Not affiliated with *that* tetsuo. Compilers, infosec, systems programming. Building compilers for ML hardware. Spiritually widgetmaxxing.
Melbourne, Australia
Joined July 2025
- Tweets2.2K
- Following2K
- Followers2.5K
- Likes33.2K
Pinned Tweet
Oh, you're writing CUDA kernels? Everyone's on Triton now. Just kidding, we're all on Mojo. We're using cuTile. We're using ROCm. We have an in-house DSL compiler targeting the NVGPU MLIR dialect but wait, Tile IR just dropped so we're going to target that instead. Our PM is on TileLang. The team lead was on CuTe but now she's back to handwriting PTX. If you're not on Pallas, you're ngmi. Our intern is building on TT-Metalium for our Wormholes. Our CFO approved an order for some big chungus wafer-scale chips so now we're porting our kernels to CSL. Our CTO is working on a kernel-less graph compiler so we won't need to write kernels anymore. Our CEO thinks we're talking about the Linux kernel. We're building Claude for dogs.
I’ll have to give this a try myself. In the past, I’ve been skeptical of some of these ultra-aggressive agentic approaches to development because the resulting software is always poor quality (I don’t care about merging millions of LoC if it’s all crap). But by all accounts, Grok Bot is a well designed piece of software. The author of pstack also has some serious technical chops which makes me feel that this is worth understanding.
Has anyone adopted pstack (or similar) for developing low-level systems software? If so, how do you find it?
tetsuo.mlir retweeted
Any startup interested in HPC engineer who worked on 2 open architectures and has experience debugging model behavior? Almost 3 but I turned down the job for [redacted] reason. Also llama.cpp maintainer / infra runner on Gemini protocol.
I need to find myself a new job.
Word on the street is that the Claude desktop app isn’t complete slop anymore. Have they changed their ways?
Schrödinger's proof: if a frontier lab generates 700 slop papers and no one reads them, has mathematics advanced.
tetsuo.mlir retweeted
Looking for a killer FDE in Sydney to join Amp. Hit me up!
We’re also open to hiring a pair.
tetsuo.mlir retweeted
Mathematicians, having traditionally lived lives of excess surrounded by women and driving fast cars, have been brought back down to earth to join the unwashed masses.
I don't want to hear about how AI could cure cancer (uninspiring, trite, boring tbh), I want to hear about how AI is going to allow me to grow a comically large penis. If you want to garner public support, frontier labs need to create a vision that people want to believe in.
Next time I break prod, I'm going to tell the team that it wasn't me, it was Codex.
Is there an established explanation for why these newer models started occasionally invoking Python to make edits instead of using the Edit tool? I'm trying to understand whether this is beneficial in some way.
If someone uses one of these bots to text me, I will load them into a canon and fire them directly into the sun.
What do you use Grok Bot, Muse, Dots, etc for? They seem similar to Hermes and OpenClaw.
I tried Hermes a while ago but couldn’t figure out what I’d actually want to do with it. Most of the things that I’d want automated in my life involve credentials that I wouldn’t feel comfortable handing to an agent, but I may just be lacking imagination.
What do you use Grok Bot, Muse, Dots, etc for? They seem similar to Hermes and OpenClaw.
I tried Hermes a while ago but couldn’t figure out what I’d actually want to do with it. Most of the things that I’d want automated in my life involve credentials that I wouldn’t feel comfortable handing to an agent, but I may just be lacking imagination.
When I was starting out, big tech used to interview almost anyone. It used to be such that, with a few months of being a disciplined Leetcode monkey, you could completely turn your life around regardless of your experience, school, connections, etc. I hate these questions as much as anyone, but there was something beautiful about that.
I'm pleasantly surprised that Opus 5.5 bucks this trend and is actually decent to pair with. The biggest improvement is that it doesn't communicate like an idiot.
My theory is that most of the focus is going towards making these models good at working autonomously and kicking humans out of the loop. I think that's why they seem to be getting worse (or at least not improving) at communicating concisely and being sensitive to second order effects of solving problems that humans generally care about like readability, managing system complexity, etc. If you look at the recent marketing material for SOTA models, a lot of it is about working autonomously for X hours, building something without human intervention like the Kimi K3 demos, etc. In short, I think the things that you and I look for in a model aren't prioritised as much as they used to be.
Afaik, most of the industry is still struggling to measure productivity gains from any of this and I suppose a lot of that is to do with the fact that they're bottlenecked by humans. So in order for the labs to justify their position at the centre of everything, they need to make good on their promise of creating "AI workers" that can do work end-to-end and replace some subset of jobs rather than helpful assistants that augment existing workers.