Must be officially released by OpenAI so leaks or off-hand comments do not count.
Extremely wide intervals("between 1000 and 100 trillion") would not be accepted: Endpoints must be within 25% of the mean. So "500-600 billion" would be acceptable because the mean is 550 billion, the 25% variation is [412, 687], and the given interval is contained within it.
Related questions
Plausible-sounding leak: https://twitter.com/Yampeleg/status/1678547812177330180
Based on paywalled content here: https://www.semianalysis.com/p/gpt-4-architecture-infrastructure
Edit: Tweet was taken down due to copyright takedown request by SemiAnalysis. Archived: https://archive.is/2RQ8X
https://twitter.com/soumithchintala/status/1671267150101721090?s=20
Speculation spread on twitter yesterday
I believe the movement in the market since yesterday is from speculation that spread over twitter ( I learned of it this morning from Mastodon ) : https://twitter.com/soumithchintala/status/1671267150101721090?s=20
Given the shrinking of the moat by the very enthusiastic and competitive environment and their stated past concerns about this in 2023 so far, I don't see this environment changing even if there were new regulations concerning LLM training and so I believe this will be kept a trade secret at least until model size for LLMs truly becomes irrelevant due to reduction in HW costs or efficiency gains from software development, which I put past the end of 2024 (we're only at the starting line of the singularity). Oops meant to buy NO. Doh!