kromem

P.S. literally the comment you were initially responding to sources it.

You: "How much will this speed up our doom?"

Your waifu: "Calculating… oh, it's actually less than 1%."

tears in your eyes the Old Yeller soundtrack starts playing

Cornell researchers found that at the current rate of AI growth, the burgeoning industry could represent 24 to 44 million metric tons of carbon dioxide emissions by 2030

The United States emitted 4.9 billion tonnes of CO₂ in 2024.

So by 2030 the AI industry CO2 release might be 0.9% of total US emissions.

That 'almost' in "Almost Incomprehensible" is doing a lot of work there.

It won't, unless you normally talk almost exactly like the model in question and then alter your word choice distribution according to the specific secret key entropy.

It's not about the variation of the words, it's about the variation of the words from the model baseline.

Like if your word choice was almost the exact same as Claude's normally, maybe you just talked to them a lot and picked up their phrases like it's not nothing.

But if you managed to be almost exactly like Claude and yet varied the possible words exactly according to a hidden entropy key, they'd know it was actually Claude with the SymthID-Text watermarking applied, as no human would end up falling into that statistical bucket.

This particular watermarking would be effectively impossible for a person to end up replicating.

Probably less so.

The hardware to run it locally would be fairly expensive and would require using a very simple model compared to alternatives. Also much more wasteful if you were only using that hardware for AI use, as you'd be distributing the hardware out rather than centralizing so it'd be often idle and when replaced create more waste than a centralized server rack.

Also, additional per user post-training seems to me both wasteful and not necessary. In context learning is often much more powerful but frequently overlooked.

If you're concerned about data retention, you'd want to select an inference provider that is listing 'ZDR' (zero data retention) as a feature.

For example, a lot of the Chinese open weight models have become quite capable and because their weights are available end up like generic vs brand name medicine where there's multiple providers serving them with different production conditions.

Because enterprise use will often be worried about data retention or sending data to China, the alternative providers usually offer things like US-only inference or zero data retention.

If you're not going to use it all that often, a la carte API use is going to be way cheaper than a subscription, probably better results than a free plan with a closed model provider, and give you more control over the process.

consumed 40% of global water supplies. But now it is.

Source? This seems really, really unlikely. Last I saw total data center water use (only some of which is AI) was still much less than even things like golf course watering.

AI Data Centers’ Water Consumption Breaks 264 Billion Gallons in 2025 as Devastating Drought Hits Nearly 63% of U.S.

Golf courses reduce water usage by 31 percent according to national survey

The report found that U.S. golf facilities applied a projected 1.63 million acre-feet of water in 2024

(1 Acre-feet = 325,851.4286 gallons, so this is 531 billion gallons)

The golf course industry points out their use is less than 0.5% of the total water use of the US, but it's twice the use of all data centers.

So for just AI to be using 40% of all water seems… really unlikely?

Not that it's surprising you're under that impression, as even in the headline above, the gist many articles about this issue try to push is that the 264 billion gallons of AI water use is connected to a drought across the country. But the actual numbers reveal that as fairly manipulative as if we rewrite the headline to "golf courses use a little under 0.5% of total US water as drought grips the nation" it's pretty absurd to suggest the first thing is significantly impacting the second and yet we'd be representing twice the water use as the original headline is discussing.

What are you using it for during the 1%??

Right, but what % of people are currently using/demanding inference right now?

Do you expect that % to change between now and 2030?

Unless you expect demand to decrease, I don't really see how the pricing of the hardware will decrease.

Let's say the Pets.com of the AI world ends up going bankrupt and their RAM hits the market. Do you expect that the demand for that RAM will be negligible such that pricing returns to earlier levels?

Your predictive model relies on companies that have hardware going out of business and then other people buying up that hardware, but isn't accounting for the levels of demand that the market will have for that secondhand hardware even if it ends up existing from failed firms.

Unless the demand shifts, the more likely scenario is that companies going out of business will be able to sell off their RAM at higher prices than they bought it at.

There'd need to be a significant inference memory reduction advance (possible) coupled with stagnating or reduced inference demand (unlikely) to see prices come back down.

Wait… how do you imagine a world where there's demand for frontier grade AI but also that the bubble has popped such that there's not demand for the chips to run frontier grade AI?

I'm really confused.