See the authors twitter, he speaks english at a rather basic level and certainly did not write this https://x.com/mohitgeryani/with_replies
In my experience, AI fixes problems by mostly adding more code.
It's a short term gain for a long term hurt.
In my experience, humans unfortunately tend to do the same.
I’m doubtful this is a meaningful result. Kernel contains a lot of legacy code and generalizations to support different hardware etc.; removing that would result in a speedup. Next are all the mitigations for hardware vulnerabilities and attacks. If removed would give a nice speedup as well at the cost of security. And then finally, just specializing the Kernel in whatever the benchmark is measuring, making it useless as a general piece of software would also make it fast.
there are more elegant ways to leverage an LLM, see AlphaEvolve: https://arxiv.org/abs/2506.13131
it's difficult to frame most coding tasks in such a way where you can trivially verify correctness.
> For context: GLM 5.1 ran the same task and reached 7.3x. Kimi K2.6 reached 5x. DeepSeek V4 Pro reached 3.3x. The models that stopped early did so because they issued no tool calls for five consecutive rounds, they concluded they couldn’t make further progress and stopped. Qwen3.7-Max didn’t stop.
By this reasoning I could release a model that lacks all the basic optimisations. Have it optimise itself for hours to reach 20x the throughput and then claim that the model is superior to the others?
I am not saying that is what happened here, but the reporting is abysmal.
Right now now I switched to the latest codewhale agent (in Rust), and it would perform much better according to his qualifications. Much better async IO implementation and orchestration, no more deadlocks as in the typical typescript tooling. It just doesnt stop out the blue, as claude, kimi or opencode.
> Over 35 hours it performed 432 kernel evaluations. Each cycle meant writing code, compiling it, running it, reading the profiling output, deciding what to change, and trying again. The model diagnosed compilation failures it hadn’t seen before, identified performance bottlenecks through runtime feedback rather than prior knowledge, and redesigned the kernel architecture multiple times when incremental improvements stopped working.
Anyone remember genetic algorithms? This might be an improvement, but it still feels a little like deja vu.
Technically birdshot from a shotgun is also randomly distributed (passing through a cone). This actually improves the chance of hitting the clay pigeon, because the birdshot spreads out and each individual ball has a chance to hit.
Genetic algo is similar. it's an optimizer that - in order to avoid local optima - will 'shotgun' an area around its current best guess.
So instead of just sampling in a close 'circle' around your current point looking for a 'down', how about we spread that out a bit? You could use a 'circle' in a regular pattern, but what does that even look like in high dimensional space? Seems it's best to use some random distribution centered on your current position.
(LLMs actually have a 'temperature' setting which introduces noise for this exact reason.)
Some of GA's claims to fame are A) it uses purely just this distribution to descend. B) It can find multiple optima.
The way I think of it is that the simplest GA is basically greedy optimization with spread.
Greedy is like shooting a rifle , which is great for sniping, but you'll miss if the target is moving fast or doing things you can't quite keep up with.
A GA -like a shotgun- introduces spread: multiple chances to hit, multiple chances to escape local optima and rough patches in the landscape.
(A really good -if slightly morbid- modern example in the wild is COVID; which managed to outwit human civilization rather handily. "Not bad for a bit of encapsulated RNA" you'd think - until you realize it was running trillions of attempts in parallel. Really, the poor governments had no chance. )