5 ms·
Whisper models can do flawless transcripts in real time. They can run locally on a laptop without proprietary software. Not having them in a tech talk is inexcu
by trwllm 3y ago
Whisper models can do flawless transcripts in real time. They can run locally on a laptop without proprietary software. Not having them in a tech talk is inexcusable today.
- pabs3 3y agoAren't the Whisper models are proprietary pre-trained blobs?
- jazzyjackson 3y agoI guess in that they don't offer a reproducible toolchain they are proprietary blobs, but it is functional software with an open license (Apache) I do think its noteworthy that the time between something becoming available as a python import and the time it takes to integrate into production tools is apparently > 1 year.
- pabs3 3y agoThe license is kind of irrelevant because its ML stuff that is impossible to retrain from the source data, because what that source data is, is a trade secret and even if it were publicly known, probably not under a FOSS license. The Debian Deep Learning Team's Machine Learning policy calls this sort of thing a "Toxic Candy" project. https://salsa.debian.org/deeplearning-team/ml-policy/ https://salsa.debian.org/deeplearning-team/ml-policy/ Of course people rarely need to retrain from scratch, but with some of the attacks on ML systems out there, that is an essential thing to be able to do if you can afford or crowdfund it.
- trwllm 3y agoSaying that ml models aren't open source because you don't have the data is like saying doom isn't open source because you didn't provide a proof the fast inverse square root works: https://en.m.wikipedia.org/wiki/Fast_inverse_square_root https://en.m.wikipedia.org/wiki/Fast_inverse_square_root LLMs are just very large magic numbers. That's all there is to them.
- pabs3 3y agoHalf of open source is licensing, the other half is source code (or data). It isn't open source without both of them.