site banner

Friday Fun Thread for March 3, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

4
Jump in the discussion.

No email address required.

I suspect it, too, is a gimped/lossily accelerated version, with the full one to be released under the «DV» brand as part of their Foundry product. I may be wrong, however, and ultimately this PR-driven nomenclature is making less and less sense, they have trained like a dozen major models at this point, and more finetunes. Same story with Google's PaLM.

More importantly, quantity has a quality all of its own. DV offers a context window of 32k tokens. If the model can be cheaply ran with such enormous contexts, this will more than compensate for some intellectual deficiency in the context-sparse text-prediction mode. You have seen effects of StableDiffusion mega-detailed prompting, and of prompt prefixes that amount to a single page – now imagine 20 pages of precise specification, few-shot examples, chain-of-thought induction, explicit caveats for every failure mode discovered in testing; basically a full employee's manual. Writing and testing these megaprompts may become a new job for ex-middle managers who have suddenly transformed into leaf nodes of the org chart – for a short while.