As someone who used to run an infrastructure API service, I suspect that the weird pricing structure on Claude actually reflects their underlying costs.
Customers prefer it when you charge $X per Y API requests but it rarely works that way. You build your system to handle a particular peak load, and then your variable costs are relatively low. Sometimes you can auto scale to handle extra load cost effectively but usually that doesn’t work for a large service.
So the weird five hour windows are essentially making sure that you can’t pay for X number of API requests and then use them all at peak time.
Yeah, although you can’t just pick one fixed time to be cheapest because then too many people will want to run then. In the long run this ends up looking like AWS’s pricing scheme. It’s super complicated and you get massive discounts by either prebuying a fixed allocation or allowing your workload to be preempted.
Zvi, it was only a couple of days ago, and maybe you'd catch it anyway - but do check Vitalik Buterin on Liron Shapiro's "Doom Debates" for the next This Week In Audio. The mutual cruxes matter but are reassuringly few.
"No, wait, we’re not done, it can always get worse". It seems you don't realise that the post from Seb Krier is satire? The thing he describes is not another event that's worse than the thing above, it's a parody of the thing above. I agree the thing above is already stupid enough. The parody thing about stealing water from British children would be ludicrous beyond belief. Fortunately Chris Bryant never said any such thing - it's just made up. Perhaps you can move it into "The Lighter Side"....
Witnessed two separate government agencies' IT offices today look skeptically at the $1 OpenAI and Anthropic offerings while pointing users instead toward NIPRGPT and AskSage. ☹️
It's too bad they didn't let individual government employees sign up for it with $1 using an email address. A huge number would just pay it out of their own pocket, and you would get a real groundswell.
"So here's Margot Sloppy in a bubble bath to explain."
Video almost seems 'solved' looking at this. Damn. Surprisingly the audio is the weak point.
The accent from 0:09 to 0:16 completely switches! Australians do not say Daullerrrs (compare 0:30 which is correct). Did the word "US" in the text make it flip...?
I think the "I'm not a lawyer" refusals are amusing, because if you ask it for help getting a lawyer, it can't help but give advice on your case, even if asked not to.
I think the critique of the open source red teaming is a bit unfair. They did red-team it before release, but this is an _open_ red-team invitation. You can't really do that before release, or else anyone could sign up to red team and get access to the model, and most likely in that scenario it leaks immediately.
Why do you believe that steering via (persona)vectors will fail at the limit?
To be clear, I see many ways that it could go wrong or prove to be insufficient, especially if we are limited to steering surface-level personality traits. But in general I think either constraining or steering the "thinking process" itself is a much more promising approach to control, than trying to impose external guardrails or behavioral monitors.
> This all passes the smell test for me, and is a blackpill for Kimi K2. I’m not sure why DeepSeek’s v3 and r1 are left out.
DeepSeek R1 is there on that chart, a little over halfway down.
As someone who used to run an infrastructure API service, I suspect that the weird pricing structure on Claude actually reflects their underlying costs.
Customers prefer it when you charge $X per Y API requests but it rarely works that way. You build your system to handle a particular peak load, and then your variable costs are relatively low. Sometimes you can auto scale to handle extra load cost effectively but usually that doesn’t work for a large service.
So the weird five hour windows are essentially making sure that you can’t pay for X number of API requests and then use them all at peak time.
Can't they charge different prices at different times of the day then?
Yeah, although you can’t just pick one fixed time to be cheapest because then too many people will want to run then. In the long run this ends up looking like AWS’s pricing scheme. It’s super complicated and you get massive discounts by either prebuying a fixed allocation or allowing your workload to be preempted.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/ai-129-comically-unconstitutional
Zvi, it was only a couple of days ago, and maybe you'd catch it anyway - but do check Vitalik Buterin on Liron Shapiro's "Doom Debates" for the next This Week In Audio. The mutual cruxes matter but are reassuringly few.
"No, wait, we’re not done, it can always get worse". It seems you don't realise that the post from Seb Krier is satire? The thing he describes is not another event that's worse than the thing above, it's a parody of the thing above. I agree the thing above is already stupid enough. The parody thing about stealing water from British children would be ludicrous beyond belief. Fortunately Chris Bryant never said any such thing - it's just made up. Perhaps you can move it into "The Lighter Side"....
Witnessed two separate government agencies' IT offices today look skeptically at the $1 OpenAI and Anthropic offerings while pointing users instead toward NIPRGPT and AskSage. ☹️
It's too bad they didn't let individual government employees sign up for it with $1 using an email address. A huge number would just pay it out of their own pocket, and you would get a real groundswell.
"So here's Margot Sloppy in a bubble bath to explain."
Video almost seems 'solved' looking at this. Damn. Surprisingly the audio is the weak point.
The accent from 0:09 to 0:16 completely switches! Australians do not say Daullerrrs (compare 0:30 which is correct). Did the word "US" in the text make it flip...?
I think the "I'm not a lawyer" refusals are amusing, because if you ask it for help getting a lawyer, it can't help but give advice on your case, even if asked not to.
I think the critique of the open source red teaming is a bit unfair. They did red-team it before release, but this is an _open_ red-team invitation. You can't really do that before release, or else anyone could sign up to red team and get access to the model, and most likely in that scenario it leaks immediately.
Section 16 (Stop Deboosting Links) seems to be missing from the main body.
fwiw i believe the polyphasic sleep post was a joke/troll
Why do you believe that steering via (persona)vectors will fail at the limit?
To be clear, I see many ways that it could go wrong or prove to be insufficient, especially if we are limited to steering surface-level personality traits. But in general I think either constraining or steering the "thinking process" itself is a much more promising approach to control, than trying to impose external guardrails or behavioral monitors.