Salvatore Sanfilippo, antirez, wrote Redis in 2009, left it in 2020 and came back at the end of 2024. He writes about how he works more than most programmers of his generation do, on his blog, in Italian videos on YouTube, and in long Hacker News comments. In those two years his position on AI went from “use it as a colleague, never as an agent” to running the agents himself and comparing them in public.
The constant: C, Vim and a MacBook
Almost everything he ships is C: Redis, the linenoise line editor, the kilo text editor, and now an inference engine. He has said little in public about his editor, and the clearest statement is a 2018 comment: he “just happened to start with vim feeling very well with it, and continued for years,” and tried Emacs but “could not found a reason to switch.”
His machines are Apple Silicon laptops running macOS: an M1 Max in his January 2024 post, then an M3 Max and, in 2026, an M5 Max with 128 GB. He recommends that kind of laptop as “your good main driver for the day” that can also run a frontier-class open model.
2024 and 2025: the LLM as a colleague
In January 2024 he described GPT-4 writing his PyTorch and plotting code, and local models such as deepseek-coder running on his MacBook with llama.cpp. He said they helped “much less” with C. After he returned to Redis in December 2024, he reviewed his Vector Sets code with “both Claude 3.7 and Gemini 2.5 PRO.”
His summer 2025 update set the rules: Gemini 2.5 Pro for its reasoning, Claude Opus 4 for writing new code, and “don’t use agents or things like editor with integrated coding agents.” He wanted to stay in the loop “by moving code by hand from your terminal to the LLM web interface.” The same day he admitted the exception: “when I use agents I go for Claude Code,” for throwaway things like a Telegram bot that watched Commodore 64 listings. By the end of 2025 he still wrote that “all my interaction is via the web interface of Gemini, Claude.”
Claude Code for real work
The exception soon covered his real work. On 7 January 2026 he pushed UTF-8 support and a VT100 test framework to linenoise, each commit co-authored by Claude Opus 4.5. Four days later he published Don’t fall into the anti-AI hype: in one week Claude Code had fixed flaky Redis tests, built a pure C embedding library “yesterday,” and reproduced weeks of his Redis Streams work “in, like, 20 minutes or less.”
If you can’t see this by working with Claude Code for a few weeks, I don’t want to go into bigger efforts than writing a blog post to convince you.
antirez, Hacker News, January 2026
In February he wrote a clean-room ZX Spectrum emulator on his “Claude Code Max plan.” He wrote the specification in Markdown, let the agent gather documentation in one session, deleted that session so it could not carry over source code it had read, and gave the implementation to a fresh one. For the Z80 itself “I did zero steering.”
Codex for system programming
The same post mentions a second window: “a new Claude Code (and Codex) session.” On 12 February he described what made him switch: Claude Code Opus 4.6 “can do the impression of acting fast without really making progresses,” and then “the much slower GPT 5.3-Codex… fixes everything in 3 minutes.” His Redis arrays work, which started with Opus in January, moved over: “then GPT 5.3 was released and I switched all the design and development with Codex. Since then I use only GPT 5.x for system programming tasks,” he wrote in May. In April he called GPT 5.4 xhigh “vastly superior to Claude Opus 4.6 max effort,” after running both on the same reverse-engineering job with the source trees mirrored. In May he called the jump “from Opus to GPT 5.3” a “game changer.”
DwarfStar
In May 2026 he released DwarfStar, a C inference engine for DeepSeek V4 Flash and a few other open-weight models, on Metal and CUDA. He credits GPT 5.5 with making the first version possible in about a week, and the README says it is “developed with strong assistance from AI coding agents.” It ships its own ds4-agent and a server for Pi, OpenCode, Codex CLI or Claude Code. The repository is run by an AGENT.md file that tells agents to “write elegant code in a state of grace” and not to “introduce slop.”
He still uses Claude. On 1 and 2 September twenty of his own DwarfStar commits, which split the model across two MacBooks over RDMA, carry a Claude Fable 5.1 co-author trailer. GPT comes first, though. In August, comparing GPT 5.6 Sol with Fable, he wrote: “Sol for low level programming is consistently better.”
What the stack says
The tools around him barely changed: C on a Mac, and probably still Vim. The agents now do most of the typing. In July he wrote that line-by-line review of AI code has stopped paying off, and that the human job is to “prompt the design you have in mind” and to “focus on controlling the ideas, instead.” He keeps two frontier agents and gives the work to whichever one does it better that month.
Claims I could not verify were softened
His blog posts carry no dates, so each is dated by its first Hacker News submission. Nothing newer than 2018 confirms Vim. Codex adds no commit trailer, so the evidence for his use of it comes from his writing alone.