Tim Zaman
@tim_zaman
Followers
29K
Following
3K
Media
129
Statuses
984
Frontier Clusters at OpenAI. Previously DeepMind, Tesla AP, X/Twitter (head of AI Infra), NVIDIA.
San Francisco, CA
Joined June 2010
Personal news - I joined OpenAI! 🎉 We're going to build and launch the largest (and most delightful) supercomputers to power frontier AI research. DeepMind is obviously truly formidable and to the moon; and I had a great time - but I missed a smaller+tighter team I found in
102
41
2K
I have never had so many belly laughs at work as the last couple days. Productivity dropped to 0%. Sora 2 is so good. Cameo's are magical at first, then quickly the remixes are just as fun.
We are launching a new app called Sora. This is a combination of a new model called Sora 2, and a new product that makes it easy to create, share, and view videos. This feels to many of us like the “ChatGPT for creativity” moment, and it feels fun and new. There is something
1
0
18
Aerosolized Herbal Formulas & Organic Supplements Age Old Herbalism: All New Technology 5:1 Formulas
0
0
4
Awesome news also for NVIDIA and the community. The existing partnership has already influenced eachother's flagship offerings to a degree that's not widely appreciated. Very exciting we now put the friendship into writing! And the gpus are nice too. <3 NVL72.
Announcing strategic partnership with @nvidia for millions of GPUs — about as much compute as they've shipped in 2025 in total — and an investment up to $100B as these GPUs are deployed:
0
1
17
LIVE5TREAM THURSDAY 10AM PT
1
1
47
This seasons models will make vibe coding roll over as the new default. Saw someone handcoding with pen and paper at blue bottle today. They know.
@tim_zaman @simonw haha nice i love that it just rolls the default over. "coding" basically assume AI assistance as the default coding now, legacy coding becomes "handcoding".
2
2
38
When running your cluster in full-user-responsibility, one failure mode is a priority explosion, where everyone schedules hi pri. 5y ago before implementing hard rules (sad!), as a last ditch effort I introduced a rotating Trophy to shame the biggest offender.. ofc didn't work.
5
1
32
When I interviewed at NVIDIA in 2016, they put me in room "Led Zeppelin". That's when you know. Room next to it - Jethro Tull.. love at first sight
Fun fact: our model is called Kimi, but our company is Moonshot — named after Pink Floyd's The Dark Side of the Moon. We're a team of scientists who love rock (Radiohead, Pink Floyd) and film (Tarantino, Kubrick). A big reason I joined was because the taste just felt right.
2
0
52
The frontier clusters at OpenAI aren't just massive: every component and protocol is bleeding edge. Cluster management boils down to navigating complexity and discovering issues no one has ever seen before. We opened up a dedicated full stack position:
12
23
308
today we are introducing codex. it is a software engineering agent that runs in the cloud and does tasks for you, like writing a new feature of fixing a bug. you can run many tasks in parallel.
1K
3K
37K
Which reminds me, In a former life I spent 2 weeks sitting in front of one - https://t.co/cCVAvnlxP9
1
1
18
Rembrandt didn't paint by himself - as artists grew, they employed a workshop/studio style. Pupils did large parts of the work, signed by Rembrandt. Over time, this master-apprentice model sustains itself by the best apparentices graduating. And retirement by the old master.
Just released Codex, a software engineering agent that can work on many tasks in parallel. It runs on its own cloud-based compute infrastructure, and can fix bugs, answer questions about your code, run tests, etc.. Feels like a step towards the future of software engineering.
2
5
53
if you are interested in infrastructure and very large-scale computing systems, the scale of what’s happening at openai right now is insane and we have very hard/interesting challenges. please consider joining us! we could desperately use your help.
643
559
12K
Great casual overview of the complexities in frontier training runs & frontier systems at OpenAI.
@atootoon here is the full video: https://t.co/RjWJR0OpJ5
0
3
33
Notes on homelab netboot node provisioning through diskless liveboot into ramfs: https://t.co/P3CnmV9t67 - finally found some time to jot it down.
0
1
37
Hmmm 2 isn't very ambitious. DGX Spark is likely to support host chaining/`HOST_CHAINING_MODE=1` natively, hence the dual port:
The NVIDIA DGX Spark supports 200GbE RDMA clustering of the GB10-based mini (~1.1L) systems with 128GB of LPDDR5X, 20 Arm cores, and a Blackwell GPU in each https://t.co/J2qNePyPf2
3
0
26
Google's OCS (2022): https://t.co/oZtID3HC06 NVIDIA Silicon Photonics (2025):
1
1
11
The backend i had incomprehensibly optimized, so i learned they had to rewrite it after I left. 😅
2
0
35
Fun patent i wrote the backend for. Multimodal semantic search through billions of frames. The search box is a formula that takes in text and images. You start with text "image of a bear". This would give frames closest to this query. You can then click on frames you (dis)like,
Clip Search with Multimodal Queries @Tesla's US20240419724A1 patent introduces a groundbreaking visual search system that revolutionizes how self-driving AI engineers find critical driving scenarios in massive video datasets. Imagine being able to simply type "car cutting off
13
12
195