tetsuo.cpp (no slop)

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.

@mitsuhiko @badlogicgames Pi Enterprise Edition.

@badlogicgames 👀

i am being bribed. i'm unbribeable.

@soham_btw I see, how do you find the Vim bindings?

@tetsuo_cpp terminal, codex/claude code with neovim if i gotta edit

giving zed a serious try

@isaacvando Thanks for writing this, this has been bugging me lol.

Every time someone posts a take like this, the replies are full of naysayers explaining how this won't work because compilers are deterministic and LLMs aren't. But this has nothing to do with why you shouldn't treat an LLM like a compiler! In defense of poor old nondeterminism:

@glcst @pavan4820 @PThorpe92 Amazing news. :-)

Everywhere I go, people keep commenting on how wild it is that I have something working with us straight from *prison*. Well, no longer. Tomorrow, at 8 am in the morning, @PThorpe92 is a free man. Preston is an inspiration to us all. He achieved so much from behind bars. I am sure he will go even higher as a free man.

@blip_tm Any opinion on Positron? :-)

AI chip companies I’m bullish on: MatX, d-matrix, Fractile, and SambaNova each has a great engineering team, and either a technical moat like processing in memory, or an architecture that can actually leverage HBM for both training and cost-efficient inference

This is becoming a real problem.

Japanese HPC engineers, who collaborated for the development of the world's Top 1 Supercomputers from Japan, including “Earth Simulator” in 2002, “K” in 2011, gathered to celebrate Dr. Tadashi Watanabe, ex-NEC & RIKEN HPC Leader and the winner of Okawa Award in March 2026. #HPC

To me, it's less about determinism and more that you don't prompt an LLM in a formal language as you would a compiler. Even a very detailed prompt or implementation plan hand-waves a lot of detail that the agent has to fill in which is why you have to review the code for anything important. I see some folks talking about formal verification for these sorts of things which I think is an interesting avenue, but from what I've seen, most of the people that talk about not reviewing generated code just ship a bunch of unit tests (which they also haven't reviewed lol).

@_winter_wonders i mean, in a narrow sense they kind of can be, by fixing the temperature and random seed (though not bulletproof apparently); if the pipelines are reliable and predictable *enough*, then people will accept the trust-utility tradeoff, esp if it's empirically better than a human

Idk if I'm missing something but I'm seeing a lot of smart security people talking abt just having AI code just never be reviewed at all as a desirable thing? Am I missing someth?

@CasualVince That’s what the dedicated crash out account is for. 😂

Being a facedoxxed account is a prophylactic against vague posting and (hopefully) crash out posts

@jindra_dev @joelreymont Hey Jindra, I'm aware of this as there was a bit of a Twitter pile onto Joel around that time. I believe this was just a mistake which he and others have since learnt from. He is a cool and thoughtful person, and is worth a follow. :-)

@tetsuo_cpp @joelreymont I wonder as this was used as Exhibit A for why Zig ran off from Github and vibe coders posting PR whatever their external brain tells them. ziglang.org/news/migrating… The PR was also refused, closed and disgraced. github.com/ziglang/zig/pu…

My first Zig PR github.com/ziglang/zig/pu…

@soroush_srd @twoscp @mattgodbolt @benrady Haha yes, as soon as I see a new episode, I listen to it immediately.

@tetsuo_cpp @twoscp @mattgodbolt @benrady I second that! I've listened to most of their episodes, some of them more than once😂 Tbh i'll listen to anything Matt puts on the internet!

If you like podcasts, @twoscp is a great listen. A lot of popular podcasts are an interview format and the focus is on their cast of interesting guests (don’t get me wrong, these are great too). But this podcast is just two skilled practitioners (@mattgodbolt and @benrady) talking shop and discussing the day to day of working as a programmer. It’s all very fun and relatable and I always walk away with something interesting to think about. :-)

@fjzeit Ffs.

i really never thought the industry would reach such levels of insanity... skiplabs.io/blog/codegen_a…

@opdroid1234 @lucasmeijer Not unlike an LLM. 😂

@tetsuo_cpp @lucasmeijer Its doing a very sophisticated implementation in that case - the inner core can do robust matmuls and yet the outer shell cant matmul to save his life

Everybody who thinks ai is conscious has to do a mandatory from scratch transformer implementation. There are only floats and multiplications.

If GitHub was built in the early 2000s.

@rogesterone Lately I’ve been super impressed with how well executed the remote editing functionality is.

@rogesterone I only use it for coding so I’m not sure about that use case. As for modal editing, I’ve been told the Vim mode is quite good. I use the Emacs bindings and find them ok as far as basic editing goes.

Zed is beautiful. In the current age of AI slop and MS Teams Copilot 360 Express Suite, they’ve managed to build a company around software crafted with love and care. :-)