Discussion about this post

User's avatar
Fukitol's avatar

Re: limits, well, start with physics. For solar panels there's a hard thermodynamic limit: at most, under 50% of solar energy received to usable electricity.

After that there's material science: no single photovoltaic material is black to the whole spectrum of the sun, so you mix and layer different materials that are alternately black and transparent to different parts of the spectrum. There's no perfect answer, and it's a box packing problem, which classical computing is bad at (and LLMs are still classical computing). Best you can do is smart guess and check, or fast heuristics but leaving efficiency on the table.

Then there's material availability: just because you can solve for optimal solar efficiency with novel materials doesn't mean you can get them at volume. Rare materials are costly to obtain and your solar panel has to deliver positive EROI.

Which leads to another problem: material concerns beyond solar efficiency. Your solar panels must be durable and functional in the target operating conditions. Here is some wiggle room if you can pump out bespoke designs and products on the fly instead of one size fits all, but still a limiting factor.

Result of all that is we're still using PVs that hit around 20-30% efficiency even though we know how to make better ones. Most cost per kwh reductions over the past couple decades have been China subsidizing their solar industry to eat the market, not engineering advancements. The physical limits were the problem, not intelligence.

Suppose we solve all that, with or without AI. Our solar panels are then about twice as good as the ones we have now. Great, but that's the end. No more improvements, ever, til the heat death of the universe.

Anyway, same basic problem everywhere else. Eventually a product is as good as it can be, and then... You have another problem: capital markets want line to go up, not flat. So a company with a perfect product and a saturated market has to raise prices and/or cut costs to maintain growth, which ironically degrades quality and efficiency makes products worse bargains over time.

So we're in this weird game where you need massive capex to do all these improvements, but promise a race to the bottom in margins, which reads: please give me money for free.

Gonna be a wild ride.

Marco's avatar

Is AI a game-changer for nuclear power as well?

41 more comments...

Ready for more?