Month 8 of building ATLAS: pushing a local small model toward frontier coding ability on a gaming PC
itigges22 · reddit · 2026-09-18
An 8-month solo research project called ATLAS attempts to close the gap between a small local model and frontier intelligence—without touching weights—by engineering the system around it, starting with code where compilation, tests, and execution feedback give verifiable signal.
V1 (Jan 2026): llama.cpp quantized local inference, RAG over the repo (embeddings + Qdrant) to compensate for small context, sandboxed execution with retries and per-attempt temperature tuning, plus a Redis/workers/K3s task queue and deployment stack.
V2 (Feb 2026): Best-of-K candidate generation with diversity-aware selection, and the "Geometric Lens"—a small neural net trained on code embeddings and execution pass/fail labels producing a scalar cost field C(x) to rank candidates (post truncates here).
More from coding & agent
- General Reasoning explains how to train agent swarms with RL and latency pressure — rosstaylor90 · 2026-11-02
- "Coding isn't the bottleneck anymore — usage windows are," says founder Kyle Gawley — kylegawley · 2026-11-02
- Demo passed ≠ production-ready: 5 questions to answer before letting agents run your business — _kellhus · 2026-11-02
- An AI agent passed 100 test chats—still not ready to run the business — djgoel123 · 2026-11-02
- Agent given total creative freedom produces a 1:44 code-drawn animated short film — ivegotnoidea1 · 2026-11-02
- Meme: Wake Up, Use Claude Code, Repeat All Day — tekbog · 2026-11-02