Yeah, but why AI-slop?

They want to cash out before the bubble implodes I guess. Let’s transfer the monopoly money to pension funds in time.

Where do you have the 3000 L of water for 1 kg almonds from? It is not in that paper. I read about 10000-15000L per kg almonds (here: https://berlin.nabu.de/umwelt-und-ressourcen/oekologisch-leben/essen-und-trinken/32632.html) but that was admittedly not a scientific publication.

Just because datacenters are terrible with questionable use, doesn’t mean that combustion engine cars aren’t really bad either. Especially in urban areas traffic is causing a massive decrease in air quality and as a consequence are certainly the cause for respiratorial issues. It can’t happen soon enough to relegate these cars into museums. But of course, EVs are just solving the engine emission problem, they don’t solvbe all the other issues cars bring with them.

According to the numbers I found above, almond is pretty comparable to meat production in terms of water intensity. The only thing saving it is that there is a lot less almond production than meat production.

If I look at German sources almonds and beef/pork have the following overall water demand for 1 kg product:

Almonds: 10000-15000 L/kg, beef: 15400 L/kg, pork needs apparently half of beef

That does not necessarily sound like “less than meat”. Granted almonds are among the most water consuming plants. Potatoes need only 290 L/kg.

https://berlin.nabu.de/umwelt-und-ressourcen/oekologisch-leben/essen-und-trinken/32632.html https://www.landwirtschaft.de/umwelt/natur/wasser/wasserfussabdruck-wie-viel-wasser-steckt-in-landwirtschaftlichen-produkten

Those aren’t full prices though. You are aware of that, aren’t you? Nor are the current “full API” prices necessarily truly enough to cover full costs. It is not unlikely that even those are far away from profitability. If they weren’t, why all the financial gymnastics?

The real costs will only be known after the bubble bursts and the venture capital billions are going to dry up and the whole circular financing schemes are falling apart that are massively distorting numbers.

If the current subsidised rates are a good deal for you, nothing wrong with that. Just don’t make decisions that are binding you to what is doomed to explode in prices in the foreseeable future. Also, if all that productive is going through the roof with AI, why is software generally getting worse and buggier. Are all those big tech companies getting suddenly more incompetent, just when they are all moving to processes that are heavily using LLMs?

This reads like a list of “Teslaisms”. Many big car producers have figured out to ditch them again, after the hype. VW is far from the only but a good example. Have a look at the upcoming ID.Polo. Or look elsewhere, Renault R5 or Twingo for example. And it is not just Europeans, Hyundai Inster is similar in this regard.

The AI bubble can’t burst soon enough but can anyone tell me how it is even possibly to burn that much money, at current token rates in a single month? How much inference would that be? Was that a huge customer with a lot of users or did they go bonkers with “claw”-stuff to exponentially increase their capacity to burn money in a bonfire?

The article sucks. That doesn’t change that most companies are still not even remotely cost controlling LLM usage like pretty much anyting else, nor that we even know what real costs, without billions in venture capital subsidies, would look like, other than substantially higher than now.

“stochastic slot machine” What a great description. I have to remember that one. :)

I think you did not get the analogy. The point was that there was a canal bubble in the US which was mostly a huge waste of effort because the technology was outdated before a network could be completed to turn useless parts into a useful network. This was neither the case for rail or road networks. Both technologies experienced bubbles and overconstruction but they yielded large networks for generations to come.

Tokens are what an LLM actually predicts, one after another. If you have very slow models that produce less than 5 tokens/s or so, you can easily follow it with your own eyes. A token is what appears at once. Often it is an entire word but it can also be parts of a word or individual letters, digits, special signs for uncommone words or special formatting, number stuff.

Makes me wonder how the Chinese and European power grids can handle it just fine so far, with Norway being not that far away from full transition for example.

If the network is a limiting factor, why not push for dynamic pricing in combinatio with wall boxes that automatically use targetted loading during better prices. If that is implemention on a large scale, they would stabilise the network rather than destabilising it.

Curious, back in the early days they did not have to force young employees to embrace the web. Now with AI they have to.

Source? Has it progressed so much? I remember substantial double digit percentages.

I know, US style capitalism doesn’t put much value in it but loyal employees can be a win for a company, but you only get them if the companies are rewarding that loyalty. The cost of having disposable employees is often extremely high. In the worst case not only a lot of informal knowledge and skill is lost but sooner or later it will end up with the competition, no matter the clauses in the contract.

Currently, yes. However, if you build your business on highly subsidised temporary prices, you are probably like the CEOs that can’t think further than their shareholders. If you believe there will be flat rates in the end I have a bridge to sell to you.

You might have a point but I’d like to add, that this view is not just shared by some random 3D printing advocate. It is shared by the SFC as well and they should know their license, I would think. Anyhow, Bambulab is now under their watch. Also the SFC is launching a reverse engineering project on Bambulab AGPL3 software. If Bambulab doesn’t like it, it can try to sue the SFC. Good luck.

https://sfconservancy.org/news/2026/may/18/bambu-studio-3d-printer-agpl-violation-response/

I switched las month from GPT-OSS-120B to Gemma4 31B. For some simple scripts I found the latter considerably mor efficient, less verbose and with better results. At the same time the sycophancy is much worse.

No way I’ll use cloud based models and feed the data base. But also local models clearly like to burn energy.