Direct Weight Editing via Codex+Astra Lifts Audio Classifier Accuracy From 84.3% to 88.75% Without Retraining
dansitu · x · 2026-11-03
dansitu tested Codex+Astra on a bird-sound classifier (sparrow/parakeet/noise): the agent pulled the float32 tflite model, reproduced Edge Impulse results locally, then directly edited weights without retraining.
Results across 12,147 windows from 820 labeled recordings:
- float32 full dataset: 84.30% → 88.75%; holdout (not used for edit selection): 84.07% → 89.00%
- INT8 full dataset: 85.25% → 89.57%
The two edits: averaging final-layer weights across time to reduce sensitivity to call position within the window, and lowering the parakeet bias by 1.0 to curb over-prediction. Model size unchanged; tradeoff is parakeet recall dropping from 96.58% to 93.38% while sparrow and noise recall improve.
More from coding & agent
- A one-night AI-built project nets 333 points and ~22k visitors — Aizkmusic · 2026-11-03
- Five manual checks before trusting an agent actually finished its job — Kodenstudio · 2026-11-03
- Open-source AI pit crew for iRacing runs 5 agents on Kimi, checks every number against telemetry — JeremyCMorgan · 2026-11-03
- Agent control plane benchmark: Smalltalk-80 cascades hit 100% compliance vs 2.5% for Gherkin — RandalSchwartz · 2026-11-03
- Karpathy: one prompt to Claude Code deploys a YouTube thumbnail-swap browser extension — amaarora · 2026-11-03
- Cherry Pi proposes layered harness architecture: agent loops compose with governed workspaces — sull · 2026-11-03