Phonon-2: 164MB open English STT beats Whisper on a laptop
Open weights that match a 2.5GB teacher and run fully offline — plus ChatGPT's free and Go tiers begin showing product ads in image generation.
Copy markdown
A 2.5GB teacher's accuracy in 164MB
Fermion's Phonon-2 scores a 5.21% word error rate across seven English datasets — past Whisper large-v3-turbo's 6.58% at roughly 10x smaller — and transcribes an hour of audio in about 20 seconds on an M5 MacBook Air (174x realtime; 6,680x batched on an H100). It ships as CC-BY-4.0 open weights with pip and Docker installs, so offline, on-device English transcription is a drop-in today.
ChatGPT's free ride starts carrying ads
OpenAI opened a US test of visual product ads — an image plus a 'Learn more' button — appearing first during image generation in ChatGPT. They run on the free and $8/mo Go tiers; Plus, Pro, Business and Enterprise stay ad-free, and ads are kept out of health and politics. If your day-to-day leans on free or Go image gen, expect sponsored results to start showing up later this month.