Gianluca Guida

I just published kotama, an experimental GOOS=tamago compiler branch that allows execution of a rich example in barely 6MB of RAM, running on an rv64imfc CPU (yes, no d or a extension). This paves the way for supporting @AIFoundryorg Minion cores! github.com/usbarmory/kota…

I hacked TamaGo exception handler to report invalid register Read/Writes without crashing the VM. I used this to extend bare metal Go LLM-fuzzing all x86_64 I/O controllers. 60 seconds. "Let me try writing to redirection entries with malicious values". Hypervisor coredumps.

Hello again San Francisco. Thanks for your usual “Why am I not living here and wtf am I doing with my life” moment. I missed it.

A good use of vibe-coding is flushing all the random projects ideas I had in my "would be fun to" cathegory. This is a project that occupied many afternoons in 2015: an assembler for intuitionistic logic and Martin Lof Type Theory, that maps to the machine. I'll probably never use it, but always wanted to see how it would look like. Here's a matmul, working now, correctly generating LLVM IR: matmul4_rowvec : Π(A_rows:β(Array(Vᵤ₃₂,₄, 4))) → Π(B_cols:β(Array(Vᵤ₃₂,₄, 4))) → Vᵤ₃₂,₄; matmul4_rowvec ≔ λA_rows. λB_cols. ⊕ᵥᵤ₃₂,₄ (⊕ᵥᵤ₃₂,₄ (⊗ᵥᵤ₃₂,₄ (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ A_rows 0)) (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ B_cols 0))) (⊗ᵥᵤ₃₂,₄ (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ A_rows 1)) (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ B_cols 1)))) (⊕ᵥᵤ₃₂,₄ (⊗ᵥᵤ₃₂,₄ (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ A_rows 2)) (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ B_cols 2))) (⊗ᵥᵤ₃₂,₄ (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ A_rows 3)) (⤓ᵥᵤ₃₂,₄ (↦ᵥᵤ₃₂,₄ B_cols 3))));

@marty188586 @corsix Yes this is my suggestion, use fans that sadly occupies a pci slot in the case.

@corsix Trying to make a fan attached to the thing. Zip ties aren't working super well...

Hello new friend

Just asked Grok to create claude instructions for a project we were refining. Didn't take it well. :-)

I am starting to believe that whoever said that if you're tired of London you're tired of life had a very different definition of life than I have.

This was a really fun one, but next year I’ll plan my usual two FOSDEM talks to be on different days :-)

@bjorntopel really great RISC-V room this year!

FOSDEM again. Sleepless night to prepare the slides, promise that won’t happen next year, and too many talks to follow. Finally!

@marty188586 @FelixCLC_ @AIFoundryorg Totally get it. Let’s fix that.

@tlbflush @FelixCLC_ @AIFoundryorg Otherwise I'm arguing in an vacuum. That's not helpful nor productive

The ET-SoC1's SIMD is quite decent. I like it. Now the problem is I need to learn the aliasing rules

@marty188586 @FelixCLC_ @AIFoundryorg Yes, that’s exactly what FOSDEM (and the AI plumbers) is about! That, and @rhatr suggesting which Belgian beer to try.

@tlbflush @FelixCLC_ @AIFoundryorg Yes I realized timezones is a real blocker (and I want to only comment on things I have real understanding on). Hope we'll have time discussing at FOSDEM

Then you should join our discussion next time! We spent a good portion of a Friday night discussing exactly this very specific topic— is it the same, is it different? — with @FelixCLC_ Now the funny thing is, since this is @AIFoundryorg we could totally have this discussion on the open. Is just that sometimes, when ideas thrown are insane, you might need — initially — a reduced friendly public :-)

@FelixCLC_ Yes. That thing is broken. But I consider the tensor side a different problem

Looking forward to reading it. I have been experimenting with this topic quite a lot recently, and I still don't have a good answer on what is the best way. I converged to starting with: "Analyze this existing project, find out coding style and architectural principles, and write it to a doc". This at least fixed the coding and commenting style. And everytime I didn't ask the LLM to think of it architecturally first, it would just start destroying all the layers, which on a project like the one I was working on -- a Mach microkernel clone fairly complete -- was just insane, and most of the time wrong. Also, small commits. Extremely focused. And asking to review it every time. But this is normal claude behaviour. In the end I ended up giving up on Claude defining architecture of the work to do, but once it was given data structures and pointers to the code, it was great. This feels at a strange inefficiency point, so I look forward to see what you have to say about this topic.

I have some rough notes for a blog post titled "Building System Software with AI". If you're someone already doing it, what are things you think I should include in the post? If you haven't used AI for systems programming or are sceptical, what would you like to know?

Claude Max 20x after first weekend: reached limit. In a weekend I finished, through a lot of steering, a project I wanted to do for years but didn't want to allocate weeks of my life to it. Feels revolutionary to our industry. It is evident that writing software has changed forever, and in a couple of years sitting at your desk writing syntax by hand will be the equivalent of listening to vynils, or playing videogames on an Atari 2600 -- you'll do it for the warmth and nostalgia. For a company to use this makes incredible sense. But. I feel I am now paying to do the thing that has been a constant free source of joy in my life: programming. And when the limit was reched, I felt I had to just pay more to continue. Ended up downgrading, as a knee-jerk reaction. Offloading my programming to a third party is something I am not ready to do yet. I will explore more cumbersome ways to have local LLMs help me programming. It's going to be harder, worse, but independent. Hopefully similar to what it was running Linux in early 1990s.

When I said that this is getting more and more fun every day, this is on of the examples of what I mean.

And with the full Go ecosystem at this level this also becomes an unbeatable AMD SEV-SNP auditing/fuzzing environment.

Okay, this thing is officially better than expected. I might have time to work on Machina again.

Okay,I can officially justify my Claude max subscription. This is an example I asked Claude to do for a actor-model with capabilities I wrote, powered by @penberg agentfs!

My holiday project has been a return to Rust, which -- end of year controversy -- I missed very much, at least in userspace. The whole RAG was always a gray area on how it worked, so I decided to focus on the R part. github.com/glguida/capsa (and yes, claude rickrolled me). Don't take it too seriously, but it works.