antirez

If you print on the gear, don't you see the different tracks? I understand that this works on desktop and mobile *app*, but not mobile browser maybe.

A remaining big problem is that the system is not using the Italian subs that I upload to do the English dubbing. I bet the model is audio to audio, so that's not possible if not as a hint. So there are errors that are NOT present in the auto translated english subs.

YouTube video by Salvatore Sanfilippo Impariamo il C, lezione 28: le funzioni con un numero variabile di argomenti

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

LLM hallucination in this case.

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

Or is it proven that any algorithm would fail even under such assumptions because in the end there will always be an edge condition?

Yes, but I believe it should be possible to use the trick of waiting for a very big gap before to cut to N tokens, and if if such gap is not found in the first M items, to just take fixed M. Something along those lines should guarantee that there is this issue AFAIK.

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.

LOL I can't resist retweeting this

Also, why the new config is not tested in a single instance before propagating? You could expect you have a toy-cloudflare that replicates the normal workload used as a canary.

Now, seriously, this is the most basic failure mode possible. At Cloudflare scale you should probably don't even ever have a critical process in a single implementation, NASA style. But the fact people continue doing this mass-propagation of updates/configs is super odd.

"We use Rust, nothing can do wrong! Let's propagate a configuration to all the servers in the blink of an eye".

Gemini 3 finds the Redis Streams bug I was talking about a few days ago like (and better) GPT 5.1 could. It's a complex bug and yet it creates a clear causal model of what is happening in the crash report.

Interestingly ChatGPT 5.1 just discovered and fixed a complicated bug in Redis Streams (of my fork: no bug in the official Redis) that Gemini 2.5 could not handle. Curious to see if Gemini 3 will be able to spot it.

I wrote a very long blog post about all I learned about HNSWs in the last year: "Scaling HNSWs" -> antirez.com/news/156

@antirez.bsky.social's C course is the first time I'm seeing YouTube auto-dubbing: www.youtube.com/watch?v=-1Zh...

I'm not sure how I feel about it, but it probably is a net positive due to being able to consume that content in English?
Still prefer original audio with English auto-subtitles though

They (YouTube) recently announced a massive upgrade of this feature. I believe it will get better and better to the point that will sound almost natural.

Required remark about Fil-C

I offer help for free, via email, to people using Redis Vector Sets in production or if they are experimenting with that. Feel free to ping me! If you don't know what I'm talking about, check this: redis.io/docs/latest/...

Good point! I generalized a bit because I saw LLMs finding similar bugs in Redis C code just from: carefully review this C code, but yep the failing test makes the goal simpler. Thanks!

YouTube video by Salvatore Sanfilippo Se l'AI non ti è d'aiuto quando programmi, hai un problema

Ciao Filippo! Mi sono permesso di citarti qui:

Se l'AI non ti è d'aiuto quando programmi, hai un problema
youtu.be/A-HyRCCa5lE

When you use LLMs to write code in a human-in-the-loop way, you ARE the agent. Multiple LLMs in different web tabs, you write the best prompts, you check the code, ask the right questions, cross-check with some other LLM. Being good agents is not trivial, but very rewarding.

Smart move. I believe this does not change the usefulness and fun of AoC at all.

Hi friends, this summer I published my latest sci-fi short story with Urania-Mondadori. I attempted to translate it in English, for the potential English speaking readers here, and I believe the result is truthful to the original. This is the ePub file.

antirez.com/sincro-en.epub

It's fantastic, isn't it? LLMs can do incredible things and can use tools in complex ways. Each time we discover some way LLMs can use tools, we call it with odd names and complicate it: MCP, Skills, ..., but it's the same f**king LLM that can do things.

Why I believe there are a lot of misleading papers about LLM reasoning.

That's the whole point. The concept exists almost everywhere but in places like Sicily and China it assumes a fundamental role in life.