Is there evidence for the claim that the training was that much cheaper? Or is is 'apparent' by inspection of the verified ingestion architecture?
I can't seem to find any evience beyond the team's statements. What have I missed?
I can't seem to find any evience beyond the team's statements. What have I missed?
Compute resources they officially should not access to given export bans, where mentioning them might lead to their export ban bypass getting rolled up.
Hypothetical: Take a large short position on NVDA, announce the market that you trained a massive model without using 10s of millions of rare-as-hen's-teeth NVDA-sentsitive resources. Settle postition then quietly settle giant compute bill. Difficult to know either way, but the market seems have taken the team at face value. I guess we'll know if and only if this reduced training cost methodology is replicated.