Nafnlaus 🇮🇸 🇺🇦

(Premodern, they relied heavily on cisterns and raincatches, but that cannot sustain modern populations)

The city of Qeshm on the island was ~40k in 2016, but only ~7k in 1976.

The plant was opened in 2013 and went through several expansions, but the population was already skyrocketing then. Main water supply was a fossil aquifer (Tourian) but it was depleting & seawater intrusion, hence the plant.

"Was the population of the island much lower before the plant"

Yes, it was. The whole region is propped up by modern trade. It was only 4-6k people in the early 20th century. And no, there is no meaningful agriculture. Some dates, but they don't need irrigation - and not much anyway. It's a desert.

People pay the huge premium because they want the best models out there. But how good the model is doesn't change its *cost to serve*, and thus how profitable it is.

A typical (competitive market) server serving said model will charge say $0,10/Mtok for input and $0,30/Mtok and make a mint on it.

OpenAI, on their *mini* model, charges $0,25/Mtok for input and $2,00/Mtok for output.

The big companies charge a massive premium on inference.

power prices, that's ~$0,30. The 5090 is $2k, and if we assume a 2 year lifespan (in reality, far longer), that's $0,40. Total cost $0,70/Mtok. In reality, a server farm will do it for WAY cheaper, not only because of the above, but also because they can do batching, which hugely boosts performance

Just to elaborate a bit more on inference serving costs: A RTX 5090 - we're talking home hardware, not highly optimized server hardware - will get 60-100 tokens per second on a medium-sized model. A million tokens every 3,5h. During that time it burns 2kWh. At retail (not cheap industrial)...

They're dumping money en masse into datacentres and R&D because (A) their usage demands are skyrocketing, and (B) they're racing in a highly competitive field against competitors (Anthropic, Google, Grok, many Chinese companies, etc).

searches in the background to gather current info before they answer.

2) Finetunes are *the easy stuff*. It's *building foundation models* that are expensive. You can finetune a model of the size in question for a couple tens of thousands. It's pocket change for a company.

"Continuous training cannot end".

There are two common misconceptions here.

1) Models are NEVER up to date. The latest Gemini models for example have cutoff dates in early 2025. Which is fine, because they don't rely on their internal knowledge - they use RAG. TL/DR, they do a ton of web...

"They are losing money on inference alone"

This is simply false. They charge an order of magnitude more per token for equivalent-sized models than the highly competitive and open market of serving inference on open models. And I myself run inference at home. It's dirt cheap. Period.

While I see no signs of any plans to occupy Qeshm at present (and haven't even seen verification that the strike occurred as described)... destroying the desalination plant, while illegal, would be an effective way to get most of the population off of the island.

If you're trying to open up shipping in the Strait of Hormuz, you have the following to deal with:

* A city of 1/2M people (Bandar Abbas)
* An island of ~150k people (Qeshm)
* A bunch of small, low/no civilian pop. islands
* A bunch of desert, with villages here and there.

Been thinking about that Qeshm strike. Beyond the options that first jump to mind:

* Never happened (PR)
* Miss
* Targeting military infrastructure
* Punishment (destroying expensive infrastructure)

... something else occurs to me:

If you wanted to evacuate Qeshm, you'd strike the water sys.

Serving inference is extremely cheap, and OpenAI's markup is massive.

You're confusing "losing money because they're spending insane amounts of money on R&D and expansion" with "losing money on serving inference". They would be quite profitable if they ditched the former.

Precisely. Which is why you must change the reward/penalty structure to daily tests, so that they *have to* study. Grading "homework" is a pointless task; you're grading an AI.

Also, point of note, it was the Spartans who were obsessed with luxury pastimes (equestrian sports, fashion, etc), and mocked Persian leaders for doing "physical work" like gardening.

"LLMs are nothing more than models of the distribution of the word forms in their training data"

LLMs are not Markov models. Try again.

There is a massive false positive rate for "catching" people, only dwarfed by the even larger false negative rate, so enjoy your terror-state.

bsky.app/profile/nafn...

bsky.app/profile/nafn...

*They're* embracing it. *Now*. Because it does their homework for them. So what's your *actual* plan to make sure they learn?

And they'll keep cheating, and the educational system collapses.

Brilliant plan. A++.

Interesting "solution" to a problem where the challenge is detecting whether people are using it. People don't shout, "Hey, I'm cheating!"

You need to deal with the fact that it exists, and students are using it en masse because it profoundly eliminates work that they don't want to do.

What's your proposal?

The current situation, where students just punch their homework into ChatGPT, is not sustainable. That's not conducive to learning.

Homework, as we know it, is a concept that just doesn't work anymore.

Still burning fuel? Lol.

Picture of a rocket with the headline "The EU atom bomb". The rocket has a flag of the EU and a round radioactive sign with three black parts on a yellow background. And a warning sign like on cigarette packs with a mushroom cloud on it. At the bottom is the exhaust fire. Around the rocket are the following texts with a bar pointing towards the rocket: Does not contain parts that can be swallowed by children under 3 years 100% recycled parts Shock picture and health warning Can only be activated when *all member states turn a key Best before date must be clearly visible The lid can't be removed completely Parts come from different member states by a fixed scheme Must be connected by USB-C to airplane or launch pad Propulsion system CO₂ neutral according to EU law 17/A Below a frame with more text titled "Additional Features:" * Automatically erects a sign at ground zero with the text, "This crater was funded by monies from the Nuclear Bomb Fund of the EU * 8 times more expensive than the standard price * For security reasons Hungary only gets a mock-up

😂
Hier eine übersetzte Variante:

Anything can look dramatic if you colour it in technicolour. I'll repeat: it's roughly Nevada-esque. explore it yourself:

earth.google.com/web/@27.2598...

But yes, it is well more rugged than Iraq, which is largely quite flat.

(It's also, BTW, beautiful. So many amazing natural wonders there)