"time to GPT-2", down to 2.91 hours | Hacker News Reader

"time to GPT-2", down to 2.91 hours

twitter.com

80 pointstosh15 comments

https://nitter.net/karpathy/status/2018804068874064198

Never used gpu spot instances before but I would have to imagine getting interrupted is pretty annoying.

it depends, our workloads can finish up in under two minutes and shut down without much effort, so we haven’t really noticed it outside of one time when we had no spot capacity.

I guess if checkpointing is set up correctly and your runtime is saved to a docker image it’s feasible. Probably not going to get a 3 hour continuous chunk of time I would assume.

When I once used Spot it wasn't that bad. You're likely to have an instance for 3 hours.

I use spot for my gmod server. Contention is higher for GPU, but I only saw disruptions maybe once every 2 months for super common instance types. We scale up and down more frequently than we get disrupted.

In addition to the two-minute interruption notice, rebalance recommendations[0] allow you to handle interruptions even more gracefully.

[0] https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/rebalanc...

As long as your workload can handle resuming again and your instances aren't heavily in-demand (looking at the eviction rates), the cost saving for us is substantial enough to take the occasional interruption.

I do wish Azure gave more than the 30 second eviction warning (like AWS) but still useable.

My favorite GPT-2 experience was setting homepages to the now-depricated website which used a 2021 LLM to generate new words (this word does not exist.com).

Its definitions didn't always make sense [2], but I kept a curated list of "my favorite computer-generated words" (and recommended to crossword/scrabble -type friends to set the same homepage).

Favorites were sent via postcards, or shared over coffee.

Mom, you were a real nayter [1] — shall miss our decades of scalpable [Z] glossip [0].

----

[1] nayter (v.): to present one's viewpoint as true rather than having exaggerated or undeserved moral beliefs

[2] ...the second definition was always static text: "a word that does not exist; it was invented, defined and used by a machine learning algorithm" with link to what GPT-2 was.

[0] glossip (n.): a term of abuse attached to words used to discredit the person or group with which the author of a story is confiding

[Z] scalpable (adj.): able to be moved with considerable ease and without needing to be lifted or deformed

[bonus] dumbfuck (adj.): (especially of an electric guitar) having a fixed flat fretboard or tuned tuner

> GPT-2 (7 years ago): too dangerous to release.

With the benefit of hindsight you can see the the charitable foundation to benefit mankind was a grift all along.

I'm not sure I agree. LLM's have the feel of an alien new technology, and especially did back then. In retrospect, it feels very obvious that small models don't pose much of a threat, but that's only in retrospect.

Nobody else should be allowed to build these ... while we build another model that is 10x more capable than the one that was a threat to humanity.

The thing about dangerous AI is that AI will by definition not be dangerous precisely until it is.

This isn't a good reason to be reckless.

OpenAI’s transformation—from a nonprofit founded to serve humanity, to a for-profit prioritizing enterprise (B2B) revenue at the expense of everyday users—isn’t simply a shift in business model; it represents a fundamental abandonment of its original mission and of public trust.

This is a fun, new speedrunning genre.

Reply on news.ycombinator.com