Yangqing Jia ✓

@jiayq · Berkeley, CA · joined 18 Apr 2009

Founder & CEO at Intent Lab. Built caffe, ONNX, PyTorch 1.0. Formerly Google Brain / Meta / Alibaba / Lepton AI / NVIDIA.

21 629Followers

385Following

701Posts total

681.8KViews on collected posts

Posts recentes

Yangqing Jia ✓@jiayq5 Aug

Excited to see the official launch of 224 Ventures! Also @OriolVinyalsML launching his new startup on the same day - the AI world is never short of excitements. 7.4K views · 24 likes · 2 reposts · 4 replies Open on X →

Yangqing Jia ✓@jiayq5 Aug

https://t.co/JlhgeihNZH 131.3K views · 452 likes · 30 reposts · 53 replies Open on X →

Yangqing Jia ✓@jiayq29 Jul

My 2 cents: one should build the right harness. If a harness dictates how the model works, it fights the model and you get brittleness. If it defines what the task is and checks the result, it enables long horizon task success. That's roughly how we got complex tasks done e2e at 11.7K views · 87 likes · 6 reposts · 2 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

@jiayq I didn't believe this was remotely possible when we first started chatting. Very excited to be on this journey with you. LFG! 562 views · 11 likes · 1 reposts · 0 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

We've spent our careers building large systems by hand, from open source to cloud infra, to our last startup. The theme was careful architecture, done one system at a time. What interests us now is building the oracle that produces a thousand of them. 3.9K views · 12 likes · 0 reposts · 2 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

We're working with design partners. If there's a system you want built, modernized, or made much faster, especially one that's been sitting on the list because it's too big to staff, send me a note for early access. If the mission resonates, we're hiring. Exciting paths ahead. 3.6K views · 14 likes · 0 reposts · 0 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

What sits above coding agents is still open: turning a vague intent into a real design, checking that the result holds up, and keeping the system evolving once it's running. And doing this all autonomously. That's the layer we're working on. 2.9K views · 9 likes · 0 reposts · 2 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

If it works, the economics of computer engineering will change forever. For 50 years the sensible move was to build one software and sell it to as many as possible, despite different needs. This can change now. The world will see a great deal more software, much more customized. 2.8K views · 9 likes · 0 reposts · 1 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

Models write code very fast now, but there's still real distance between code that runs and software you'd put in production and own for years. I don't think that distance is a limit of the models. It looks to me like a missing layer. 3K views · 9 likes · 0 reposts · 1 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

What matters most to us: all three came out of one general system. One method, applied to three very different problems. 3.1K views · 14 likes · 0 reposts · 1 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

3/ It also builds a super fast file system for a world where agents, not people, use the infra. Filesystems are notorius for its complexity. In this example, with formal verification, we catch a coding agent bug by identifying, and then fixing, a transient corrupted state. https:

0:03

3.7K views · 17 likes · 0 reposts · 3 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

2/ From a one-line prompt, it creates a database engine, autonomously going until all 6 million SQLite compatibilty tests passes. It worked from zero existing doc or code, ran every role a strong team would: architecting, coding, testing, review, QA. See roles in action: https://

0:07

5.2K views · 19 likes · 0 reposts · 3 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

1/ The fleet re-engineered the TRT-LLM inference engine and made it the fastest engine we know of for GLM 5.2: 102 to 647 tokens/sec on 2x Grace Blackwell nodes, a 6.3x speedup. This is with the fleet building and implementing optimization end to end. https://t.co/K0nnBhNlVh

9.7K views · 23 likes · 1 reposts · 1 replies Open on X →

Yangqing Jia ✓@jiayq28 Jul

We are starting Intent Lab, building an autonomous team we call "fleet" that turns intent into production software. Today we are sharing some early results: the fastest GLM5.2 inference engine, one shot database creation, and a fully verified agent filesystem. 280.2K views · 602 likes · 91 reposts · 76 replies Open on X →

Yangqing Jia ✓@jiayq22 Jan

@bingxu_’s VibeTensor is a rare verification of a fundamental question: can AI write complicated system code? The answer is “yes but” - it’s getting there very quickly, and if we get more principles right, then AI will fully beat human coders. It’s like AlphaGo in Jan 2015. 12K views · 45 likes · 5 reposts · 2 replies Open on X →

Yangqing Jia ✓@jiayq22 Jan

Just open-sourced VibeTensor — the first deep learning system fully generated by an AI agent, with 0 lines of human-written code: https://t.co/1kZ9v2tLW1 It’s a working DL system with RCU style dispatcher, a cache allocator and reverse-mode autograd. The agent also invented a ht

55.3K views · 237 likes · 42 reposts · 10 replies Open on X →

Yangqing Jia ✓@jiayq23 Mar 2025

A vivid example of long lasting specs. IIRC the padding config is designed for backward compatibility with PyTorch (2018) -> Caffe2 (2016) -> TensorFlow (2015) -> DistBelief (2011 or even earlier) -> Matlab (unknown when it was introduced) https://t.co/wvVmQLZpB6 14.4K views · 33 likes · 3 reposts · 4 replies Open on X →

Yangqing Jia ✓@jiayq21 Mar 2025

Hooooly cow PyTorch this is probably the worst API for a `pad` function I've encountered in my life. Why!? Or, you know what, I'd rather not know lol https://t.co/UHDkhounmu

131.2K views · 458 likes · 11 reposts · 25 replies Open on X →

Em comparação com contas do mesmo porte

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

Mediana de visualizações3 801esta conta924mediana para 10K–100K

Alcance, %17.57%esta conta3.62%mediana para 10K–100K

Engajamento, %0.40%esta conta1.52%mediana para 10K–100K

Métrica	Esta conta	Mediana para 10K–100K	Proporção
Mediana de visualizações por post	3 801	924	4.11×
Alcance (visualizações ÷ seguidores)	17.57%	3.62%	4.85×
Taxa de engajamento	0.40%	1.52%	0.26×

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

280.2K28 Jul

9.7K

5.2K

3.7K

3.1K

2.8K

2.9K

3.6K

3.9K

562

11.7K29 Jul

131.3K5 Aug

7.4K

Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.30%28 Jul

0.26%

0.43%

0.54%

0.49%

0.34%

0.36%

0.38%

0.39%

0.36%

2.14%

0.81%29 Jul

0.41%5 Aug

0.41%

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

What the audience does

Likes57.5%2 075 in total

Reposts5.3%192 in total

Replies5.3%190 in total

Quotes2.2%81 in total

Bookmarks29.6%1 069 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.

Posts recentes

Em comparação com contas do mesmo porte

Growth & engagement

Views per post

Engagement rate per post

What the audience does

Contas semelhantes