Yuandong Tian ✓

@tydsh · California, USA · joined 19 Dec 2009

Co-founder of @Recursive_SI. ex-Meta FAIR Director. ex-Google. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.

47 244Followers

950Following

1 168Posts total

778KViews on collected posts

Posts recentes

Yuandong Tian ✓@tydsh2 Sep

We got exact the same comments when we published Coconut😀. However, even with explicit CoTs, the model can still learn to hide the true thinking process. The ultimate solution is always to understand how the model works. 13.8K views · 95 likes · 3 reposts · 9 replies Open on X →

Yuandong Tian ✓@tydsh2 Sep

new: OpenAI & others quietly using loop transformers that don't show their 'thinking' when scaled up a leap forward on performance, but sparking concerns inside & outside OpenAI re: security as this takes off https://t.co/rBVuez1wUK

515.7K views · 731 likes · 56 reposts · 38 replies Open on X →

Yuandong Tian ✓@tydsh1 Sep

😅 so our StreamingLLM (the original attention sink paper) strikes back again? 😅 -- Caveat: it is in post-training setting. 23.2K views · 132 likes · 4 reposts · 4 replies Open on X →

Yuandong Tian ✓@tydsh28 Aug

Interesting findings by our automatic system😆 It finds that some FlashInfer kernels have numerical issues, provides a PR and accepted by human experts🤓 20.9K views · 104 likes · 5 reposts · 2 replies Open on X →

Yuandong Tian ✓@tydsh27 Aug

A fun small win for automated research: We found and helped fix some edge cases that could affect inference performance in vLLM and SGLang. In our last blog post, we described a reward hacking judge we developed for performance optimization tasks. While applying the judge to 85.5K views · 413 likes · 44 reposts · 12 replies Open on X →

Yuandong Tian ✓@tydsh21 Aug

赠Zeyuan Allen-Zhu🫡 世言大模惟精调，君执初心问真章。动力学中观涨落，表达性里辨行藏。万里独征云和月，所慕岂在功与赏。他年若问登临处，格物致理照新航。突き破れ、扉の向こうへ！（source: 鋼の錬金術師 ED，《扉の向こうへ》） 46K views · 248 likes · 9 reposts · 7 replies Open on X →

Yuandong Tian ✓@tydsh11 Jun

Early results from Recursive 🚀🚀 SotA results from our open-ended knowledge discovery system: 1️⃣NanoChat 5min pre-training (0.9372 bpb -> 0.9109 bpb, 2.8% lower Bits-Per-Byte than long-standing community SoTA) 2️⃣NanoGPT SpeedRun (79.7s -> 77.5s, 2.8% faster than long-standing 72.9K views · 369 likes · 37 reposts · 8 replies Open on X →

Em comparação com contas do mesmo porte

6 posts dos últimos 90 dias, ao lado da faixa de 10K–100K seguidores. aparece para muita gente, mas poucos desses espectadores reagem.

Mediana de visualizações34 617esta conta924mediana para 10K–100K

Alcance, %73.27%esta conta3.62%mediana para 10K–100K

Engajamento, %0.57%esta conta1.52%mediana para 10K–100K

Métrica	Esta conta	Mediana para 10K–100K	Proporção
Mediana de visualizações por post	34 617	924	37.5×
Alcance (visualizações ÷ seguidores)	73.27%	3.62%	20.2×
Taxa de engajamento	0.57%	1.52%	0.37×

Outras contas desta faixa → Comparar com outra conta → Como estas referências são construídas →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

72.9K11 Jun

46K21 Aug

85.5K27 Aug

20.9K28 Aug

23.2K1 Sep

515.7K2 Sep

13.8K

Last 7 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.57%11 Jun

0.57%21 Aug

0.56%27 Aug

0.53%28 Aug

0.60%1 Sep

0.17%2 Sep

0.78%

Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 1.52%.

What the audience does

Likes62.3%2 092 in total

Reposts4.7%158 in total

Replies2.4%80 in total

Quotes2.5%85 in total

Bookmarks28.1%945 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.