Plausibly most of those lines of code don't exist now either, if people threw them away. And the others might not be any good. Or they might be things that already did exist—either because the AI generate them previously or because it memorized part of its training set.

I spent a lot of the morning talking to GPT-5o Mini about desiccants, passive solar collectors, and candidate approaches to 3-D printing of glass and ceramics, and it generated many pages of text, but most of those pages of text will get deleted without anyone else reading them; large parts of them are just wrong, and I'll need to check the non-wrong parts against the research literature and rewrite them from my own perspective so they don't sound like an impatient sales pitch.

It did give me some pretty good ideas, though:

- Nitrates (of magnesium, calcium, yttrium, lanthanum, etc.) are good precursors for metal oxides for bonding ceramics, and have special virtues for SHS.

- Zirconyl chloride is the usual water-soluble precursor for zirconia for this purpose.

- Titanium oxysulfate is the usual water-soluble precursor for titania for this purpose.

- Advection of supercritical steam through a crucible with salt may be a viable way to salt-glaze ceramics if you can mitigate the HCl problem.

- Acidification of an object molded from zirconia-filled waterglass may be able to leach out the alkali, making it possible to sinter the shape into a continuous zircon object.

- When acid-leaching iron out of a heap of crushed terra cotta, sulfuric acid has the problem that it can clog the heap with gypsum particles, if calcium is present.

- You can electrodeposit iron at an acidic pH as well as a basic pH.

Like, none of these are novel, right? But they were new to me, and they turn out to be correct.