exploring, experimenting, investing
Palo Alto, CA
Joined March 2019
- Tweets587
- Following1.2K
- Followers422
- Likes2K
Pinned Tweet
a peak inside the inner workings of math ai
Some more details on our work @HarmonicMath : arxiv.org/abs/2510.01346
By the way, we’re hiring Research Engineers. This is a high-leverage opportunity to join a small and world-class team. High agency is required. DM me leads
when your two worlds come together unexpectedly, incredible things happen!
humanfia 🤝 princeton
We have now also completed a full Lean formalization of the Hamilton-Perelman proof of Thurston’s Geometrization conjecture!
Ben, Yuan, Ziyang and I completed this in collaboration with the Humanfia team at NVIDIA @juihuichung @LigengZhu
Mark Kim retweeted
🚀 Humanfia × Geometrization: Perelman's Proof, Now Fully Formalized in Lean
The Humanfia team and Yuan Liao, Ziyang Qin, Ayush Khaitan , Ben Chow, and have completed a full Lean formalization of the Hamilton–Perelman proof of Thurston's Geometrization conjecture! It goes beyond Poincaré and classifies every 3-manifold. Mathematicians needed years to digest this proof. Now a machine has verified it line by line within weeks!
The progress of Humanize agentic math:
26-07: Perfect score in IMO 2026
26-08: 670/672 in PutnamBench
26-09: 100% in PutnamBench and 1st place in LeanEval (Surpassed by Axiom later)
26-10: 1st place again in LeanEval with upgraded Humanize workflow
26-10: Completed the Lean formalization of the Hamilton–Perelman proof of Thurston's Geometrization conjecture.
We are increasingly convinced: when everyone has tokens of similar intelligence, the agent workflow is what decides how far AI can go in frontier mathematics.
Tokens are「transforming」 the math.
🌌Geometry library: github.com/qinz1yang/differe…
⚙️ Humanize: github.com/humanfia/humanize
#AI #LLM #Humanize #Mathematics #Lean4 #Agent #Humanfia #Geometrization
Mark Kim retweeted
Two years ago I argued we needed an autoformalizer.
Now the time has come to formalize all of mathematics.
We need to unite behind a single Mathematics Autoformalization Project that turns all known mathematics into one Lean library.
@jdlichtman @patrickshafto
Mark Kim retweeted
🚀 Kernel Design Agents to Optimize Kimi Delta Attention (upto 2.96x)
When the KDA (Kernel Design Agents) project was initiated, its name happened to coincide with another KDA (Kimi Delta Attention). More than once, people asked us , "Can we use KDA to write a version of KDA?" Today, right at a moment with a full moon good for shooting, we are thrilled to share the latest results of KDA(gent)-v0.6: using KDA(gent) to optimize KDA(ttn)!
KDA-v0.6 achieves this through
* Humanize2 Flow & LLM Collaboration Optimization: Introducing multi-round iterative evolution with Flame Chase, GPT-5.6-Sol, and Fable-5, paired with an automated Workspace cleanup mechanism to significantly broaden the operator search space.
* Multilingual Domain-Specific Skill Stack Support: Fully supporting CuTe-DSL, CUDA C++, as well as recent Agent-Native CAKE IR and TIRx. Combined with customized diagnostic tools (such as CuTe's dedicated IKET pipeline analyzer and TIRx's CPU-side numerical simulation and static checks), achieving deep performance bottleneck diagnostics.
* Self-Evolving Kernel Wiki: Building a dynamically updated knowledge base that automatically cleans up invalid/erroneous implementations, corrects classification tags, and streamlines search indexing, significantly improving the Agent's knowledge retrieval quality and code generation accuracy.
KDAgent ultimately optimized the KDAttn kernel to 2.96x of FlashKDA, and we face a lot of unexpected hacks during the optimization. More details shared in our blogs nvlabs.github.io/kda/blog/20…
Release results of KDA for KDA: github.com/NVlabs/kda/tree/2…
Community & Production Deployment:
All kernels are open-sourced for community verification! For deploying NVIDIA Agentic CUDA acceleration in production environments, we recommend using the production-level version rigorously verified by FlashInfer.
By the Team: Dongyun Zou, Yixin Dong, Hongyi Jin, Junxian Guo, Yahui Cui, Avery Huang, Zihao Ye, Junru Shao, Changye Li, Zijian Zhang, Sihao Liu, Song Bian, Ligeng Zhu.
The man, the legend!
For those of you following the recent developments around Navier–Stokes, this should be an incredible lecture.
This Friday at Harvard, Javier Gómez-Serrano, who has been working with Google DeepMind on the problem, will present and analyze OpenAI’s recent breakthrough.
I spoke with him today, and he’s currently working through the solution in preparation for the lecture. This will be a good one.
For those of you following the recent developments around Navier–Stokes, this should be an incredible lecture.
This Friday at Harvard, Javier Gómez-Serrano, who has been working with Google DeepMind on the problem, will present and analyze OpenAI’s recent breakthrough.
I spoke with him today, and he’s currently working through the solution in preparation for the lecture. This will be a good one.
Carlota Perez covers this well in Technological Revolutions and Financial Capital.
Every technological revolution since 1771 has run roughly half a century.
Installation is fast because financial capital rushes into the first half of the S-curve, causing a bubble.
Deployment is slow because every business has to assess the cost of adopting the new technology and the economy built around the old paradigm has to be rebuilt.
Sam Altman admits he was wrong on AI's timeline and says society and the economy will adapt more slowly
"I thought when we got to GPT-4, which was back in 2023, that very quickly after that there was going to be much more disruption, software businesses up for grabs right away, than it turned out to be."
"I think I was wrong about a few things, but one in terms of the speed: the economy just has so much inertia."
"People keep doing the same things, buying from the same company, wanting to use their tools the same way. I think this is actually a positive in many ways, and it's going to make this big transition go smoother and slower. I'm grateful for it."
"But it means we've all been too ambitious on timelines. Even with this incredible technology, society and the economy will adapt more slowly."
It makes a ton of sense once you connect it to what neural networks actually do.
They capture data through successive layers of mathematical equations.
Recent progress in AI has made “mathematizing” the world dramatically easier and cheaper.
If we can mathematize human behavior at a sufficiently granular level, we can, in theory, build accurate human simulations. Those simulations become the building blocks for simulating the world itself.
We’re moving from top-down models to agent-based ones.
markkim.bearblog.dev/mathema…
markkim.bearblog.dev/superin…
Had a similar experience with Simile where the longer you dwell on the simulation concept, the more you realize that there's something here.
And didn't realize Andrej was involved, wild.
It crystallized when I learned SharkNinja is experimenting with AI-generated "synthetic personas" even before doing focus groups. The leading consumer products team, at least on the public markets, doesn't even turn to real humans first.
We'll see how effective this type of research is by 4Q26 when the first synthetic-influenced products land. The future is strange.
Wrote this last February after seeing a demo of TTT being applied to video.
markkim.bearblog.dev/test-ti…
Test-time training shifts compute from pre-training to adaptation at inference. It handles distribution shifts more efficiently and amortizes training cost across the model's entire lifetime.
‼️huge ssi news.
ilya is about to take his first tentative steps out of the age of research and back into the age of scale. it’s time to smell what ssi is cooking.
ssi have built a small reasoning engine that can compete with much larger training runs because his data is better curated to meta learning. but, more importantly.
we’re about to step into the era of TTT (test time training. gradient descent happening in real time to solve your problems). so instead of a context window you get actual learning.
and because it’s so sample efficient it can be trained on hard to verify tasks that other paradigms can’t touch. everyone else’s weights are frozen, they struggle out of distribution. ssi have created something that has a bundle of knowledge but can truly learn in real time and use that to your advantage.
current approaches are trying to hack their way to ‘learn’ with memory tricks, this thing will updates its weights, remember key lessons, and finally feel like a human level reasoner. this is a huge paradigm shift from the king. early results are very impressive. we can stop watching memento on repeat.
it’s learning all the way down, the descent is real.
Something interesting that I learned today.
TPUs will naturally struggle with training Kimi K3 and other linear architectures due to the mismatch in runtime between the MXU and VPU.
The MXU needs to wait for the sequential vector work from the VPU before running the matrix multiplication.
Inference shouldn't be as big of an issue though except TTFT.
The Helios rack is gonna be insane when it comes out.
Meta's engineers are on site helping design it
Anthropic will help with mega kernels
xAI has already deployed AMD chips with expert parallelism and RL workflows
Exclusive: Advanced Micro Devices and Anthropic have signed a deal for tens of billions of dollars’ worth of artificial-intelligence servers, strengthening AMD’s competitive position against Nvidia wsj.visitlink.me/IQPuOi
Mathematizing new areas and being able to do so cheaply is one of the most exciting futures for AI.
Neural networks can become the substrate to capture complex phenomenon. Humans may not understand the data, but AI will pick up on subtle patterns that's captured in its weights.
You build hierarchies on top of one another and that's how you build models of the cell and solve problems like the Three-body problem
Most existing math will be properly analyzed and digested soon by AI.
The next step will be the mathematization of new domains.
When the mathematical abilities of AI are much more sophisticated, cheaper, and faster, it will be able to capture domains that are too complicated today: biology, AI, physics, etc.
The quantity and depth of new, applied math will quickly surpass what we have today.
goat when it comes to s-1 analysis
SpaceX filed its S-1 last week, and my partner @ausw2000 read all 318 pages (and made all the charts) so you don’t have to. The result: three businesses in one — Space, Connectivity (Starlink), and AI. AI is set to become the largest revenue segment within a year.
The company did $18.7B in FY25 revenue, up 33% YoY, and is reportedly expecting a $1.75T valuation that would make it a top-10 company globally.
Another datapoint: Napoleon was tutored by Laplace.
Napoleon considered himself to be a “scientist in uniform” and would study physics, engineering, metallurgy, and celestial mechanics.
He would often summon Laplace to discuss physics.
Aristotle tutored Alexander the Great.
And he took over the world.
For 2,000 years, that was the deal. 1-on-1 tutoring was the cheat code, and only the royals afforded it.
Bloom proved it in 1984. 1-on-1 tutoring routinely lifts learners from the 50th percentile to the 99th.
Now with AI, you have a tutor for anything you want to learn. Quantum mechanics at 2AM. French conjugation. How the postgres query planner works. How to negotiate a term sheet. Literally everything.
The upside is that it's infinitely patient & infinitely knowledgeable. It also cost less than most of the subscriptions you will get.
Learning has never been this easy in human history. It's in fact now your fault if you are not learning something new everyday.
Marc Andreessen on what just changed:
Ecstatic to be supporting Nikola and the team!
DeepInfra has raised its $107M in Series B funding 🚀
AI is moving from training to production-scale deployment, and inference is becoming the system constraint.
DeepInfra was built for this shift — scaling high-throughput inference for open-source and agent-driven workloads. Grateful to our investors and partners, co-led by @500GlobalVC and @gharik
Mark Kim retweeted
The Future of Mathematics Symposium is being held on May 1st-2nd!
We have a truly exceptional group of speakers, with multiple Fields Medalists including Terence Tao.
Tune in for livestream 9am-5pm PST (link below)
Terence Tao opens this interview describing how Kepler brute-forced planetary laws from Brahe's dataset. Tried random relationships for 20 years until one fit.
Now imagine millions of AI mathematicians doing the same with observations of human behavior. Formalizing what's easy to observe but tedious to capture.
That's where I think mathematical superintelligence actually leads: markkim.bearblog.dev/mathema…
The Terence Tao episode.
We begin with the absolutely ingenious and surprising way in which Kepler discovered the laws of planetary motion.
People sometimes say that AI will make especially fast progress at scientific discovery because of tight verification loops.
But the story of how we discovered the shape of our solar system shows how the verification loop for correct ideas can be decades (or even millennia) long.
During this time, what we know today as the better theory can often actually make worse predictions (Copernicus's model of circular orbits around the sun was actually less accurate than Ptolemy's geocentric model).
And the reasons it survives this epistemic hell is some mixture of judgment and heuristics that we don’t even understand well enough to actually articulate, much less codify into an RL loop.
Hope you enjoy!
0:00:00 – Kepler was a high temperature LLM
0:11:44 – How would we know if there’s a new unifying concept within heaps of AI slop?
0:26:10 – The deductive overhang
0:30:31 – Selection bias in reported AI discoveries
0:46:43 – AI makes papers richer and broader, but not deeper
0:53:00 – If AI solves a problem, can humans get understanding out of it?
0:59:20 – We need a semi-formal language for the way that scientists actually talk to each other
1:09:48 – How Terry uses his time
1:17:05 – Human-AI hybrids will dominate math for a lot longer
Look up Dwarkesh Podcast on YouTube, Apple Podcasts, or Spotify.