<antirez>

I didn’t expect DwarfStar 4 (https://github.com/antirez/ds4) to become so popular so fast. It is clear that there was a need for single-model integration focused local AI experience, and that a few things happened together: the release of a quasi-frontier model that is large and fast enough to change the game of local inference, and the fact that it works extremely well with an extremely asymmetric quants recipe of 2/8 bit, so that 96 or 128GB of RAM are enough to run it. And, of course: all the experience produced by the local AI movement in the latest years, that can be leveraged more promptly because of GPT 5.5 (otherwise you can’t build DS4 in one week — and even with all this help you need to know how to gently talk to LLMs).

I started working on the new Array data type for Redis in the first days of January. The PR landed the repository only now, so this code was cooked for four months. I worked at the implementation kinda part time (kinda because many weeks were actually full time, sometimes to detach yourself from the keyboard is complicated), and even before LLMs the implementation was likely something I could do in four months. What changed is that in the same time span, I was able to do a lot more. This is the short story of what happened.

The proof of work is the wrong analogy: finding hash collisions, while exponentially harder with N, is guaranteed to find, with enough work, some S so that H(S) satisfies N, so an asymmetry of resources used will see the side with more "work ability" eventually winning.

But bugs are different:

1. Different LLMs executions take different branches, but eventually the possible branches based on the code possible states are saturated.

2. If we imagine sampling the model for a bug in a given code M times, with M large, eventually the cap becomes not "M" (because of saturated state of the code AND the LLM sampler meaningful paths), but "I", the model intelligence level.

Those who cannot remember the past are condemned to repeat it. A sentence that I never really liked, and what is happening with AI, about software projects reimplementations, shows all the limits of such an idea. Many people are protesting the fairness of rewriting existing projects using AI. But, a good portion of such people, during the 90s, were already in the field: they followed the final part (started in the ‘80s) of the deeds of Richard Stallman, when he and his followers were reimplementing the UNIX userspace for the GNU project. The same people that now are against AI rewrites, back then, cheered for the GNU project actions (rightly, from my point of view – I cheered too).

Here LLM and coding agents can find:

1. Exhaustive documentation about Redis commands and data types.
2. Patterns commonly used.
3. Configuration hints.
4. Algorithms that can be mounted using Redis commands.

https://redis.antirez.com/

Some humans claim this documentation is actually useful for actual people, as well :) I'm posting this to make sure search engines will index it.

Anthropic recently released a blog post with the description of an experiment in which the last version of Opus, the 4.6, was instructed to write a C compiler in Rust, in a “clean room” setup.

The experiment methodology left me dubious about the kind of point they wanted to make. Why not provide the agent with the ISA documentation? Why Rust? Writing a C compiler is exactly a giant graph manipulation exercise: the kind of program that is harder to write in Rust. Also, in a clean room experiment, the agent should have access to all the information about well established computer science progresses related to optimizing compilers: there are a number of papers that could be easily synthesized in a number of markdown files. SSA, register allocation, instructions selection and scheduling. Those things needed to be researched *first*, as a prerequisite, and the implementation would still be “clean room”.

In my YouTube channel, for some time now I started to refer to the process of writing software using AI assistance (soon to become just "the process of writing software", I believe) with the term "Automatic Programming".

In case you didn't notice, automatic programming produces vastly different results with the same LLMs depending on the human that is guiding the process with their intuition, design, continuous steering and idea of software.

Please, stop saying "Claude vibe coded this software for me". Vibe coding is the process of generating software using AI without being part of the process at all. You describe what you want in very general terms, and the LLM will produce whatever happens to be the first idea/design/code it would spontaneously, given the training, the specific sampling that happened to dominate in that run, and so forth. The vibe coder will, at most, report things not working or not in line with what they expected.

I love writing software, line by line. It could be said that my career was a continuous effort to create software well written, minimal, where the human touch was the fundamental feature. I also hope for a society where the last are not forgotten. Moreover, I don't want AI to economically succeed, I don't care if the current economic system is subverted (I could be very happy, honestly, if it goes in the direction of a massive redistribution of wealth). But, I would not respect myself and my intelligence if my idea of software and society would impair my vision: facts are facts, and AI is going to change programming forever.

* For years, despite functional evidence and scientific hints accumulating, certain AI researchers continued to claim LLMs were stochastic parrots: probabilistic machines that would: 1. NOT have any representation about the meaning of the prompt. 2. NOT have any representation about what they were going to say. In 2025 finally almost everybody stopped saying so.

* Chain of thought is now a fundamental way to improve LLM output. But, what is CoT? Why it improves output? I believe it is two things: 1. Sampling in the model representations (that is, a form of internal search). After information and concepts relevant to the prompt topic is in the context window, the model can better reply. 2. But if you mix this to reinforcement learning, the model also learns to put one token after the other (each token will change the model state) in order to converge to some useful reply.

I’m taking a few weeks of pause on my HNSWs developments (now working on some other data structure, news soon). At this point, the new type I added to Redis is stable and complete enough, it’s the perfect moment to reason about what I learned about HNSWs, and turn it into a blog post. That kind of brain dump that was so common pre-AI era, and now has become, maybe, a bit more rare. Well, after almost one year of thinking and implementing HNSWs and vector similarity stuff, it is time for some writing. However this is not going to be an intro on HNSWs: too many are present already. This is the “extra mile” instead. If you know HNSWs, I want to share with you my more “advanced” findings, especially in the context of making them fast enough to allow for a “Redis” experience: you know, Redis is designed for low latency and high performance, and HNSWs are kinda resistant to that, so there were challenges to expose HNSWs as an abstract data structure.