) Labs (语用科技) in shanghai, focusing on the research of next-generation agents across digital and physical worlds. thanks to Gaorong Ventures and HSG (红杉中国 & 高榕创投) for co-leading this round, and to Tencent (腾讯) and Shanghai Engine Fund (上海未来产业基金) for the support. @pragmatik_labs · pragmatik.com

Junyang Lin avatar
Junyang Lin@JustinLin610·10h

notably this year spatial reasoning has been a clear gap between vl models (especially open weight ones and o a g). it is a bit hard to use current vl models to build very high quality physical agent, and this might be one of the reasons

Junyang Lin avatar
Junyang Lin@JustinLin610·10h

6.2 looks impressive (but as a boston fan i don't feel good about this example lol...

GPT-6.2 Astra is so good at spatial reasoning asked it to map player positions onto a top-down court

0:17
Junyang Lin avatar
Junyang Lin@JustinLin610·Nov 17

oh when u build eval, make sure it is without (big) mistakes (fewer samples is totally fine). having been tired of wrong evals leading models to wrong directions

Junyang Lin avatar
Junyang Lin@JustinLin610·Nov 17

anthropic.com/research/inves… just read this and i think it is very inspiring. we care too much about the so called capabilities but care too less about the unintended actions that might cause big troubles. if anyone thinks about new evals, you can build towards this direction

Junyang Lin avatar
Junyang Lin@JustinLin610·Nov 14

he is really a good person. no matter what happened, he doesn't deserve all these.

Junyang Lin avatar
Junyang Lin@JustinLin610·Oct 14

most cases i saw for argon are about visualization. how does it perform on other tasks?

Junyang Lin avatar
Junyang Lin@JustinLin610·Oct 3

it is always super hard to find a good angle to take a satisfactory picture. just step back and blow up like 3x or even 5x, u will find that the picture looks much better

Kaiyue Wen avatar
⟳ Junyang Lin retweeted
Kaiyue Wen@wen_kaiyue·Oct 2

Do you really need to finish every rollout in RL for LLM? Not if you make your critic better—and trust it! We propose Actor-Critic with Action Chunking (AC2): use a learned critic to score chunks of tokens => only need partial rollouts ⇒ train faster than GRPO!

Post attachment
Junyang Lin avatar
Junyang Lin@JustinLin610·Oct 3

just checked hf daily papers sept ranking, dpsk v4.1 flash only has 191 upvotes? ridiculous, it is the greatest paper of sept