@GaryMarcus ...which is sus because the initial release could simply have artificially *juiced* it just before IPO to help their numbers a bit more. But if so, it seems like it wasn't near enough.
Ramp data comes out Wednesday, which is more indicative of actual spend. Should be interesting
@GaryMarcus The OpenRouter crew is fickle. You'll see models take over the entire usage share, then they drop off weeks later. Whatever is hot and new takes off. Opus 5.5 was a really impressive model initially before they apparently nerfed it (another major problem in this industry),
@AnsenKim@edzitron It's more like this movie I saw, about a bus that had to speed around the city,--keeping its speed over 50--and if its speed dropped, it would explode.
I think it was called "The Bus That Couldn't Slow Down".
@Ringsofproviden@arakharazian It was. That's why Q3 growth is slowing, and @arakharazian's projection of Q4 growth shows even more slowing. Usage of Claude models is expanding even as tokenmaxxing contracts--it's offsetting their growth. God help them if a better cheaper model from Goog/Meta lands.
@SenSanders Private Credit is using life insurance and annuity premium money to fund AI. When insurers fail the taxpayers of every state MUST bail them out--that's the law in all states. State regulators need to take action NOW, esp VERMONT, @BernieSanders !
@StocksNBaseball@arakharazian Could be shenanigans, not ruling it out...just saying it's not from AWS, specifically. I've heard good things about Claude Opus 5.5 - maybe they took some Fable magic and moved it to their lower model so they get a usage bump without cutting prices.
@StocksNBaseball@arakharazian I pulled the OpenRouter data for tokens processed by all models from August and beyond. On a relative basis, the patterns for all providers for the ANT models show a similar pattern. So I don't think this looks like shenanigans from a single, or set of providers.
@weeklyshowpod@edzitron@weeklyshowpod , this was a FANTASTIC interview, happy to see @edzitron's fine work elevated.
Request--PLEASE continue pulling on the insurance / pension thread here. There are some severe consequences if the AI bet goes south. @TomGober5 and @rdubitsk are the experts here.
@HomoSapienLCY@StocksNBaseball@arakharazian I wonder if Ramp can see whether these new tokens were generated by a small number of customer accounts, or if there was a larger distribution that might indicate more organic growth. You'd think so.
@ravenloft43@arakharazian July was peak tokenmaxxing among large tech companies. Partly from FOMO, partly from a desire to learn how effective, partly to capture their own prompt and trace data to build their own models. But there were zero spending caps. Everyone is now reducing spend.
@arakharazian Thank you for sharing! Can I make a request? On the link for "Get the Data", could we have it in vertical form (date, provider name, model name, model version, measure) please? Would make it easier to drop into various analytical tools if we want to create our own charts. TY!
@arakharazian@chamath@theallinpod@arakharazian - Is there a reason Ramp doesn't release the weekly token usage/spend data on a weekly cadence, rather than bundling it up in a monthly release with the adoption data? Just curious. Thanks for all you've been doing to share data!