So Gemini was the most non-deterministic model of them all and now we get this one with temperature at 1 and max thinking. It’s so random that it’s hard to justify putting in my setup right now.

> the internet in 1988

60k computers ( mostly at institutions ) in 20 countries

Everything was slower though. Turkey as a whole country had one 9600bps link to Bitnet at the time. Internet was accessed through Bitnet gateways. Systems (CPUs and I/O in general) were also much slower.

Slower and unstable. I spent a lot of my freshman year in college on Bitnet chat and iirc about every 30 minutes there would be a "netsplit" and a bunch of folks in the chat would disappear. Maybe it was our universities connection, which I think was direct to UIUC. I've posted here before that back then I thought Bitnet chat was magical. Things like being in a chat room with students in Berlin while the wall was falling felt so futuristic to me.

Much slower. Most campuses in the US were connected with 56K dedicated lines. The NSF backbone had just upgraded to T1.

ftp.wustl.edu would manage about 1 KBps and I was sitting one hop away from it at UIUC.

Insomnia paid off a lot back then.

Probably relates to some of the political controversies surrounding the source NPR here: https://grokipedia.com/page/NPR_controversies

I am both bemused and disappointed that the top comments are all political. Kind of plays into what Billionaires like Bezos would want - the commoners bickering and not united.

I don’t think I ever could do all the 25 so this is fine with me.

Arthur’s classifier will only be as accurate as their retrieval. The approach depends on the candidates to be the correct ones for classification to work.

OP here. This is true. If you make your min_score .99 you can have very high confidence in copy-pasting the label, but then this is not very useful. The big question is then how far can you get from 0.99 while still having satisfying results?

Thanks for the article and approach. How did you come up with min_score at the end? Was it by trial and error?

I don’t get it. Why no American accent?

Suggest adding 2024 to the title

My experience was while it had longer context it didn’t do as well on coding tasks. I use CC after being on Cursor.

Same with me, I had signed up for a year plan on Cursor, used it for like... a week, then tried CC and it was so much better I signed up for the max plan and then never looked back.

Whatever Anthropic is doing, it’s working amazingly well.

And codex seems to be catching up. But gemini-cli today was really bad on an elixir edit. I’m not sure what causes it to be that bad.