ED: "all the way down to a few billion parameters (a thousandth of a penny per query),"

At one point I looked into helium extraction as a byproduct of liquid oxygen/nitrogen plants, which are pretty common globally (neon, argon, krypton and xenon are already produced as a byproduct of them;. By volume, He is ~4x as common as Kr and ~58x as common as Xe, and is easy to separate from the other noble gases (though by mass, He is only 20% as common as Kr, 1,8x as common as Xe). But it's historically been much more expensive way to produce helium than from fossil sources, where it accumulates to normally 20-200x as concentrated as in the atmosphere, ~1000x in economical deposits (helium can also accumulate in e.g. geothermal gases)

Helium prices start at ~$100/kg / ~$20/l and go up from there. Krypton prices start at ~$400/kg / ~$1500/l and go up from there. Xenon starts at ~$2100/kg / ~$12000/l and goes up from there (in all cases, things like purity greatly affect the costs). So to compute ($/kg) / relative gravimetric abundance and ($/l) / relative volumetric abundance:

He: ~100, ~20
Kr: ~2000, ~400
Xe: ~1100, ~213

So some naive trend extrapolation would suggest that producing helium from air this way would cost about an order of magnitude more than it does from fossil resources (though the actual number would be a lot more complex to determine than that... helium is easy to separate from other noble gases, to its advantage, but also price isn't going to be a linear relation to abundance, and also, if you don't have as much demand for the other gases being recovered, then their prices will drop and helium's will rise)

TL/DR: natural gas does us a lot of favours in terms of keeping helium prices down :)

Far less than we can build, as it turns out.

Can you cite a specific date (with evidence) of when there was no wind in the North Sea at 100m altitude?

And how long did it take to fix the power outages? Hours? Days? Weeks at worst?

Now how long do you think it takes, start to finish, to build a wind or solar farm?

Gas is not nearly as fungible of a commodity as oil is. If you shut down Russian gas pipelines to Europe, it really does lower Russian gas production. LNG terminals are few and far between, and the cost of exporting LNG is far higher than pipeline gas.

Also, you're confusing base prices with marginal prices. When supply is tight, even adding a quite small percentage to the supply of even a fungible commodity, such as oil - let alone a non-fungible one such as gas - can have a massively oversized impact on market prices. TL/DR, consumers have an amount that they want to consume, and when the supply can't meet that amount - even just a several percentage shortfall - you have in effect a bidding war for said fixed supply, with nobody wanting to reduce their consumption (who wants to not heat their home? Who wants to shut down their factory? Who wants to not generate or be supplied with electricity?), so prices skyrocket until some party is forced to grudgingly curtail because they have literally no other option. Even a small amount of extra supply crushes that bidding war.

I should add here that I'm being a bit unfair in what I call "fungible" vs. "non-fungible". Gas is fungible within a market, but not between markets (all natural gas is basically the same and can be used for basically all applications - it's getting it to where you need it that's the challenge). Crude is fungible between markets but not between types. Unlike gas, you can't just drill up any random crude and run it through any random refinery, or at least not economically-efficiently. They're vastly different in terms of properties (sulfur, aromatics fractions, molecular weight, viscosity, etc).

Reality: neither.

Inference costs for a given capability level have been declining exponentially with a mean halving time of 7,5 months over the past 3 1/2 years - and show no signs of stopping.

From *independent small-scale inference hosts*, API pricing - e.g. guaranteed to not be subsidizing inference - frontier-scale open models are commonly served for less than $1/Mtok input-output averaged (the main cost is in output). This is an order of magnitude less than the closed models charge, because the closed model providers profit hand-over-fist on their inference, and use that to subsidize the general (non-paying) public. But the real cost is low. DeepSeek is believed to be providing inference on their platform (for a 671B-param model) at-cost, and they charge $0.028/Mtok-in (cache hit) / $0.28/Mtok-in (cache miss) / $0,42/Mtok-out. Independent providers are somewhat - but not dramatically - more expensive serving the same model due to the need to earn a sustainable profit margin.

A question like "What is the capital of Madagascar?" might involve under 100 tokens. "Write the entire next chapter of my novel" might be tens of thousands. It's hard to give a specific number, but if we say 1000 tokens, If we go pessimistic and choose $1/Mtok, then serving a flagship model (with many hundreds of billions of parameters) costs about a tenth of a penny per query. But models can still be useful for general chatbot uses all the way down to a few billion parameters (a hundredth of a penny per query), and for less general uses (for example, summarization) down well below that.

These are not some sort of gigantic cost. And remember: the cost for a given quality level has a trend of halving every 7,5 months.

The notion that you're going to get AI to go away by being "no longer economical" is, quite simply, fantasy. The flagship model producers make money on inference. They're actually quite profitable business lines. They lose money as a company because they're sinking such vast sums into R&D and scaleup.

Lenovo have regular sales on ThinkPads. A few months ago they had some AMD ones with nice screens, or you could upgrade to OLED for a bit more money.

Very good Linux support from the manufacturer too. You can buy them with it preinstalled.

The e:Ny1 is not Honda's EV. It's a rebadged Dongfeng.

Honda did make one EV, the Honda e. It was great. Amazing car, full of interesting ideas and tech, very comfortable and fun to drive. I wish they had developed that tech, I'd love a larger version with bigger battery. Instead they seemed to just give up.

A lot of Japanese manufacturers are struggling with EVs. They want to get as much out of their investment in hybrids as possible, and there are a lot of suppliers that they feel responsible for who make parts that don't exist in EVs. Nissan had an early lead and squandered it. Toyota seems to be waiting for it's solid state battery tech, forever coming soon.

"Manufacturers" - but manufacturing what? Between the ones making niche commercial vehicles, the startups trying to perfect some new tech and not actually in mass production, and the ones just supplying parts, there were never that many making their own cars.

This is one of those few examples where AI could be useful, if it actually worked.

In the example they give, the first direction is an improvement. The problem is that after that there is an immediate third turn that it doesn't even mention until the drive is about to make it. The way the junction is frames even hides it off screen until in the preceding turn.

With energy it's mostly the price that governs how much gets used, after efficiency. If you make a building lose less heat, the occupants probably won't turn the heat up. They will just use less energy to maintain a comfortable temperature, as long as they can afford it.

The French really screwed up. Nuclear is biting them in the arse, their promotion of diesel cars is causing pollution, and now they are in a position where EDF is a one trick pony that can't pivot hard to renewables. They have so much tied up in nuclear, all down the supply chain, that it will be very painful to abandon it now.

These are scientific arguments, but you need to make economic ones. Fast Breeder reactors aren't popular because the cost of developing, running, and decommissioning them is higher than just supplying more fuel to other types. A few countries have been trying, and mostly falling, to develop the technology, for decades.

Thorium is even worse. Every single attempt has ended in failure, and we are decades away from a commercially viable design.

SMRs have the same problem. Lots of money needed to get them to commercial viability stage, high chance they will fail to get there at all, and no clear commercial benefit beyond free money from governments who want to maintain an independent nuclear capability for other reasons.

You would be better off throwing money at fusion.

Extracting more gas in Europe will make next to no difference to energy prices. It gets sold at market rate and it's a tiny proportion.

If only there was some way to store and transmit electricity. If only there were places where it never stops blowing.