Applied math, supercomputing, trains, banjos, county courthouses, college football, county courthouses, Steely Dan, "HPC," mismatched, black cats
BHM/ Bmore/ A2/ STL
Joined August 2021
- Tweets6.6K
- Following864
- Followers937
- Likes117K
leighton retweeted
Your AI agent has a to-do list. Waiting shouldn’t be the longest item.
Cerebras’ Andy Hock joins theCUBE to discuss why inference speed matters for reasoning models and AI agents.
Where would less waiting make the biggest difference?
Watch the full conversation with Andy Hock and theCUBE:
youtube.com/watch?v=H2GnroPz…
leighton retweeted
The main thing I notice AIs do that humans don't is write lots of scripts running searches and verifications.
They don't work in a small search space, and they aren't using a gigabrain directly on a large search space. They're just writing lots of scripts to search large spaces that would be annoying to do.
When I watch them work, I mostly think, "Oh I probably should've been doing that. But I'm just so lazy, I would never crank out 1000 lines of Python doing a bunch of Grobner basis or group theory calculations just to test a hypothesis." Even if it would've been more reliable and less work overall.
Your computer runs billions of clock cycles per second, and AIs will actually use it for thinking.
I also think: If math had "industrial value", then this would've been systematized and you could train people to search for math solutions like an AI. Maybe Bel is superhuman, but lots of people could be an Astra if the computational approaches were written down and taught.
It's the difference between "stare at a paper, think, hope a pattern comes to mind" vs. "What is a good search space? Let me feed that into an SMT solver or CAS or custom C++ searcher." Of course, many people do that, but not as fluently and consistently as AI.
What does a superhuman level of mathematics look like? Here is one elementary example I could give to nonexperts.
Daniel Kane and Terence Tao showed (arxiv.org/abs/1702.00912) that, given a collection X of sets, the total number of ways to write A as a disjoint union of B and C, with A, B, C from X (disjoint means non-overlapping), is at most cardinality of X to the power 1.72598…, and this power is sharp.
To show this, they came up with an innocent-looking function B(x,y,z) = (xyz)^c of three variables, satisfying a certain arithmetic inequality and having prescribed boundary conditions. Proving that arithmetic inequality was really a headache.
There was a question whether one could improve this bound, i.e., maybe add some logarithmic correction terms, and one could kind of imagine what should be done (see comments on terrytao.wordpress.com/2017/…), but it was not clear whether it would be easy to do.
Now superintelligence improves this bound by introducing a certain monster function of three variables (see screenshot). Even if I had seen this function in my dreams, I would never have dared to verify that main arithmetic inequality.
The candidate function produced by SI gives a bound matching the known lower-bound examples up to a constant factor, including the sharp logarithmic correction.
All the details are here paatai.github.io/ai-proofs/b…
Our 2023 Gordon Bell Finalist with KAUST on seismic processing sc23.supercomputing.org/2023…
My first time really writing deep-in-the-weeds kernel code for our architecture. I'd like to think this was a stake in the ground on the power of SRAM-heavy arches for scientific workloads.
DRBE, my main focus for the last two years. I've been deeply fortunate to spend the last couple years drive this capability forward and now see it as a personal passion project
darpa.mil/news/2025/drbe-dev…
leighton retweeted
“If you’ve ever asked AI to do something and then drummed your fingers waiting for it to respond, you’ve run up against slow inference. To pick up the pace, Cerebras developed its own inference platform from the semiconductor wafer up.”
Cerebras is proud to be named in @FastCompany 's 2026 Next Big Things in Tech list.
My grandfather was a genealogist in his retirement. Since my son was born, I've wanted to continue with my granddad's work to document as much family history as possible. Thanks to Sol and Astra, my son now has 181 identified direct ancestors and 862 leads on more
leighton retweeted
Excited to partner with @GimletLabs to bring the ultrafast speed of @cerebras to their cloud!
Up to 3,000 tokens/sec, massive scale, and a shared vision for what the fastest builders in AI will unlock. 🚀
cerebras.ai/press-release/gi…
I know I'm not the first person to say this but Opus 5.5's improvement over Opus 5 is genuinely stunning
Johannes Doerfert was an incredible computer scientist, community builder, teacher, and contributor to LLVM and OpenMP. Thinking of his family and his colleagues at the lab and across the LLVM community today
It is with great sadness that we share the news of the passing of Johannes Doerfert. Johannes was one of the most prolific and respected contributors to the LLVM compiler project, and his loss will be deeply felt. blog.llvm.org/posts/2026-09-…
leighton retweeted
What folks say about math on here is very telling re confusion about overall AI trajectories. Could not be further from the truth that "math is solved."
As is going to happen and is already happening with lots of fields, the distribution of human effort is going to shift to different tasks, including interpreting and scrutinizing AI-generated solutions, as well as using them as fodder for further research (see: scientificamerican.com/artic…).
There is not some ceiling of the totality of possible math knowledge that we've lurched closer to. To a first approximation, we have functionally infinite headroom. Your modal expectation should be for that to remain the case even as AI tools accelerate progress.
leighton retweeted
When we started @cerebras, we knew we would have to solve difficult problems in yield, power, and cooling. What became clear was that making wafer-scale work would require rethinking almost the entire system.
Today, we answered questions from the community about those engineering decisions, how we run larger models, and what it takes to program the hardware.
2:15 Why hasn't NVIDIA gone wafer-scale?
4:10 How we build a working chip out of an imperfect wafer
8:07 What happens when a model won't fit on one wafer?
11:34 Why we chose SRAM over HBM
14:26 How much of a barrier is NVIDIA's CUDA moat?
Today was my son’s first day at daycare.
I don’t really know what my point is other than that we must always remember the point is to help humans and make the future of our species more prosperous than we can imagine. That’s what matters
