The GPT-3.0 "davinci-instruct-beta" models have been returning non-deterministic logprobs as early as early 2021. This is speculation. CUDA itself often has nondeterminism bugs.
text-davinci-001 and text-davinci-002 were trained through FeedMe and SFT, while text-davinci-003 was RLHF; the models themselves have more variance at high temperature.
S3 is a highly reliable storage media. You should use a CDN that transparently caches S3 artifacts to save on reads, such as AWS CloudFront or Cloudflare R2 are good choices.
Definitely worth getting excited about, though it seems the current cost of fuel production as of Aug 8 including crude oil cost is $3.40/gal minus taxes and distribution.[0] Prometheus will likely cost less to distribute, since it doesn't need to be shipped across seas, funneled through pipelines, and between refineries, it can be produced next to solar or wind generation.
For reference, Prometheus costs $1.86/gal to operate and they are aiming to reduce the cost of the machine so it can be produced including capex at $3.00/gal.
> More generally, a job that requires fast and unexpected changes, significant skill, and is not very repetitive will be one a hacker will excel at. Repetitive, simple jobs are a waste of a good hacker, and will make your hacker bored and frustrated. No one works well bored and frustrated.
> The good news is, if you get a hacker on something he particularly likes, you will frequently see performance on the order of five to ten times what a "normal" worker would produce.
I agree with this post, though there's another factor to consider: SSD endurance. A system that swaps a lot will write more, decreasing endurance. You might need to buy a new SSD every 5 years.
under the free software philosophy (free SOFTWARE), art assets are allowed to be non-free, but stallman believes the game code itself should be free. considering games rely heavily on art, this model still lets you sell games.