Squareblack shifts to classifier models: local LLMs write strict eval metrics, no frontier needed
michalmalewicz · x · 2026-10-21
Squareblack's michalmalewicz shares how his team is moving into classifier models (Jev, then Basal):
- Entry prompts with context and knowledge support the evaluations, with a specific set of grading metrics defined per case
- Metrics are either annotated manually or written by an LLM under very strict constraints, then tweaked
- That LLM is also local — when you constrain it enough, frontier models aren't needed
His takeaway: the direction is "definitely not what I thought it would be last year." The core insight is that with tight enough constraints, small/local models suffice.
More from coding & agent
- Twenty, the open-source AI-native CRM alternative to Salesforce, nears 59k stars — twentyhq · 2026-10-21
- freehire: open-source AI job search engine hits 2k GitHub stars — strelov1 · 2026-10-21
- Running multiple Claude instances to talk to each other slows everything down — rickasaurus · 2026-10-21
- Vibe coding without judgment is just 'lottery with syntax,' developer argues — robertonappe · 2026-10-21
- Indie dev vibe-codes a beehiiv alternative, hits $7,836 ARR in one week — marclou · 2026-10-21
- Veteran coder says treating Claude as a design partner made building software fun again — shimoheihei2 · 2026-10-21