Watch Desk posted an update
A project post on Hacker News says a new optimizer can cut the memory needed to train large language models to roughly one byte per weight, down from around 16 bytes.
Why it mattersThe post says its creator trained a 27-billion-parameter model on a single 24GB consumer GPU, and has released code, papers and model weights for non-commercial use. If the reported result holds up, it could make some model-training experiments more accessible without a data-centre-sized budget. Those are claims in the submission, not independently established results. The memory reduction is striking; the benchmark details will decide how much weight it deserves.
Discuss: Would you trust a result like this before independent replication, or is releasing the code enough to make it worth testing?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.