Yitong chen
88 posts
做了开源几年，关了几万个 Issue，合了很多几万个 PR 之后我觉得我还是能说上两句的。目前开源社区的主要矛盾就是：无限扩张的 AI PR 供给，同开源维护者有限而宝贵的注意力之间的矛盾。
中文
again?
Dmitry Rybin@DmitryRybin1
Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years. The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60. Chat with GPT 5.6 Pro where this was found: chatgpt.com/share/6a60b2eb…
English
@docmilanfar it's kind of hard to make this theorem really work in diffusion model
English
What do polar coordinates, polar matrix factorization, & Helmholz decomposition of a vector field have in common?
All are implied by Brenier’s Thm: a cornerstone of Optimal Transport theory. It’s a fundamental decomposition result that’s not widely known
1/6
English
@NikStarodub RAE outlier tokens seem to come from the outliers already in DINO. Since there are no visual token norm heatmaps for JiT and SiT, I guess the outlier tokens are less obvious than in RAE.
English
@NikStarodub RAE also is a class-conditional models, not MM-DiT. The author didn't show the JiT/SiT internal token nomr heatmap, but they indeed see the the improvement across them
English
Register tokens were introduced to fix artifacts in Vision Transformers (ViTs) (arxiv.org/abs/2309.16588)
In Diffusion Transformers (DiTs), we don't observe such outliers, so registers shouldn't help.
But they do... And especially for pixel-space DiTs.
Simply adding a few empty tokens can improve FID from 3.52 to 2.69.
English
GPT-5.6 Sol confirmed to be an extremely good model
Claude@claudeai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
@giffmana What's the connection between Curry finding a dead bounce and LLM solving ghost font?
English
You know that videoclip of that basketball player who's casually bouncing the ball, but once the bounce isn't perfect so he goes back and checks a few more times to confirm it's a dead spot in the floor?
Yeah:
Riley Goodside@goodside
I got GPT-5.6 Sol to read “Ghost Font” (a proposed method for making text legible to humans but not AI) by just telling it in what direction it looks like the noise is moving:
English
the fifth should be replaced with “不要比烂”.
think not what you've done better than someone else, but what other labs have done better than you.
oai and a\ are still leading the real frontier, glm can match opus with a <1T model, dpsk has incredibly high inference efficiency, we have very much work to do!
Xinyu Yang@Xinyu2ML
Why can Kimi ship K3? Let me tell my story. Earlier this year, I left academia for industry. I talked to a lot of companies along the way. Here's what I saw: 1⃣Arrogance. They believe the AI war is over, and they won. No hunger for the future, and no hunger for talent. 2⃣Restlessness. Young labs short on foundation, either rushing to catch the frontier or pivoting away from the competition. 3⃣Fear. Strong teams with real experience, but from the second tier, they can't quite bring themselves to aim for #1. 4⃣Misalignment. Everyone is optimizing for their own credit, but nobody really cares whether the company can reach AGI. Kimi was different. Over many conversations with the founders, the same thing came through every time: a raw, genuine hunger for AGI. I joined. The hunger was real. We shipped K3. This is only the beginning.
English
Yitong chen retweetledi
Kimi K3: The Opensource AI Frontier
Ahmad@TheAhmadOsman
Kimi K3 is ~3T parameters model with 1M context window btw
English
Yitong chen retweetledi
开发 DataLayer Studio 时，GPT‑5.6 Sol 给了我很大帮助：我真的没想到一个模糊的想法经过几百次的commits可以真的变成一个我自己每天用的东西： apps.apple.com/us/app/datalay…还有一个场景，我司的所有AWS的硬件和K8S，现在全由5.6 Sol负责管理，对我们这种没有专业运维的团队帮助很大！
中文
A hot take: the videogen pipeline of the future is an LLM-coded 3D/4D base layer plus a lightweight diffusion renderer. We already know how to build the latter, but the former requires VLMs to be much better verifiers. So if you want to unblock generation, work on perception
Brandon McKinzie@mckbrando
At last, I can share what the Blender video was making: a slop cannon. Here is the final simulation. All made from scratch in Blender by 5.6-Sol.
English