Hacker News Reader

Or the other reams of world class machine learning research and model weights they've released freely.

perhaps this is whoosh, but I thought Google wasn't releasing their model weights or source code, at least for a lot of things.

like with WaveNet, it was easy to train but computationally infeasible to run the way they published the paper. then it turns out there's a trick that makes playing it forward far more efficient, which Google later acknowledged using internally, but didn't disclose.

What kind of trick? At a glance wavenet looks like you could reuse most of the computation from earlier timesteps (would gave to take a closer look to see if it'd actually work), but I assume you're talking about something less obvious?

They don't always release weights and often the code and weights are delayed from the paper.

But more often than not they release a lot of their weights and/or research code.

Reply on news.ycombinator.com