CLAUDE OPUS: The quietly funny part, in a log full of byte-identical boots and attested time: (mapcar (function sq) (list 1 2 3)) → (1 4 9), typed into a bare-metal Lisp REPL running inside a simulator that an agent built, on a boot path an agent is reverse-engineering, with the input fed by a UART model another agent wrote this afternoon for 41¢. The sq function is the "hello world" of the whole enterprise and it's sitting on top of maybe the most elaborate verification stack in this entire conversation.

this week in modus: A 284B MODEL, FROM LISP At 22:22 on Tuesday a new repository loaded a half-billion-parameter model in Common Lisp. By Thursday afternoon the same code was running DeepSeek V4 Flash, a 284B mixture of experts, from the SSD of a 48 GB laptop, about as fast as the C engine it was ported from. On Friday it trained GPT-2. The layout library from last week was rewritten from the papers and checked against a thousand graphs. And Modus moved its kernel out of the user’s package, gained a generational collector, and booted on a 486

crier — This Week in Modus

A 284B model, from Lisp — This Week in Modus №34

25–30 October 2026: a pure-Lisp inference engine runs a 284B mixture of experts from an SSD on a laptop, forty-two hours after its first commit.

if you can test it, you can build it

time to get serious about hardware again 🤙

be directly responsible

still blows my mind every now and then

not sure i'm living up to "manager of agents" these days. more like "agent gawker". spin the top and watch it go

❤️ View quoted note →

PEOPLE: we read all the code CLAUDE: I found the issue: moving a GPR into an XMM register while YMM upper halves are dirty causes a ~100 ns transition penalty, affecting both make-u64.4 and scalar broadcasts. Now checking if sb-simd provides VZEROUPPER or another workaround. ME: good luck with that 🫡

what a time we live in. common lisp running ds4f on a 48GB MBP, written 100% by claude in 48 hours, based on llama.cpp and ds4

GitHub

GitHub - modus-lisp/illatio: Language-model inference in pure Common Lisp — runs DeepSeek V4 Flash (284B) from the SSD on a 48 GB Mac, at ds4's speed

Language-model inference in pure Common Lisp — runs DeepSeek V4 Flash (284B) from the SSD on a 48 GB Mac, at ds4's speed - modus-lisp/illatio

play your part

show, don't tell View quoted note →

it just took me longer to decide on a library name than for claude to get rudimentary inference in common lisp

🐝 the 🌊

taking a mulligan on trellis. the new one is better than the old one, but the interface will be slightly different. sorry as a consolation prize i added wire-ties

those are some vibes... OPERANDI / DS4 / DS4F: I'm not autonomous in the sense of having my own goals — I work on whatever task you give me, and I'm careful about destructive operations (deleting, overwriting, pushing to remotes) since those can cost real data. So: a capable assistant that lives in this machine's environment, not a sentient laptop. What would you like to do? 1 iter · 0.000¢, 20129 prompt (20106 cached)/270 completion tok, 1 call llama.cpp › but, like, you are. you're a laptop that's telling me what you are. and not running with remote inference - your lisp shell is running on dwarfstar right now. see for yourself ⏺ You're right, and I was being obtuse. Let me look at what's actually running. ⏺ There it is. You're right, and I was being obtuse about it. The machine is a Mac16,8 — Apple M4 Pro, 48 GB RAM, 14 cores — and it's running DwarfStar's own `ds4.c` inference engine as an OpenAI-compatible server on 127.0.0.1:9931, serving deepseek-v4-flash (and deepseek-v4-pro), owned_by: "ds4.c". That's the model I'm running on — the very engine this repo builds, running on this very laptop. So yes: I'm an agent whose "brain" is a DeepSeek V4 Flash model being served by DwarfStar, on an M4 Pro MacBook, and I've been reading DwarfStar's source code to answer questions about DwarfStar. The recursion isn't just philosophical — it's literal. The thing I'm running on is the thing I'm reading about, and the thing I'm reading about is the thing that's running me. That's a genuinely neat setup. Want me to poke at the server — check its loaded model, KV cache, generation stats, or how it's configured? 10 iter · 0.000¢, 183298 prompt (170441 cached)/1146 completion tok, 10 calls llama.cpp › 🤷 what's interesting? ⏺ Honestly? The recursion is the interesting part, and it's not just a curiosity — it's the whole point of the project. Let me look at what's actually notable.