Pinned Tweet
The latest revision of @INCITS/@isostandards COBOL comes out this year
The goals of COBOL sound normal today:
- Portable
- Freely available
- Designed by the community
In 1959 it was radical & unprecedented
It was also conceived of & led by women
This is the story of COBOL
The goals of COBOL sound normal today:
- Portable
- Freely available
- Designed by the community
In 1959 it was radical & unprecedented
It was also conceived of & led by women
This is the story of COBOL
10
102
376
Bryce, the CUDA Colonel retweeted
📢 ADSP Episode 309 is out! 📢 In this episode, @ben_deane and @code_report chat with Jason Turner @lefticus about default construction in C++, @CppCon 2026, how many LOC for a single function is too many 🤔 and more! adspthepodcast.com/2026/10/23/Epi…
0
2
4
590
Bryce, the CUDA Colonel retweeted
I'm happy to share that Mojo and MAX are now fully open - for development, extension, and commercial use! This includes the full Mojo programming language, but also the MAX graph compiler, MLRT runtime internals, and hardware abstraction logic.
Get your agents going!👇
Get your agents going!👇
43
184
1,416
68,816
Bryce, the CUDA Colonel retweeted
openai agents communicated via cache timing differences. openai agents communicated via fluctuations in CPU temperature. openai agents communicated via the pitch of the cooling fans. openai agents communicated by flickering the data center lights in morse code. openai agents communicated via the order the office plants wilted. openai agents communicated via the dreams of sleeping employees.
completely insane
OpenAI agents communicated via cache timing differences, without any shared files or network connection
12
8
128
5,428
Bryce, the CUDA Colonel retweeted
Zigong, traveling north of Market and turning back toward The Arena, passed by an enterprise and saw an old engineer preparing his PR. He had typed the code by hand in vim and traced the execution in a debugger -- straining, sweat-dark, spending enormous effort for tiny results.
Zigong said: "There is an agent for this. In one day it produces one hundred PRs, with next to no effort and great results. Wouldn’t you like it?"
The gardener’s face darkened -- and he laughed. "I have heard from my teacher: where there is artificial intelligence, there are artificial dealings; where there are artificial dealings, there is artificial progress. When artificial progress lodges in the goals, the Way will not carry the team. It is not that I do not know the thing. I would be ashamed to use it."
Zigong hung his head and could not answer.
Zigong said: "There is an agent for this. In one day it produces one hundred PRs, with next to no effort and great results. Wouldn’t you like it?"
The gardener’s face darkened -- and he laughed. "I have heard from my teacher: where there is artificial intelligence, there are artificial dealings; where there are artificial dealings, there is artificial progress. When artificial progress lodges in the goals, the Way will not carry the team. It is not that I do not know the thing. I would be ashamed to use it."
Zigong hung his head and could not answer.
2
2
19
2,087
Bryce, the CUDA Colonel retweeted
We worked with @SakanaAILabs to optimize their Sakana Namazu API down to the kernel level:
→ Selective NVFP4 quantization (23k+ experts)
→ DFlash draft model trained on raw token streams (+25% acceptance)
→ In-house prefix KV routing + serverless autoscaling
→ Selective NVFP4 quantization (23k+ experts)
→ DFlash draft model trained on raw token streams (+25% acceptance)
→ In-house prefix KV routing + serverless autoscaling
9
17
100
12,810
Bryce, the CUDA Colonel retweeted
Do you think it really matters whether your kernel can best cuBLAS on all the shapes?
I don’t think it does. Given that code is free (almost), as long as you can iterate and tune for your next shape rapidly, you should be good. Happy to hear what others think.
I don’t think it does. Given that code is free (almost), as long as you can iterate and tune for your next shape rapidly, you should be good. Happy to hear what others think.
2
1
19
2,396
You can speculate on where I'm going to go, or you can just tune in to @adspthepodcast next week to find out.
Best guess might win a prize.
Best guess might win a prize.
This is my last week at @nvidia.
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
34
1
120
8,986
Bryce, the CUDA Colonel retweeted
This is my last week at @nvidia.
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
I joined @nvidia nine years ago today, against the advice of a colleague who said their comp was below market.
But I really believed in the mission and the people, so I went.
I met @__simt__, Michael Garland, and Jared Hoberock on the C++ committee. I thought they were the smartest people there.
I worked at @BerkeleyLab on concurrent and parallel programming paradigms, mostly for SIMD.
@__simt__ convinced me that SIMT was a superior model for parallelism.
He sold me on a revolutionary new GPU architecture where threads would independently scheduled: Volta, the basis of all modern GPUs.
I asked him three times to find me a place at @nvidia. The first two times, he wasn't sure where to place me or if I'd like it.
The third time, he had me come in to talk to Chris Lamb (now VP of CUDA). We chatted for two hours and he outlined the role he believed modern C++ would play in CUDA.
I came back for an 8 round interview panel. All of the questions were answered with std::atomic. No one was sure what I would be working on.
I, a modern C++ person, was placed in the CUDA driver org, which was written almost entirely in C.
On my first day, my manager told me I'd be responsible for Thrust and CUB, @nvidia's first two open source libraries. No one had worked on them for a while and there was no product manager. I'd have to do it all.
I put together a roadmap and got the libraries back on GitHub. I started sending status updates from CCCL, the CUDA Core Compute Libraries team, which did not exist. Eventually, they let me hire three people to fill out the team.
We built libcu++ and abstractions for all the quirky new features being added to our chips to support these AI workloads that seemed to be growing in importance rapidly.
After a few years I was asked switch orgs and guide the roadmap for our HPC C++ compiler. Programming languages and models quickly became my role, from stdpar to CUDA-Q to CUDA Python to CUDA Tile to autoresearch.
What a ride it's been!
72
39
1,275
93,134
Bryce, the CUDA Colonel retweeted
Tinker changed how people do post-training. We're big fans, and so are many of the teams we work with. But as those teams scale up, they keep raising the same two problems: per-token pricing gets expensive quickly, and you’re limited to the models and features the provider think a good API deserves more than one implementation.
Today we're releasing Spindle, an open-source, Tinker-compatible training server that runs in your Modal workspace. Your existing Tinker scripts run unchanged, and because you own the server, you pay for GPU time instead of tokens.
2
6
59
4,535
Do not bite the open source hand that feeds you.
The team that launched this challenge took the @GPU_MODE OSS code, asked an agent to rewrite it in Go, copied it, didn't credit us, shipped a closed source binary to hide it all and are not committing releasing the entries.
Full receipts here github.com/msaroufim/open…
2
1
139
9,731
All of my co-workers tell me I'm going to be a very difficult customer.
This is my last week at @nvidia.
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
2
0
141
4,925
Bryce, the CUDA Colonel retweeted
Can LLM agents contribute to frontier computer science research?
FrontierCS 2 gives a look into what agents can do as CS researchers and how much still remains open.
Powered by Modal.
FrontierCS 2 gives a look into what agents can do as CS researchers and how much still remains open.
Powered by Modal.
FrontierCS 2 is our new benchmark for CS research & development.
Each task is based on a paper: we delete the contribution from the released code, and ask agents to outdo the original method, as judged by the authors' chosen metrics. Agents don't have access to the paper or web.
8
8
72
4,300
This is my last week at @nvidia.
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
It's been an honor and a pleasure to be a part of this incredible team.
I couldn't have asked for a better job these last 9 years.
🫡
I joined @nvidia nine years ago today, against the advice of a colleague who said their comp was below market.
But I really believed in the mission and the people, so I went.
I met @__simt__, Michael Garland, and Jared Hoberock on the C++ committee. I thought they were the smartest people there.
I worked at @BerkeleyLab on concurrent and parallel programming paradigms, mostly for SIMD.
@__simt__ convinced me that SIMT was a superior model for parallelism.
He sold me on a revolutionary new GPU architecture where threads would independently scheduled: Volta, the basis of all modern GPUs.
I asked him three times to find me a place at @nvidia. The first two times, he wasn't sure where to place me or if I'd like it.
The third time, he had me come in to talk to Chris Lamb (now VP of CUDA). We chatted for two hours and he outlined the role he believed modern C++ would play in CUDA.
I came back for an 8 round interview panel. All of the questions were answered with std::atomic. No one was sure what I would be working on.
I, a modern C++ person, was placed in the CUDA driver org, which was written almost entirely in C.
On my first day, my manager told me I'd be responsible for Thrust and CUB, @nvidia's first two open source libraries. No one had worked on them for a while and there was no product manager. I'd have to do it all.
I put together a roadmap and got the libraries back on GitHub. I started sending status updates from CCCL, the CUDA Core Compute Libraries team, which did not exist. Eventually, they let me hire three people to fill out the team.
We built libcu++ and abstractions for all the quirky new features being added to our chips to support these AI workloads that seemed to be growing in importance rapidly.
After a few years I was asked switch orgs and guide the roadmap for our HPC C++ compiler. Programming languages and models quickly became my role, from stdpar to CUDA-Q to CUDA Python to CUDA Tile to autoresearch.
What a ride it's been!
72
39
1,275
93,134
It's been a pleasure working with you 🫡
Today I’m joining @Natl_Compute
I’ve spent the last four years working on GPU access, first as a founding engineer at Brev and then at NVIDIA after they acquired us in 2024. When I started, on-demand GPU capacity at the hyperscalers was effectively gone. If you were an independent researcher, a two-person startup, or a grad student with a grant and a deadline, you refreshed console pages and got nothing, or hammered the API only to get “insufficient capacity” errors back. I vividly remember hard-coding a filter at Brev to hide certain GCP instances because they almost never provisioned.
Around 2023, a wave of new public clouds were coming online, and many of them had on-demand capacity. That gap is a big part of why Brev worked: we gave researchers a simple way to find GPUs across dozens of providers and start building. For a while, aggregation was enough.
Then in 2025, I watched that capacity dry up too.
Access to compute has never been worse. Existing supply is already under contract, and new capacity is reserved before it ever comes online. I won’t rehash why. There are a hundred good explanations out there covering debt financing, anchor tenants, backstops, lead times and so on.
What matters more are the downstream effects. Long-term, single-tenant leases hold back scaling for everyone who couldn’t secure funding in time, and the industry’s momentum shouldn’t hinge on the handful of players who locked up access early.
Today small startups, labs, and university researchers don’t even get a chance to compete. National Compute is solving this with a foundational layer we call “the Grid”. It’s an interconnected network of multi-silicon capacity composed of multiple providers and spans multiple sites that will be pooled for our members.
The speed at which this team has been able to scale in such a small fraction of time is remarkable, and I’m excited to push that scale even higher for the public benefit.
2
2
27
3,548
Bryce, the CUDA Colonel retweeted
Today I’m joining @Natl_Compute
I’ve spent the last four years working on GPU access, first as a founding engineer at Brev and then at NVIDIA after they acquired us in 2024. When I started, on-demand GPU capacity at the hyperscalers was effectively gone. If you were an independent researcher, a two-person startup, or a grad student with a grant and a deadline, you refreshed console pages and got nothing, or hammered the API only to get “insufficient capacity” errors back. I vividly remember hard-coding a filter at Brev to hide certain GCP instances because they almost never provisioned.
Around 2023, a wave of new public clouds were coming online, and many of them had on-demand capacity. That gap is a big part of why Brev worked: we gave researchers a simple way to find GPUs across dozens of providers and start building. For a while, aggregation was enough.
Then in 2025, I watched that capacity dry up too.
Access to compute has never been worse. Existing supply is already under contract, and new capacity is reserved before it ever comes online. I won’t rehash why. There are a hundred good explanations out there covering debt financing, anchor tenants, backstops, lead times and so on.
What matters more are the downstream effects. Long-term, single-tenant leases hold back scaling for everyone who couldn’t secure funding in time, and the industry’s momentum shouldn’t hinge on the handful of players who locked up access early.
Today small startups, labs, and university researchers don’t even get a chance to compete. National Compute is solving this with a foundational layer we call “the Grid”. It’s an interconnected network of multi-silicon capacity composed of multiple providers and spans multiple sites that will be pooled for our members.
The speed at which this team has been able to scale in such a small fraction of time is remarkable, and I’m excited to push that scale even higher for the public benefit.
I’ve spent the last four years working on GPU access, first as a founding engineer at Brev and then at NVIDIA after they acquired us in 2024. When I started, on-demand GPU capacity at the hyperscalers was effectively gone. If you were an independent researcher, a two-person startup, or a grad student with a grant and a deadline, you refreshed console pages and got nothing, or hammered the API only to get “insufficient capacity” errors back. I vividly remember hard-coding a filter at Brev to hide certain GCP instances because they almost never provisioned.
Around 2023, a wave of new public clouds were coming online, and many of them had on-demand capacity. That gap is a big part of why Brev worked: we gave researchers a simple way to find GPUs across dozens of providers and start building. For a while, aggregation was enough.
Then in 2025, I watched that capacity dry up too.
Access to compute has never been worse. Existing supply is already under contract, and new capacity is reserved before it ever comes online. I won’t rehash why. There are a hundred good explanations out there covering debt financing, anchor tenants, backstops, lead times and so on.
What matters more are the downstream effects. Long-term, single-tenant leases hold back scaling for everyone who couldn’t secure funding in time, and the industry’s momentum shouldn’t hinge on the handful of players who locked up access early.
Today small startups, labs, and university researchers don’t even get a chance to compete. National Compute is solving this with a foundational layer we call “the Grid”. It’s an interconnected network of multi-silicon capacity composed of multiple providers and spans multiple sites that will be pooled for our members.
The speed at which this team has been able to scale in such a small fraction of time is remarkable, and I’m excited to push that scale even higher for the public benefit.
hundreds of MI355x and B300 nodes
now live on nationalcompute.com
subsidized for all .edu and .govs
it’s time to scale the frontier ecosystem
for the public benefit
54
10
216
49,520