Junyang Lin

building p7k @pragmatik_labs | ❤️ 🍵 ☕️

[Junyang Lin @JustinLin610]

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

[Junyang Lin @JustinLin610]

i don’t think it must have more upvotes. the key issue is why those above it are seemingly no better or even much worse. this tweet is not against hf daily papers but against the filthy corruption. so many papers with a lot of upvotes but no real discussion, that might be something that u should think how to improve. i know a lot of people just asked their friends or even bought upvotes, and clicking an upvote has no cost and effort at all, just like today’s github stars.

tbh i love hf daily papers as it helps a lot with my very limited attention and time to catch up with today’s ai. but now I just feel sad it is becoming worse

[Junyang Lin @JustinLin610]

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

[Junyang Lin @JustinLin610]

@ivanfioravanti idk hahah! i just think it is meaningless to have this type of paper without building good models as well as good products from scratch, but apparently it is too costly. i now focus on building company and business.

[Junyang Lin @JustinLin610]

@kugwzk1 links? i missed them sry. i ll take a look!

[Junyang Lin @JustinLin610]

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

[Junyang Lin @JustinLin610]

nyc style

[Junyang Lin @JustinLin610]

@bookwormengr somehow too personal. i don't like things big but too sparse taking up memory and storage. moreover, it might be good for ppl but maybe not that helpful for reasoning, which is the bottleneck of today's intelligence

[Junyang Lin @JustinLin610]

@bookwormengr temporarily i got no interest in it but i think i need to learn more about it.

[Junyang Lin @JustinLin610]

@kugwzk1 which is better

[Junyang Lin @JustinLin610]

@bookwormengr here is what i thought when i first met them:

v4
csa, hca: what and why?

v4.1
ced: yoco finally comin' out, nice but too risky
csa: together with ced? make sense
engram: optional and fine

other things i don't care that much

[Junyang Lin @JustinLin610]

@lily_gpupoor shanghai hahah

[Junyang Lin @JustinLin610]

i mean, the ranking is corrupted and causes me much trouble filtering papers

[Junyang Lin @JustinLin610]

@bookwormengr v4 is too complex, v4.1 is complex but understandable

[Kaiyue Wen @wen_kaiyue]

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!

[Junyang Lin @JustinLin610]

@altryne @thsottiaux @clairevo damn it is a real button

[Junyang Lin @JustinLin610]

lol

[Sebastian Raschka @rasbt]

Update: OpenAI just added a Jev clone 😆

[Junyang Lin @JustinLin610]

hope smarter than opus 5.5

[OpenAI @OpenAI]

GPT-6.1 Sol: near-Astra intelligence for a fifth of the price.

It’s the most cost-efficient model for its performance available today.

[Junyang Lin @JustinLin610]

who the hell r u

[Junyang Lin @JustinLin610]

@TianbaoX cool