crazypython

It would be interesting to see a variant of this that used a small language model to measure entropy.

It lacks batching support (n>1), unfortunately, which is necessary for Loom-like applications

What are the tradeoffs between cpp2 and Carbon?

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.

"™" has no legal meaning. "(R)" means a registered trademark.

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.

How did you come up with the long-term memory idea?

Setting norms of asking folks not to abuse it reduces the likelihood people will step out of line with those norms.

> 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.

BigScience (a coalition including Huggingface) is training and releasing a 175B language model and finishes in 2 month.