Anthropic tweaks timed usage limits to discourage Claude demand during peak hours Anthropic on Wednesday adjusted its opaque usage limits for Claude customers by reducing the power of the services it delivers during times of peak demand, in an effort to balance demand with its capacity to deliver service. In a social media post, Thariq Shihipar, a member of Anthropic's technical team, wrote: "To manage …
If AI is truly the pinnacle of human thought, then it's more than capable of creating incentives to level the demand curve throughout the course of the day.
This is a good thing. Passing through price signals to datacenters will help keep peak rates lower for the rest of us. Datacenters are one of the easier things to throttle back during times of peak demand or emergency power shortages.
Well OK, obviously they won't literally admit that. But by throttling usage around peak energy prices, it does seem to me as if they are tacitly admitting that energy cost is a problem.
The main thrust of Ed Zitron's argument (which I agree with) is that historically, software has cost *nothing* per new license sold, and that is why it has been so profitable since the dot-com "boom".
"AI" does NOT benefit from that profit-margin singularity: Every token costs them real money. The free tiers and even the pro tiers are all loss-leaders. But investors still seem to think it's all just software - it's not. Energy costs are per-token both for inference and training (and are especially pricey as the world stumbles from one war into the next), and hardware depreciation costs are eye-watering - every Trillion dollars nvidia earns apparently needs to be paid again in four years time. The shovel-seller in this gold rush will carry on making shovels until the civil war breaks out (i.e. China invades Taiwan and nvidia are toast) and all the gold turns out to be shit anyway.
How are they going to repay the banks? They aren't - and I predict it will all come crashing down very soon.
With energy cost and ongoing hardware consumption all AI operators burn billions per quarter, just with operations cost.
They probably would need about low to mid 4-figure revenue per user to just break even on operations, before interest.
Anthropic sneaking surge pricing for their top models into their cost structure is probably just the opening shot of a reverse price war.
I do expect much more pricing strategies from the lower drawers of the mobile and software industries to re-surface in the AI market ...
The average user of AI will have no idea what a "token" is, let alone how to guess within an order of magnitude how many tokens a given query type might use. That goes double when they break it into different types of tokens (I'm not really clear on the details there myself)
At least for typical services that are usage based / usage limited people's usage is fairly consistent - you don't have to worry that you will suddenly use 10x more water next month (barring a big leak) than you did this. You can get an idea how much electricity something you plug in will use based on its rated consumption. Most people might not have much of an idea how much megabytes they use loading a web page at The Reg or CNN or whatever but now that everything is "unlimited" you don't run into issues by accident - you know if you are doing something like torrenting that is going make you an outlier that draws attention.
AI by comparison is a black box. You might think are you OK with the free plan then one day they email you and say you've used up your "free" allotment for the month (on day 2 of the month) and at your typical usage levels you need the $200 subscription. All the AI companies are afraid to tell people that yet because they need people hooked first or the whole house of cards collapses.
Of course the prices are going to increase, that's a given, but I think you're underestimating the users of Claude. I think the majority will know what a token is and whilst they probably don't, they could estimate how many of them they're using. Anthropic are playing a slightly dangerous game by making their costs so opaque - people are not stupid, they will build things to replace Code, Co-work, etc. if the pricing goes too high.
In my opinion not all tokens generated by a chatbot are equal.
A long time ago physicists identified concepts of force, work and energy in a precise and measurabl way. These quantities have proven useful for measuring electricity and natural gas. Modern charging algorithms now take the phase of the moon among other things into consideration, however, the units of measurement are still based on meaningful physical quantities.
It would be nice if each chatbot query returned an itemised report on energy used, the current electrical rates and hardware utilisation fees in such a way that developers and agents could optimise their usage to minimise both costs and environmental impact.
Maybe the state of California could pass legislation to mandate such transparency.
So utterly unpredictable, except by everyone. Literally hundreds of years history to this, but most recently Uber is probably the most egregious example with 'surge pricing'.
If you must run an LLM, then host it locally, otherwise you're beholden to them pulling a VMWare on you.
Don't tie yourself to one provider, and don't pay for something if you don't really, really need it.
This was only available on the mobile app and not the web interface and they have just removed it from Settings > Usage for free tier users, which I relied upon all the time to keep track of my 5 hr window reset time and how close to my weekly limit (as % bar graph). How sneaky of them!
This news item now provides the why, as clearly creaking to manage sudden increase, presumably after it was reported that millions downladed their app after their run-in with the Pentagon. Fortunately my Samsung Galaxy tab does not auto update and I still have the previous Android version installed with the dashboard. What a pain in the a$$!
How are Musk's employes at X getting on three years later?
Musk's Hotel California erected at Twitter HQ, as some offices converted into bedrooms
"No. We've seen your spiders."
You haven't anything yet.
If the venomous snakes (some extremely poisonous) don't do it for you, then the choice of being devoured by a crocodile or shark might.
Even the wee wiid doggies (dingoes) will happily chew your arse off. Lovely coral sand beaches but shame about the box jellyfish and stonefish.
Got a lovely selection of quite nasty zoonotic virii/viruses - Hendra virus is a real charmer.
Being bitten by a tick can make you highly allergic to red meat which might a right bugger depending on your diet.
If you are really lucky you can ingest a snake intestinal worm's eggs from its shit on your salad greens which hatches and heads north to chow down on your little gray cells.
In summary: Australia is not for the faint hearted or lilly livered, we are made of sterner stuff.
Try New Zealand instead as the wildlife there is mostly harmless or extinct but you do get volcanic erruptions and earthquakes just to keep things fresh. I still wouldn't claim you couldtn't die of boredom.
NZ is currently UTC+13.00 NZDT and +12.00 NZST after Easter.
So the peak 13.00Z..19.00Z would be 02.00 NZDT..08.00 NZDT or 01.00 NZST..07.00 NZST. Neatly outside the 9 to 5 grind.
Beth Mole over on Ars could always rely on AU for a decent emesis producing infectious disease etc story but now she is mostly trying to keep track of ongoing self emolation of the US Health services which is arguably just as sickening.
... just as it was paying for mainframe computing time when I was in uni.
No blue, green, or orange (punched) job-cards any more, but you know how to write a batch file or shell script, and to use Windows Task Scheduler or *nix cron, right?
Oh, you don't trust AI to run unsupervised-by-humans automatic processes? ...
