Connor Leahy on EleutherAI, Replicating GPT-2/GPT-3, AI Risk and Alignment

It was interesting to start the TPU Podcast discord server. I was annoyed Connor showed up and convinced the talented researchers to spend their time on his server. But now I get to claim correctly that Eleuther never would have started without me, so I’ll put that up on the trophy shelf. Which happens to be next to Shockley’s trophy shelf, across the hall from MySpace.

It was a wild and fun ride. I ended up becoming a published, cited researcher thanks to Eleuther. https://twitter.com/theshawwn/status/1480479272078913539?s=2...

I’ve always wanted to do a postmortem writeup of my attempt at starting an open source AI lab. It was always my goal from the very moment I pressed the “create server” button, and it’s at least mildly interesting that it was successful — in spite of my efforts, not because of my efforts.

I do feel a pang of jealousy every time I see an article like this, and it would be nice if they’d occasionally link to our server (https://discord.com/invite/x52Xz3y) rather than mentioning it, but (with effort) I managed to transmute those feelings into amusement. It would have been better if I felt amused from day one.

I'd recommend reading the blog post about their first year, it's really fun and informative

https://blog.eleuther.ai/year-one/

If this is open source is there a good way to get involved?