Fetching from the wire…
01
02
03
04
05
06
07
08
OSS2026-08-21 · source-backed
Decoder-only transformer with RMSNorm and rotary embeddings, using note-level tokenization NOTE(pitch, delta_onset, duration, velocity) so one forward pass advances the music by a complete note rather than an event fragment. Trained on a few hundred thousand MIDI files, ~300M note events. DPO moved pairwise preference from 24.55% to 69.05%, judged by Gemini 3.5 Flash. simedw.com 558 points on HN. Aggressively cleaned classical data beat larger noisier sets, which is the transferable lesson.
Each link below shares sources, entities, or timing with this story.
Gemini built by Google / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Gemini built by Google); both cover Flash, Gemini; overlapping topics (data, flash).
Linked by a graph relationship (Gemini built by Google); both cover Flash, Gemini; overlapping topics (beat, flash).
Linked by a graph relationship (Gemini built by Google); both cover Flash, Gemini; overlapping topics (data, flash).
Gemini built by Google / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Gemini built by Google); both cover Flash, Gemini; earlier Flash coverage from 2026-08-08.
Linked by a graph relationship (Gemini built by Google); both cover Flash, Gemini; earlier Flash coverage from 2026-06-19.
Claude Code benchmarked against Gemini / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Claude Code benchmarked against Gemini); both cover Flash, Gemini; earlier Flash coverage from 2026-07-23.
Gemini competes with ChatGPT / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Gemini competes with ChatGPT); both cover Flash, Gemini; earlier Flash coverage from 2026-03-17.
Vercel uses Gemini / Shared entities / Earlier coverage
Linked by a graph relationship (Vercel uses Gemini); both cover Flash, Gemini; earlier Flash coverage from 2026-08-14.