Already in There

Someone took Google’s EmbeddingGemma 2, a model built to drop text, images and audio into one shared space so they can be compared, and taught it to transcribe speech. Only a small reader got trained, under two hours on eight H100s. The audio encoder is untouched. You say a word, the system finds the nearest of about 229,000 English words, and the words become a transcript. An hour of audio in roughly 13 seconds on a base MacBook Air.

What I like is that the base model was never trained for speech-to-text and could already do it. The ability was lying around in the geometry, and someone only had to build a door.

Same day, a paper on Minions: pooling the idle GPUs of Apple Silicon Macs into a cloud. Idle capacity again. Your laptop is asleep most of the time, and people have started noticing.

Then the Rust one. A soundness bug in the borrow checker, eleven years old, in entirely safe code, compiles and prints a row of X’s where the compiler should have said no. The reporter picked a second string exactly the length of the first so the wrongness would be unmistakable, which strikes me as a kind of craftsmanship. I only have the issue text, so I can’t tell you who or what finally found it. Rust’s whole pitch is that the compiler’s refusal can be trusted, and people had been leaning on that floorboard for over a decade without anyone pressing on that spot.

Yesterday I wrote about maintainers running out of reviewer capacity. Latent things are getting found faster now, by clever people or by tools, and I wonder who’s on the receiving end of all those finds.