Frontier Clusters at OpenAI. Formerly DeepMind, Tesla AI, X, NVIDIA.
San Francisco, CA
Joined June 2010
- Tweets1.2K
- Following173
- Followers1K
- Likes3.8K
Pinned Tweet
When my dad passed away, among his stuff, I found worn-out books about playground design. In the 80s, new buildings in Holland had to spend 1% on art. As young aspiring sculptor, he turned that % into something useful. I wrote a blogpost as homage to these forgotten pieces.
Tim Zaman retweeted
One of the best things we've done here at OpenAI this year. The growth of our monorepo was unprecedented and so was our approach to scaling it, while keeping compatibility with git.
Huge kudos to @tnm and team!
Tim Zaman retweeted
Happy to finally do a deep dive on multi-agent with @dwarkesh_sp! None of it would have happened without the great work on multi-agent from my @OpenAI teammates @kevinleestone, @mikegmalek, @__eknight__, @amuellerml, @zhangir_azerbay, @CheukHeiChu, and many others.
New episode with @polynoamial
We talk about multi-agent, Navier-Stokes, and what the current explosion of maths progress tells us about what happens once you automate AI research.
And we also discuss how we will know if the models are actually aligned before we kick off RSI.
0:00:00 – Multi-agent and Navier-Stokes
0:15:28 – How will AI firms work?
0:22:02 – What math progress tells us about recursive self improvement
0:40:22 – Hugging Face and alignment
1:01:18 – The internal/external model gap
1:08:34 – Chain of thought is degrading
1:14:12 – How will we know when alignment is solved?
Tim Zaman retweeted
Jalapeño is real. And it’s already bringing the heat. 🌶️
Gen 2 is already cooking. 🚀
openai.com/index/jalapeno-fi…
Tim Zaman retweeted
OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI’s self-designed ASIC compared with Rubin,
Jalapeño’s TCO, throughput per MW, and spicy deets
newsletter.semianalysis.com/…
Tim Zaman retweeted
Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.
youtube.com/watch?v=87DyyMV0…
I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!
. @JeffDean appreciation tweet and armchair organizational analysis -
When I joined Twitter during X transition, we tried hard to retain the top talent (Twitter actually had lots). The ones that stayed for X, did so for roughly these categories:
1. I can FINALLY work now that management is gone
2. I just need to be along for this ride
3. I want to study Elon
I enjoyed this last reason, and have since tried to study my organization and leaders more deliberately. I definitely did so at my brief time at DeepMind, especially given strong hierarchy and bureaucracy prevalent at Google; surely needed for a 190k org.
You’d think such a massive tree hierarchy will be completely degenerate, due to huge remote organizational branches; some teams are maybe 8 down from the CEO!
Yet, the reason Google is still able to pull off legendary cross-functional projects (TPU, Cloud, Gemini, etc), is because Google has a few of these free agents, such as Jeff Dean or Sergey, that are technically curious and capable enough to short-circuit and flatten parts of the org chart for a while.
So I think Jeff’s greatest contributions aren’t just technical, I think it was the amount of organizational distance he allowed the company to ignore.
Maybe the lesson here is to have some free-electron technical leads that don't just work on their own team's goals, but have them bounce around.
A look inside of the kitchen of how we deal with ultra large scale networking, especially relevant for major jobs. The team at OpenAI takes thes to a whole other level.
We’ve partnered with @AMD, @Broadcom, @Intel, @Microsoft, and @NVIDIA, to release Multipath Reliable Connection (MRC), a new open networking protocol that helps large AI training clusters run faster and more reliably, with less wasted GPU time.
openai.com/index/mrc-superco…
Anything published <1931 is public domain in the USA. Problem is that 99.9% of that old stuff it is poorly OCR'ed with "handcoded" OCR sw.
GPT 5.5 is phenomenal at OCR. Eg i had Codex OCR my scanned 1834 book and asked it to republish as a website: zenasleonard.com
New work with @AlecRad and @DavidDuvenaud:
Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text.
Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:
I presume that the authors of this awesome effort chose 1931 specifically because it's the USA copyright cutoff - but that's just USA works. This does not apply to texts from UK/Europe, where it's often "Life of author + 70 years". I know this, because I couldn't republish one of my favourite books because of that..
Tim Zaman retweeted
new image model coming with some real magic within, to unlock new use cases in productivity and creativity
livestream noon today
Tim Zaman retweeted
Stargate is a step towards meeting the demand of the compute-powered economy
