None of the flaws in this talk had anything to do with bad entropy. They were all firmware flaws (and in the Trezor case they managed to glitch it to get the seed -- so generating the seed yourself adds no extra security and might actually reduce it if you generate a bad seed).

I'm sure bad entropy with hardware wallets is a real problem, but clearly there are far more practical problems.

Try computing ECDSA by hand and we'll talk.

Instead just generate the word list, which is what you need anyway to put your handmade key into the hardware wallet. Pick the words by some truly random method, like Von Neumann coin flips or picking words out of a hat. The last word is (partly) an 8-bit checksum made by cryptographic hash of the other bits; the easiest method without using a computer is probably just to try different words in the hardware wallet until one works. With the Ledger at least you'll have to re-enter the first 23 words with every attempt, so on average you'll be inputing 128 * 24 = 3072 words; at ten seconds per word it's 8.5 hours of focused effort.

I'm confused. How is it not 128^24?

All you're doing is using the same 23 words each time, and then finding an 8-bit checksum to match.

(Each word is actually 11 bits of entropy, so the last word is 3 bits of entropy plus an 8-bit checksum on all the random bits. You could either determine those 3 random bits in advance and narrow the list to the 256 words that match those 3 bits, or just pick 23 words and randomly pick from the full list for the last word until one works; this is a bit simpler but risks having to try even more than 256 words.)

Ok, I got it now :)