I'm thinking that because of automatic programming certain complicated and elegant programs we wrote in the past may basically become a form of art for future generations. Like the manuscript books of the middle age for us.

Biggest mistake in AI coding era: to believe that specifications should be either natural language OR something else. The best combo is a natural language high level specification (the intend), plus code (as it gets written) documenting the finer behaviors.

Day 2 of Monster Scale Summit! Today, we have @skamille.themanagerswrath.com, @antirez.bsky.social, @tigerbeetle.com (Joran), @dominiktornow.bsky.social, @teivah.dev, Avi Kivity, and lots more. Guaranteed 🔥🔥🔥 www.scylladb.com/monster-scal...

New blog post: "GNU and the AI reimplementations" antirez.com/news/162

Who thinks "clean room" is needed to reimplement and put it into a new license does NOT understand copyright. Clean room is a trick to make litigation simpler, it is not mandated by law: rewrites are allowed. The new code just must not copy protected expressions. Linus was Unix-aware.

Every morning there are two news: the news about the release of some major upgrade to some AI, and the news about the lack of DeepSeek v4 release.

Implementing a clear room Z80 / ZX Spectrum emulator with Claude Code: antirez.com/news/160

Users report that asking Sonnet 4.6, via the Anthropic API, "What's your name", it reports "DeepSeek" with a high frequency. Labs cross-fine-tune on chain of thoughts and use other models as RL signal consistently. Also the pretraining, a *key* step, is on mostly public data. Anthropic disappoints.

A clear room emulator of the Z80, the Spectrum, and CP/M written by Claude Code. Vey skeptical with the compiler experiment (more complex for sure) made by Anthropic because of the fundamental flaw of not providing the agent with the specifications / papers.

You see, I'm ok with AI usage in many ways. Yet I can't understand why people that don't have deficiencies in written expression use it for writing. Emails. Blog posts. Comments. Why? We only lose something in this way. We want your voice.

Today I had to fight with GPT 5.3 to defend my position on the complexity of a specific command of the new Redis type I'm adding (released soon I hope). It had a great point about the worst case, but the typical case was as I claimed. We reached an agreement mentioned both... :D

You know what happens with Nvidia ban for the Chinese market? That 1.5 technologically advanced and capable humans said "maybe we can use our own GPUs". You know what's going to happen with Anthropic-style AI usage bans, right?

After 19 years, the version 2 of the Picol interpreter is out. Features [expr] in ~40 lines of code, floats, globals, can run mandelbrot.tcl, and so forth. The code is now more functional and more readable at the same time. 654 lines of code now. github.com/antirez/picol

On the stress induced by automatic programming (English audio and subs available): youtu.be/id9QG-mQSOo?...

Of all human feelings, envy is the one I despise the most.

Claim: OpenAI is leaving a lot of money on the table (and a lot of Codex users) for not having a plan between 20$ and 200$

Flux2.c is now Iris.c (greek goddess Iris, messenger of the gods and personification of the rainbow), and adds support for the zImage Turbo model as well. Calling projects after company products names is a bad long term idea. github.com/antirez/iris.c

Imagine creating the UX for HBO Max and deciding to make the screen almost black each time somebody moves the cursor or acts with the TV remote, or each time the show starts, for 10 seconds. How broken products design is, in 2026?

WTF the Qwen3-TTS encoder/decoder compress wav files of 100 times... Compression is now GPU-bound, no longer algo-bound, for the most part. The same is happening for images and videos as well, just not practical because of speed for now.

The 20$ codex plan is worth more than the $200 Claude Code plan.

The way those transcribing models work, with audio -> FFT -> MEL -> Conv2D -> self attention feed to what is, basically, a decoder only LLM (autoregressive with the emitted tokens AND the audio embeddings generated by the encoder) is one of the MOST fascinating things in AI, IMHO.

Qwen3-ARS transcription models inference in C is ready: github.com/antirez/qwen... This is by far the most practical model for quick inference on low end CPUs. Perfect for things like transcribing bots, voice-driven UIs, and so forth.

Now we are talking. 2x audio length speed for the 1.7B Qwen3-ASR large (1.7b) transcription, on my very modest Hetzner 13th Gen i5 server. The pause you see initially is the encoding stage. -S 0 means: no segmentation, one offline pass.

When you target MPS for local inference, and you got an M3 MAX, you are biased into a high performance world that does not exist for others, and the BLAS / AVX / NEON targets will suffer. So for Qwen-ASR I'm NOT going to include an MPS engine on purpose.

Software is created for accumulation of knowledge. AI is not going to cancel this fact. Forgot the idea that programs will be prompts (specifications). The details is what really matters, and they are hard to capture textually than in the code. New projects will be spec + code, evolving.

You see Anthropic releasing a "fast mode". At the same time you see Codex slowly doing in 5 minutes something that Claude Code rushed for 2 hours trying to get done. Speed is very relative.

Now Flux2.c also supports the base model (if you have time and a >= M3, and the inclination), and different schedulers, and it is quite fan to run, also at reduced steps, since 50 are... a bit too much to wait.

Now flux2.c is consistently faster than official PyTorch MPS and not far from Draw Things (14s vs 19s at 1024x). DT uses 6bit quants while flux2 uses BF16 (2.7x the weights). Happy with the result so far. github.com/antirez/flux...

voxtral.c decoding speed is now near to Metal hardware limits for ~80%.

The HN commenters who simultaneously argue "it just decompressed GCC" and "the output is way worse than GCC" don't seem to notice they're refuting themselves. I didn't expect better than that, given the normie-alization of the site in recent times.

I wrote a SPEED markdown file with instructions for Claude Code (Opus 4.6) with a loop process in order to improve the speed of the Voxtral speed with the Metal backend and left my home to pickup my daughter, have lunch, ... Back home, the code is 2x faster, and it is still going.

No mercy for abstractions that suck. Don't let your LLM be tempted.

Voxtral (from @MistralAI) transcription quality is quite incredible, the way it handles the punctation and all the rest, makes transcribed audio messages so much more understandable. I implemented a few fixes in the FFT and now there is no longer a skipped tokens issue in voxtral.c

Maybe you were wondering why serious, grown-up folks would spend nights writing a Commodore 64 well made demo or game, in recent years. Now that your dear LLM can write a serious C compiler in Rust, but can't write a well made C64 game/demo, maybe you are starting to get it.

Yesterday @MistralAI released an open weights transcription model able to work in real time, Voxtral Mini 4B. Today, following the Whisper.cpp lesson, here is a C inference pipeline ready to use as a library, I hope you'll enjoy it: github.com/antirez/voxt...

Mmmm... 44ms per token. Not stellar but nice.

Anthropic ADs post: they realized that with Claude Code they have a solid business model and that ADs-model is not comparable. They made a business decision, and used what they refused (ADs) to improve company image. As simple as that.

I'm pretty confident that Claude Code, since it spawns multiple sub-agents for sub tasks, has lower performances.

Italian is an intrinsically understandable language, by people and machines.

Explain me one thing: how a sane human being that saw Claude Code / Codex / ... at work for a few months can have *any* doubt about the ability of those systems to work as a personal assistant, given enough access to your stuff? It requires a lot less capabilities than the other main task.

With this prompt trick, you can use flux.2 4b klein (distilled) as a handy and powerful super resolution model. In the example 128x128 -> 1024x1024.

Welcome to TgTerm: run your coding agents via Telegram for fun and profit: github.com/antirez/tgterm

Remembering to run stuff with tmux, tunnels, termux or another SSH client on the phone is ways too much. So I thought I could build a Telegram bot that gives me access to terminals running on my computer when I'm just with my phone around. Security must be handled. I believe I'm going to use it.

Anthropic may be near to release a new model, however, as things are today, after weeks of testing, I'm totally sure Codex with GPT 5.2-xhigh is definitely a model capable of handling way more complex tasks.

Un video sulla vita e le opere di Norbert #Wiener, grande scienziato, inventore della #cibernetica e matematico sopraffino, raccontate sfogliando qualche sua opera. youtu.be/dSPd7excxC4?...

The new AGENT markdown file in the Flux2.c project -> github.com/antirez/flux...

As I did time ago with the Claude account, let's stop also the ChatGPT subscription I did for an error via iPhone to redo it via the web site. I don't think the 30% cut is fair and, as a customer, I want to express it with my choices.

Second time my MacOS crashes because of a combination of Ghostty memory leaks + MacOs kernel terrible handling of OOM conditions.

"What's your workflow with coding agents?" Well, this, among the many others.