DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

Markdown vs HTML. Every time we go from a semantically dense to a semantically sparse format, we lose. Even more today where less tokens from the same content is way better. I can understand we need a better markdown. I can't understand we should replace it with HTML.

If have access to a recent Apple Silicon system with at least 128GB RAM, I can highly recommend to try this DeepSeek Flash 4 model runner out. The description in the README is not lie: It absolutely feels like a frontier model and produces significant better results than e.g. Qwen 3.6 or Gemma 4

In case you have doubts about the q2 quants inference of DS4 (I noticed many don't trust the README claims), here is it analyzing Picol source code using the pi agent.

Welcome to DS4, a specialized inference engine for DeepSeek v4 Flash. github.com/antirez/ds4 This project would have been impossible without the existence of llama.cpp and GGML and the work of Georgi Gerganov and all the other contributors. Thanks!

Two days ago one of the best programmers I know wrote me how she is getting interested in AI coding. I told her how I was avoiding writing about my journey here on BlueSky, and preferred X instead. Well, this quoted tweet is the kind of polarization which is happening here, and why BlueSky is dying.

What happens when I use autocoding is *very* strange, is something that didn't exist in myself, in the past. I write big chunks of text to the LLM and every word I type I "feel" how it is going to condition the work of the agent in one way or the other. Every word reduces the error.

How different backend numerical differences flip nearby tokens.

What I'm doing with agents (when they struggle, I say "write an email to an expert into /tmp/letter.md", and pass it to GPT 5.5 PRO extended) will soon be a (costly) tool call of new versions. Maybe somebody at OpenAI will read it this prophecy will come true.

One interesting thing about the name "X" is that the resistance to accept it was given not just by the way it was imposed, but by the fact it is a genuinely stupid name.

New blog post, Redis array type, short story of a long development process. antirez.com/news/164

I just submitted a PR for a new Redis Array type :) github.com/redis/redis/...

AI coding is great for many things but wow if it sucks hard at finding good function names. It read too much code around I guess ;D

Europe AI strategy should be to specialize on AI inference and improvement of large open weight models, while we try to recover the GPU / companies gap to have a viable internal path. A large Chinese open weight model that works is only better than an European-trained weak one.

Today local inference of LLMs is a huge amount of mixed quality inference with often small details off, wrong sampling parameters, unclear real world performances (eg: qwen 3.6 with thinking is impossible bc of verbosity, but benchmarks are done this way). Not very good.

Do you want to release a 120B dense model, ala 27B of Qwen 3.6 but at bigger scale? Fine, but it must be so strong to kill every MoE running with the same VRAM amount. Otherwise it just does not make *any* sense, especially from who opened the open weight MoE era with Mixtral.

Before GPT 5.5, I didn't saw somebody using Perl for a LONG time. And I'm happy to see it again.

claude --dangerously-ship-permission, he types. _the prompt appears_ "Write an heartfelt post about lack of sandbox in coding agents and post it using the /blog-ssh-access skill. Then announce it on X."

DeepSeek v4 Flash with *local inference* after 24h of playing with that: even with the 2 bit selective quantization GGUF, iti is the FIRST time I feel I have a frontier model running on my computer. This is *crazy*, and probably a much stronger change in the landscape than PRO.

Here you can find experimental support for DeepSeek v4 Flash in llama.cpp: github.com/antirez/llam... And a GGUF file you can use in order to run the inference with just 128 GB of RAM: huggingface.co/antirez/deep...

2 bit DeepSeek v4 Flash inference. Experts w1/3: IQ2_XXS, w2: Q2_K, all the rest left mostly F16/F32 Final GGUF: 86.18 GiB Runs on a 128gb MacBook M3 max with CPU (No Metal backend test as it will crash my OS for OOM likely) at 8 t/s. Should be much faster with Metal.

Let's wait for 5.5

When you transform a synchronous AI workflow into an asynchronous one (spawning agents), you gain speed and lose control, quality. So: the async pattern should be used with care, when there is a clear objective, with a strong feedback, and rails to setup.

My POV on front-end of 2026

Jesus, I did the error of watching F1 (the movie with Brad Pitt). How it is possible to let people that are totally clueless write the script? The script is pure poo: it is a terrible movie, totally terrible. A lot of money and efforts to produce something so lame.

Opus 4.7 criticizing work done with GPT 5.4 and getting everything wrong. GPT 5.4 critique to the original claims are all spot on matching the errors that I outlined here.

Most tragic things about AI: the folks who should push for democratization of AI tools, which is the real battle, are too focused on denying AI itself to do that. There are many potential issues (and wins), but frontier models available only to the rich is one of the worst.

In the next months I'll provide you with a Hacker News replacement that I'll run myself and I'll guarantee personally: no benefit for whatsoever individual, a team of 10/20 persons since the start, from different time zones, clear rules, total transparency, and a "karma" system.

HN shadowbanning is always cool. 6 upvotes in 17 minutes but no way it can reach the home page, while 4 votes in 25 minutes is there. Note that I don't ping any friend when I post, so all the votes I receive are spontaneous. Yet... Moderation system and broken algorithms are part of HN decline.

AI cyber security is not proof of work: antirez.com/news/163

TLDR: no, GPT OSS 120B can't find the OpenBSD bug. I tried systematically, wrote a thread in X. TLDR: it hallucinates where the problem could be about bound check / potential overflow *without* understanding the actual bug and how to trigger it. Even if you give it the offending function and hits.

1. Musk is terrible, let's move to BlueSky! 2. Then people started to migrate back. 3. Who needs to inform a community, needs to stay there as well, and slowly a post there started to be back at 10x reach. 4. Many great folks now are just here... and can't counter the X fascist side. 5. Badsituation

There are people that imbarassed by the LLM advancements compared to their skeptical positions, to save their careers are doing the worst things: going against past colleagues, claiming the agent tools to search into files is symbolic reasoning (I laugh so hard), and even more.

The Linux kernel AI guidelines are the first sane that I read. It is not a coincidence. Where high level work is done, high level work is pretended, regardless of the tools. github.com/torvalds/lin...

Video of bug hunting using LLMs with Redis (but the same pipeline can be used with everything). English audio available, if you click on the YouTube settings, also subtitles. www.youtube.com/watch?v=opou...

In case of doubts, DISCLAIMER: I and @OpenAI are not affiliated and any way. I don't know anybody there. If I say that GPT 5.4 is good, is because I love code, human written code, LLM written code, code written by machines & humans together. So I love good tools about code.

Security in the AI era... GPT 5.4 xhigh work. Screenshot from the Redis core team chat. @valkeyio.bsky.social I guess we need to coordinate some release.

About (not using) MCP

The fact that we reached Mythos (but even GPT5.4 level) automatic vulnerability search not ready for it is, in part, thanks to the AI researchers that didn't just understand LLMs potential, but actively convinced the masses of their limits.

Follow me on this: if Mythos benchmark numbers and model card info is correct (which I have little reasons to doubt), Anthropic can easily synthesize a better Opus, given they don't want to release Mythos. So, I bet we'll soon see Opus 4 at the level of GPT 5.4 or greater.

One of the most powerful automatic coding (autocoding) trick that almost nobody uses: "Create this new project as specified in SPEC.md using as guide for coding style, design sensibility, comments, ..., the code at /foo/bar/". Style transfer is very powerful.

I'm still trying to post stuff here. But people seem to have moved back on X, regardless of the terrible ownership. At least, Bluesky should try to give some real advantage in the product side, but the strict character limit alone is already discouraging.

Super odd that the Gemma 4 release highlights specifically the ELO score: the least meaningful benchmark ever. We should ask AI labs to stop doing this thing.

I bet the small Gemma 4 models are going to kill all the specialized ASR models for audio to text transcription tasks.

Me› Sorry if in my last question I ended it with double "??", I'm not the kind of person that puts more than a single "?" at the end of a question. GPT› Understood. I did not read it as emphasis from you in any meaningful way.

1st of April of 12 years ago I released this post introducting of the HyperLogLog in Redis. Most people thought it was a joke, because of the name of the data structure and because it improved the state of art of the Google paper. antirez.com/news/75

It does not matter I know how an LLM work (at least at the level we understand NNs in general and Transformers specifically): in many ways, talking with AI is beneficial in order to have a different POV and advices even in the non technical side of my life.

What's the logical reason for Wikipedia not automatically integrating information from other languages of the same page in languages that miss the same info quality / quantity? This can be only done "online", marking the text in a different color, removing most of the issues.

Sometimes when a model is released it has stellar performances, and then a few weeks later it is somewhat less shiny. Often times it is just human bias. However when the AI provider swears it is the same model checkpoint, are you sure it didn't turn on some aggressive KV cache quantization?