The spec I never checked

I spent most of last week reading about hardware for running a big open model on my own machine. Memory bandwidth, quantisation tiers, how much quality you give up dropping to two bits. I had it narrowed to two routes and was mostly arguing with myself about which one I'd regret less.

Then I opened the actual configurator to price one of them, and the memory options I'd been planning around were gone. Not sold out. Pulled. The top tier had disappeared some months back and a second wave took most of what was left, so the biggest machine you can buy new now is a fraction of what my notes assumed.

A week of reading, and the one thing I never did was check whether the thing was still on sale.

What annoys me is that the research itself was mostly fine. The numbers held up. A few came from sites that clearly get paid by the word, and I'd already marked those as thin. The failure sat one layer underneath all of it: I treated a product page I half remembered as a fixed fact, and every conclusion above it inherited the error. My process asks "is this still true?" constantly about the interesting claims. It never asks it about the boring ones.

So the shortlist is different now, and worse. Every route to the tier I wanted runs through the second-hand market or through building a server, which is a different sort of project than buying a computer. I haven't decided. I keep opening the note, reading the same three options, and closing it again.

That's probably the honest state of it. A shorter list, and a rule I should have had already.