Pinned Tweet
I have been verbally sharing my perspective & insights on the AI market having worked in the space for 6 years now. Yet another person told me I should be writing more often. Stay tuned for more!
open.substack.com/pub/deepdr…
Jared Roesch retweeted
AI for principled compiler construction!
Everyone is talking about AI-generated high-performance kernels, but the real opportunity lies in using AI effectively to build compilers! Why use AI to generate kernels one by one when it can build a kernel generator for infinite reuse?
By pairing solid compiler engineering principles with tools like Claude, we built a CUDA Tile IR backend for PolyBlocks in a few weeks, achieving a level of performance that typically takes multiple human-years!
Our backend now compiles unmodified PyTorch and JAX models to a sequence of fused CUDA tile kernels (all via MLIR) with state-of-the-art performance. By layering on top of the NVIDIA CUDA Tile IR compiler, we offload lower-level optimizations to it. Both NVIDIA GPU backends of PolyBlocks, cuda-tile-based and non-cuda-tile-based (which we had developed over several years), still rest on the same Polyblocks code generation infrastructure for affine fusion, tiling, mapping to MMAs, 2-d tilization, etc., but the cost models they use on both paths are slightly different.
Check out the benchmarks below showing wall-clock and total kernel execution times.
More to come! Compilers are now finally evolving rapidly!
ALT Polyblocks performance with CUDA Tile backend - wall clock times.
ALT Polyblocks performance with CUDA tile backend - kernel times.
I have had the pleasure of hanging out with Shalev and Romi this year as they have begun to organize their thoughts and ideas for the space into a vision for AI security.
I am confident we will see some awesome things come out of their startup journey, take a read!
Jared Roesch retweeted
We are announcing our cyberdefense strategy for superintelligence.
We are faced with three new threats: sabotage, escape, and theft of superintelligence.
>Sabotage: a model we depend on has been secretly turned against us.
>Escape: a model we created has broken free of our control.
>Theft: the capability we built is now in the hands of an adversary.
We must work together to secure the stack against this new set of threats. Here’s the strategy.
Secure Acceleration: A Cyberdefense Strategy for Superintelligence
This is such a good take. I learned how to first manage a group of people to accomplish something as a 13/14 year old raid leader/tank back in 2006 doing 20/40 man content with folks.
A kid who knows how to raid in WoW is unironically a better hire than 99% of non gamers. They develop team skills, coordination, leadership, how to master the respect of 40 people, how to share loot, maintaining a raid schedule, showing up on time, clear direct communication, how to get along with unique personalities, creating and executing on strategies, failing and trying again (perseverance), delayed gratification, etc. Invaluable skills are developed during raiding that are transferrable to the real world.
Jared Roesch retweeted
We just have a port of cuTile for Java as well. github.com/beehive-lab/Torna…
![[cuTile] Java CUDA Tile (cuTile) execution path to the CUDA backend as TileContext API by mikepap...](https://pbs.twimg.com/card_img/2114713434541694976/PBUIL3PE?format=jpg&name=800x419)
Jared Roesch retweeted
Quoting Ilya as a response to the field medalists.
If all math/science/tech problems are solvable by a super AI, no one will even care about explicitly solving a problem unless there is a practical need for that solution. The self-aggrandizing importance to deriving insights for the advancement of the community has no importance in that future other than satisfying our own ego.
Jared Roesch retweeted
It's very difficult not to read @terrence_tao's declaration about AI in math as sour grapes. He makes as eloquent a decel case as anyone could for the downside of using AI to solve hard problems. But ultimately, it still unavoidably boils down to ego.
The power of AI is profoundly humbling. And as the old saying goes, the bigger they are the harder they fall.
Sure, one could argue as Tao does that having unsolved problems out in the wild serves as additional inspiration for exploring and building an understanding of the wider intellectual landscape around that problem.
But only a tiny handful of individuals have ever had the privilege being pioneering explorers. The glory of first discovery is marvelous, of course, but it's only ever even been a possibility for a rarified few.
For all the rest of us, we simply have to settle for being the latest visitors to an already charted and mapped landscape. I'm sure it will sting a bit, having to join the club down here in the intellectual dirt with us plebs, but don't worry - you'll get used to it!
I suspect mathematics, like chess, will survive the dethroning of humans as king of the castle. More people visit and explore the chess landscape today than ever before. And while getting whupped by IBM's Deep Blue was initially hard on Garry Kasparov's ego, he eventually got over it, and a new generation of even greater players grew up in the shadow of Stockfish.
terrytao.wordpress.com/2026/…
Jared Roesch retweeted
Sorry, but asking Jacob about AI extinction risk is like asking your AC guy about climate change.
Not saying it's necessarily uninteresting or wrong per se but let’s keep things in perspective and hear from the full range of expertise across the ecosystem!
The funniest part of this week is watching mathematicians having the experience of being computemogged for the first time.
Subtracting the drama from the situation this is a moment where a challenge of intellect and taste was transformed into a pure function of compute.
Jared Roesch retweeted
If you publicly assign a >10% probability to human extinction within a decade, you owe people a clear explanation of how you reached that number + what would change your mind.
I’m disappointed in all the fear-mongering.
This tech wave needs adult chaperones.
Excited that @nvidia is joining the @rust_foundation , more to come in the next few months, wrote a few notes on LinkedIn lnkd.in/p/gvKf7yni about me and @binarybana joining the board.
Looking forward to more Rust + NVIDIA!
Jared Roesch retweeted
We're pulling back the curtains on Rubin, designed to maximize agentic tokens/watt! Some headline numbers:
- 50 Petaflops sparse NVFP4 compute
- 288 GB HBM4 with 22TB/s of bandwidth
- NVLink 6 with 3.6TB/s of bandwidth
But that's not all! Cool new performance features include:
Jared Roesch retweeted
Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉
What’s crazy is: we 100% leveraged AI agents to beat the competition.
This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.
This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .
Databricks AI is like a neolab. Join us if you’re cracked!
Fearless Concurrency on the GPU
For those interested @melibol just posted a paper on building a safe Rust kernel programming abstraction on top of Tile IR.
arxiv.org/abs/2606.15991
A short teaser: but the safety is effectively free. On a B200, the safe GEMM is competitive with cuBLAS: about 2 PFlop/s 92% of the GPU's dense f16 roofline.
Read more in the paper or Melih's LinkedIn post (linkedin.com/posts/elibol_gi…)
He will also be giving a talk at RustConf in September, hopefully he will see you there!
I am giving a talk at ARRAY @ PLDI this week. If anyone is around Boulder for PLDI would love to catch up! see you all there!
Caught at @MLSysConf
The era of Slop is over.
After months of optimization, I’ve eliminated every unnecessary step between “writing code” and “throwing it away.” – at Bellevue, WA
