Sometimes during Redis command design work I feel that CISC / RISC tension inside myself :D

I have had enough of Telegram services run by unknown individuals to do the audio transcription task in my Telegram channels. Wrote one using my botlib, ffmpeg and whispher.cpp. May be useful for you as well, perhaps? github.com/antirez/whis... Transcription quality is WAY better.

My blog at antirez.com used to run well in a very old Linode instance. Now I dismissed Linode/DigitalOcean in favour of @Hetzner_Online and the page generation is now 5 milliseconds (I mean, local HTTP request to completion). No web framework + Redis performs well.

What is happening in the United States is horrible. Half the Americans is in the right side, and it is the side that can restore and make the country sane again. Act now (without getting killed), do what you can to fix this mess. Get back your country.

I mean, flux2.c already needed to implement the Qwen3 pipeline to produce the text embedding. So, why not implement the full inference? Its' 2026.

I love so much the non sensical images generated by diffusion models when you embed the empty string as conditioning text. A combination of plausible things from their representations emerge randomly.

It is two weeks that I refine and extend the specification for the new Redis type (top secret for now, but not a niche type like vector sets, something really broadly applicable). I believe starting with Monday, it will be time to begin the coding phase.

After another round of optimizations flux2.c now even faster, but still far from DrawThingsApp that is literally a *beast*, kudos to liuliu (on X) and the rest of the team. Denoising a 1024x1024 image in 14 seconds (~27 for flux2 / PyTorch with MPS backend) is really impressive!

Not many people really understand what interpolation in AI is: it is the process in which folks that were completely LLM skeptics are interpolating their opinion every day to reach a "B" opinion that still has the same vibe but that factually fixes how wrong they were.

My bet is that AI policies will be soon avoided by most companies / projects in favor of quality centric policies.

I agree with Macron words but not with the sunglasses style (I understand it's for an eyes condition, but, go classic, dude). It is exactly the kind of situation where you want to look European and not the fucking Top Gun.

Flux2.c now supports the official img2img VAE-encoded inference, and it is like having a small Nano Banana in your computer, pretty impressive! Also the MPS target is almost as fast as PyTorch. github.com/antirez/flux...

Olé, if not for slower loading of weights, now flux2.c is as fast, using the MPS target, as the highly optimized PyTorch MPS backend. github.com/antirez/flux...

Codex seems semantically more powerful than Claude Code. Maybe Claude is a better overall agent, a lot more "respectful" of the files, less prone to do a mess. But GPT5.2 is more capable. I have no doubts... I wonder why Claude Code is *so much* more popular.

Spectacular PR to Dump1090 from my friend Demetrio: github.com/antirez/dump... We went out for a beer a few days ago and he returned back and did this :)

My split terminal right now: Left: Implement the MPS pipeline in bf16 without resorting to CPU for the Flux.2 inference. Right: Continue working at Redis design document for a NEW FUNDAMENTAL DATA STRUCTURE :D No, not something niche as Vector Sets this time. A bold new general thing.

Now it is faster especially on Apple Metal, and runs on 16GB Linux machines with BLAS target, if you use the --mmap switch.

Imagine writing a pure C, dependency free, fast enough Flux 2 Klein 4B inference library. In a weekend. Thanks to AI writing the code, and the human steering it towards the right direction. github.com/antirez/flux...

Imagine designing the safe tensors file format, and deciding for a JSON header, in a binary file that mostly contains blobs.

And... during the design stage, I decided not to continue with this idea. Back at the blackboard to find the next one. It is important to say "no" to yourself sometimes.

Working to some PR for Redis that is going to be non trivial and completely AI-developed under a strict specification. It's the same kind of spec I write for myself, but if this goes as I hope, instead of working 2 months, I'll have it ready much faster.

Claim: in the domain of programming, the biggest limitation of current LLMs is their ability to "see" images and reason about them spatially. People mistake decent layouts (in frontend code / apps) for this capability, but they are mostly vision impaired. Gemini is better at this

My latest video: the only SECRET!!111 you need to know to work well with coding agents. www.youtube.com/watch?v=TJ6r...

Recent changes in Dump1090: - Ground position messages decoding. - Merged improvements to phase correction from Malcolm fork. - Fixed UX in webserver mode. github.com/antirez/dump...

Do you use Redis Streams? What do you think about my latest PR to Redis? Thanks! github.com/redis/redis/...

Make sure to check lobste.rs comments on my blog post, if you are interested in the topic of AI coding. They are better than HN comments about the same post.

I started using /active because of this, where your post is currently #2 news.ycombinator.com/active

HN worst feature ever: if there are a lot of comments compared to votes, the post gets suppressed. Imagine a site where people (in theory very informed human beings) discuss stuff, that does everything to suppress controversial things.

New blog post: Don't fall into the anti-AI hype. antirez.com/news/158

I published on GitHub the GTE-small embedding model pure C inference library, written with Claude Code, here: github.com/antirez/gte-...

It is impossible to understand how the Italian government is ok with what it is happening in Venezuela. Maybe Meloni forgot that the white minority there is of strong Italian heritage, the fact Italians massively helped building Venezuela, and the relationships we had for decades.

Just to say that I wanted to run the GTE-small embedding model in pure C, without any dependency, and Claude Code wrote the 917 lines of code needed in 4/5 minutes. Perfect match with the huge codebase needed by Python to perform the inference. Same speed (!) as PyTorch.

In a single session with Claude Code (Opus 4.5) I implemented, in my "linenoise" library: UTF-8 at the right complexity point (I had two PRs in the past: one too simple, one too complex), and my dream: a testing framework simulating a VT100 terminal!

So thanks to Claude Code if my PRs will get into Redis, "make test" will pass again on slower (i5, in this case) computers. Incredibly boring tasks to do, for humans, to detect timing related failures in tests. But not for the AI, that unstoppably debugged the problems.

So thanks to Claude Code if my PRs will get into Redis, "make test" will pass again on slower (i5, in this case) computers. Incredibly boring tasks to do, for humans, to detect timing related failures in tests. But not for the AI, that unstoppably debugged the problems.

Recently I got a Dell Latitude 7420, refurbished. I installed Fedora with Plasma. And after a few days of using it: 1. I severely overestimante the MacBook experience. This hardware is a lot less capable, but the keyboard is better, the plastic feel and the small weight are awesome.

First LLM coded Redis patch? github.com/redis/redis/...

So right now for MANY new projects I would: 1. Create the first core alone or with LLM web interface assistance if needed. Shape *every* line. 2. Start building more with agents, using specification.md files, to put the agent in the right design direction. 3. Coffee as it codes. 4. Review. GOTO 2

I believe we are no more into the stage where the LLM + human in the loop is the best approach for many projects. Sure, reviews are needed, but at this point what coding agents can do is simply incredible. Last few weeks were as shocking as the first ChatGPT version.

StackOverflow chart of death interpretation: LLMs killed it so fast because it was already losing users because of the social issues generating inside the site. Otherwise, SO could be the place where certain quality conversations continued to happen, and also a great source of LLM training.

Wars for oil, again. Since I'm a small child I see the United States bombing other countries to control oil. This also shows that Trump is not just generally a disaster for Americans but *at least* the foreign policy is less aggressive. It's 360 degrees bad.

At this point the average level of discussion within Hacker News threads when the topic is AI is: a bunch of normies drinking beer in some café and cursing the government, their football team coach, and their ailments.

Needed context

Reflections on AI at the end of 2025 antirez.com/news/157

In this video of mine, you can hear the new YouTube AI dubbing if you select the English audio track. It is *incredibly* good, and a huge step forward compared to the past version. www.youtube.com/watch?v=cvWb...

🇪🇺❤️ 🇮🇹🇦🇹🇧🇪🇧🇬🇭🇷🇨🇾🇨🇿🇩🇰🇪🇪🇫🇮🇫🇷🇩🇪🇬🇷🇭🇺🇮🇪🇱🇻🇱🇹🇱🇺🇲🇹🇳🇱🇵🇱🇵🇹🇷🇴🇸🇰🇸🇮🇪🇸🇸🇪 = 🕊️🏛️📜

There is a new port of zx2040 to the PicoCalc! github.com/ohm69/zx-pic... @flohofwoe.bsky.social

Running the Redis test several times today for a fix I'm operating in Redis Streams: it's one of the few times my computer spins the fan, other than when I spin llama.cpp or when I use Davinci Resolve. I'm 100% positive that, I would start it again, the Redis test would be C coded

Random idea for when you need a totally deterministic LLM and can't control the kernels: when there are N top logits (N>=2) separated by less then a given epsilon (chosen so that reordering may produce such delta), among all them, take the token that is lexicographically greater.

One morning on tests in complex C code: GPT 5.1 > Gemini 3.0 Yes, it's just a specific use case: code reviews. And a single person testing. But to say, it is not obvious how is leading right now. If GPT 5.1 can correctly reason on things Gemini 3 PRO can't I could say the landscape is complex.