iiinova — where your digital self keeps evolving

I'm thankful I was thinking like that. And I believe many other systems and companies that failed in between, all considered, now would not blindly buy the "lineraizable at all the costs" argument anymore.

070

So a system that takes some care to avoid common failure modes, or that just limits the outcome of failures, if it pays this compromise to have stellar properties when things don't go bad, it is a system that has a place.

180

Back in the days I realized that a system can't be characterized just for the way it handles whatever complex failure modes can be simulated, but it is actually the sum of how it behaves across all its lifetime of execution.

170

That's not to say that linearizability is not a good property for certain use cases, but that safety is actually a tradeoff, and one can be still vulnerabile to complex partitions/restarts scenarios while still having different degrees of real-world safety for most setups.

190

But I believe that not following the vision that wanted, in the late 2000s, every database to be able to guarantee strong consistency properties, was one of my best views that basically *saved* Redis from the same fate of many other systems that failed.

1110

antirez @antirez.bsky.social · 03/10/2026

[Thread] That's interesting, on HN there is a "from Redis creator" news right now (community site, the DwarfStar real site is GitHub) and somebody replied that since it is "from Redis creator" people should probably we skeptical because Jepsen Redis results.

3121

Those models can be either *small* LLMs that are not let think, but instead the logits at the last position of the prefill are used to categorize in classes or, when they are BERT-alike, they compress the input meaning and project a class: in both cases they decide a lot less than an LLM.

080

antirez @antirez.bsky.social · 03/10/2026

Calling classifiers "decision models" is a crime against machine learning. Those models decide *less* than LLMs, as you can think the CoT as a form of serial processing of information that such classifiers lack.

5231

antirez @antirez.bsky.social · 01/10/2026

*** STOP DEATH PENALTY IN THE USA ***

1363

antirez @antirez.bsky.social · 01/10/2026

Wrong advice for junior developer: look at the AI generated code, since you need to mature as a programmer. Right advice: save time with AI coding and write by hand "get-better" programs: interpreters, games, ray tracers, ...

5772

antirez @antirez.bsky.social · 30/09/2026

Note how ds4-agent does not permit a cache miss from happening. This is a byproduct of the fact we have a serious problem IMHO with the current local inference pipeline.

020

antirez @antirez.bsky.social · 30/09/2026

Totally understand it is not for everybody but this is an upper limit. Most people will be absolutely ok with a single 100$ plan. Some even with the 20$ one.

001

This in turn will make the DGX Spark and Strix Halo a lot more compelling, so do not assume that the hardware you got can be valuated in the void. It depends a lot on the class and exact strengths of models that are released.

2130

Don't look at generation tokens per second as the main metric for local inference. Soon Cinese open model providers will realize that the game is all about making the thinking phase as short as possible, and 20/30 t/s will be enough if you have fast prefill.

6371

ANTIFA for the win of course. Btw left-win anti-AI orientation is allowing right-win folks to dominate the narrative around AI. Yet another self-inflicted goal of left-win: in this way we will deliver the world to fascists.

0120

antirez @antirez.bsky.social · 28/09/2026

I claim that if you are a normal person that like me don't have "fuck you money", so unable to spend large sums of money randomly, you should do all your programming work with two max accounts (OpenAI / Anthropic for instance) in the worst case, and never buy tokens via API.

8720

The junk problem is an illusion, most of the programs written before AI are terrible crap. Now it is much simpler to write much better code. The crap produced will appear particular subpar since it can re-done better faster, and will be more isolated and forgotten.

4141

antirez @antirez.bsky.social · 28/09/2026

Now "slop! sl00p! SLOP!!!111". Just tell the truth: you are rightfully scared about how to pay your invoices and bring forward your family, and this is a correct and respectable concern. But say things as they are.

3335

antirez @antirez.bsky.social · 28/09/2026

The most ridiculous thing about the AI-coding age is that in the last 20 years programmers accepted otherwise all the kind of shit: terrible frameworks, languages, layers of complexity everywhere, everything super slow. Very few protested, since destroying the field didn't involve their paycheck.

3544

antirez @antirez.bsky.social · 27/09/2026

Good programmers state they use models like GPT 6 Astra without obtaining good results: it makes me feel I live in a parallel universe. But the explanation is not that hard: good programming in the past and now requires a different skill set, even if there is some overlap.

8595

antirez @antirez.bsky.social · 25/09/2026

I'm worried to see people that think the right move is to question AI companies and narratives around AI regardless of logic and the actual significance of events we are living. They act just as an inverted flock, but still a flock. Sometimes it's fear of friends judgement.

4192

antirez @antirez.bsky.social · 23/09/2026

Trivially known for years but since apparently every year everything about machine learning is new to newcomers...

4311

antirez @antirez.bsky.social · 21/09/2026

Jev may have its (narrow) use cases but the hype you see around is the perfect representation of the fact the greatest part of the AI bubble don't know what is important and what is not. Jev is a minor thing happening on AI compared to all the rest, yet the hype exploded.

15914

antirez @antirez.bsky.social · 17/09/2026

Started yesterday as a joke without any code reference if not papers. Because of fillets already better than TinkerCAD for certain stuff. There is something odd, here: in theory open source should explode right now, but we don't have, apparently, folks motivated to do wonders.

3320

antirez @antirez.bsky.social · 16/09/2026

At this point DwarfStar contains many fast fused kernels for important model families: feel free to steal everything you want from there according to the MIT license, in order to improve your own implementation.

0654

antirez @antirez.bsky.social · 16/09/2026

Witch hunting level: some guy uses AI to reverse engineer an M4 GPU driver for Linux and part of the community that should be for the open source, for the hacking, for the liberation and freedom is against him.

415716

antirez @antirez.bsky.social · 15/09/2026

The irreconcilable misunderstanding about AI is that for some of us is the way to remove suffering, inequality, limits from the human race. For others, a tool: and, right now, the capabilities are at tool level, giving the illusion that the tool is the point.

2320

antirez @antirez.bsky.social · 14/09/2026

Not sure if this makes sense on the DGX Spark. You have 128GB there, however the extra speed could be very tempting for the Spark users, as well as having a lot of RAM free for other stuff. For ROCm it makes sense as many folks have a Framework Desktop with 64GB.

290

antirez @antirez.bsky.social · 14/09/2026

Qwen3.8 Flash Next is now supported in DwarfStar, covering 64GB Mac systems very well and with very fast inference of 50~70 t/s and > 1400 t/s prefill. For now this is Metal only. Thanks to @ivanfioravanti.bsky.social for all the cool work in the PR. N-grams on SSD like for DS4.1F.

3778

antirez @antirez.bsky.social · 14/09/2026

It drive me nuts that now people use AI to write tweets. If you think that this way you will get more popular and so forth, think twice. Write something authentic. AI is great but it is easy to misuse. Your tweets must be your more cared thoughts.

51024

antirez @antirez.bsky.social · 13/09/2026

With the last commit into DwarfStar now you can use DeepSeek v4.1 Flash in a single DGX Spark as well, with SSD streaming. Around 9 t/s generation. It works also dual-spark RDMA at ~22 t/s.

0656

antirez @antirez.bsky.social · 13/09/2026

Conspiracy theories are, very often, illogical. If AI companies were worried by open weight models (they probably are btw) the logical response would be to *not* slow down the development of frontier AI, to try locking the advantage. Stop with nonsense.

3212

antirez @antirez.bsky.social · 12/09/2026

DeepSeek v4.1 Flash support is now pushed on DwarfStar "main" branch on GitHub, and this is a YouTube video (in English language) where I test both the SSD streamed and the dual MacBook m5 max 128GB setup during a coding session: www.youtube.com/watch?v=ogs8...

youtube.com

Let's test DeepSeek v4.1 Flash 2 bit with DwarfStar (ENG)

YouTube video by Salvatore Sanfilippo

1467

antirez @antirez.bsky.social · 11/09/2026

About Anthropic banning minors from Claude.

5321

antirez @antirez.bsky.social · 11/09/2026

I wonder what face they put on when something really surprising happen to them. The risk is of a severe muscle tear.

1130

antirez @antirez.bsky.social · 11/09/2026

I'm a simple man. If a YouTube video cover has a stunned face on it, I don't watch the video.

816013

antirez @antirez.bsky.social · 10/09/2026

DwarfStar running DeepSeek v4.1 Flash on a 128GB M5 Max at steady 15 t/s. I didn't expect with SSD streaming it could be so fast. Recent SSD streaming changes to retain the right experts surely helped, but also maybe DS4.1 uses the same experts more. Will push online when ready QA > ASAP.

29413

antirez @antirez.bsky.social · 10/09/2026

The third chapter of WOHPE opens like that. Written mostly during 2020, published July 2022. My book was among the most successful sci-fi books published in Italy in the latest 10 years. Yet, I believe that it deserved a bit more.

240

antirez @antirez.bsky.social · 10/09/2026

P.S. I believe the credits should go exclusively to Córdoba–Martínez-Zoroa and AI.

010

7. So Tao claim is surely false from the POV of the net knowledge. The open part remains in the math trajectory, but it starts from premises that look very unreal (see the previous points). The source of such claim is not guarantee of correctness. Remember Stochastic Parrots.

270

6. AI ability to produce math at a super-human level in certain areas is now quite clear. Who claims math need to remain the same as yesterday really requires stronger arguments than the ones in the Tao thread I read.

170

5. Closed companies providing proofs can only add to mathematic, they can't remove. The math community will continue to have all the past tools plus the new proofs, plus, likely, access to advanced AI. So if the math community will be willing to make progresses, it still can.

380

4. Tao reading requires a use of the AI tool that nobody in the math community would pick. If Tao position is not in extreme minority, the generated proofs will not remain unaddressed and unstudied. There will be efforts to study them in-depth, producing math results.

170

3. In the past, to address Fermat, Poincaré new techniques where developed. But this does not mean that with AI in order to study / simplify more math no new techniques will be developed. There is no reason this is a process that can happen only with exclusion of AI.

1100