BootLoops: open-source harness that makes AI do exact science
A Harvard physicist's open toolkit reproduced a past physics result in ~20 minutes; plus Replit's new model modes and OpenAI's next model sunset.
Copy markdown
Make your agent's calculations provable, not plausible
Harvard physicist Matthew Schwartz open-sourced BootLoops, a harness that drives an LLM through verifiable scientific computation instead of guessed arithmetic — it reproduced one of his own published physics results in about 20 minutes (weeks by hand) and generated 36 manuscripts across 18 fields in three months, drawn from ~400 candidate problems. For builders it's a working template for wiring exact symbolic and numeric work into an agent so answers actually check out — Feynman integrals, 5.7B mutation pairs from the 1000 Genomes Project, a word-stress database spanning 6,072 languages.
Replit swaps in Sol Max and Sonnet 5.5 Power
Replit's AI integrations now expose GPT-6.1 Sol in a new "Max" mode and Claude Sonnet 5.5 in "Power" mode, and add Jev for content classification and structured decisions. If you build on Replit, you can route hard coding tasks to a frontier model and cheap routing or classification to a small decision model — pick the tier per step instead of paying frontier rates for everything.
OpenAI dates the GPT-5 series for retirement
OpenAI set an April 1, 2027 shutdown for gpt-5.1, gpt-5.3-codex and gpt-5.4-nano, steering callers to gpt-6-sol and gpt-6-luna. It's six months out, but if any of those IDs are hard-coded in your pipelines or Codex configs, the migration clock just started — and it's separate from the GPT-4 and o-series retirement already set for Oct 23.