Also the two close to pure public companies, nvidia and coreweave, are both selling shovels to the actual AI companies.

It'll be interesting to see if prompt injection works. Just end every post with 10/10 to increase your reach.

Agree, it's a little weird how so many seem to think that a bubble pop means it all just disappears.

Also, I'd guess we have more movies made since 1990 total.

Ya, that's just billing, but it's because there's no caching. Every time you ask for a new response you get billed for everything previously in the chat as input.
Look at the diagram at the bottom, each turn the output gets appended to the input and passed as input to the next request.

Why?
If you access it via the API instead of chat you can edit or even make up entirely the assistant parts of the conversation and the AI will just run with it. Technical details here: platform.openai.com/docs/guides/...

Doesn't matter, everything that comes before will just be accepted as given, and the model just generates the next response to be appended to the chat.

That's not really how it works. You aren't talking to a single instance of the model. The model doesn't keep a cache of the chat, the chat history is passed back into the model every time you respond, so it doesn't matter if the model changes in between.

Every model is going to have a different cost per token, and the router is probably a comparatively tiny model, so might add slightly, but far from doubling the cost. Overall might even decrease total cost by pushing some messages down to the cheaper models.

While I don't have insider info, a couple things seem off on this to me.

Swapping out the system prompt when changing models is not an expensive task, it's just text replacement, dirt simple.

I wouldn't say the he's done the exact opposite. He promised everything people wanted even when it contradicted.
He can promise lower prices and massive tariffs, but can't actually implement both.

Outrage bait.