strata-cluster
C++One 125B mixture-of-experts model split across two home PCs over plain gigabit Ethernet. An unofficial, experimental fork of the Strata inference engine.
Ph.D. · AI research & tools · vintage computing · photography
After a Ph.D. in electronic and computer engineering and research at the Istituto Italiano di Tecnologia, I now work in applied AI R&D in Genoa. I run large language models on recycled hardware, study when a fast decision model should stop and think, restore retro computers and still write code for the Commodore 64, these days with AI agents next to me. Away from the keyboard, I photograph the Ligurian coast in natural light.
One 125B mixture-of-experts model split across two home PCs over plain gigabit Ethernet. An unofficial, experimental fork of the Strata inference engine.
When should a fast decision model stop and think? Code and data of the paper: a learned actor plays Doom and hands control to a reasoning vision-language model only when a gate opens.
AI agents that write, debug and even playtest Commodore 64 assembly games, through a feedback loop with the VICE emulator and Python visual checks.
The Laya System One decision model against Gemma 4 E2B and Gemma 3 270M fine-tuned with LoRA on the same data, in FreeDoom: accuracy, calibration, latency and real games.
A multi-agent system for writing a book together, inspired by Pirandello: the characters are autonomous agents with their own memories and voices.
Can a small LLM move from brute-force memorization to intuition and tool delegation? A proof of concept on Qwen2.5 with LoRA and structured pruning.
arXiv:2610.09683 · cs.AI
About twenty earlier papers (2000–2006) on video coding, surveillance and telemedicine. Indexed as both Gian Luca Bailo and Gianluca Bailo.
The day after its release, Anthropic’s newest model built an eight-way scrolling C64 platformer in about two hours, drawing every pixel from scratch. The 1982 machine still landed one punch: a bug only a human eye could see.
Towards AI
A second PC on a plain network cable made a 4-bit model that did not fit in RAM 87% faster at the same quality. Long prompts got slower, and where one PC was already enough, the second added only power draw.
Towards AI
38 to 48 tokens per second from a Xeon without AVX2 and two 2017 GPUs, and what three home PCs taught me about where local LLM speed really comes from.
Towards AI
Towards AI
A “System One” decision model and a small general-purpose LLM, trained on the same 979 examples, play the same Doom levels on the same GPU. The answer is less exciting than the launch posts, and more useful.
Towards AI
Codex vs. the Commodore 64: a benchmark nobody funded, on a machine nobody can game — and what a 1 MHz CPU says about an agent that a leaderboard cannot.
Towards AI
Rung seven of the ladder, measured on release week: banded colour RAM, a raster HUD split and a dozen multiplexed sprites on a 1 MHz CPU — built by an agent that measured before it drew.
Towards AI
Hardware is declared obsolete against a general-purpose yardstick. Inference is not general purpose. Pick one task, and parts written off years ago come back with most of their value intact — for a third of the money.
Towards AI
Everyone will publish the benchmark scores. Here is a less glamorous and more useful question: what does it actually cost to run on a home machine — and why that number is not up for negotiation.
Towards AI
Towards AI
My most downloaded photos on Pixabay: sea, sky and storms in natural light. Click a photo to see it large. More on Instagram and Pixabay, free to use.







