He demonstrates without a doubt that 128 bit of hash is not enough, yet you can still find many projects still publishing MD5 hashes. The long tail of bad crypto is extremely long.
Sometimes whether a cryptographic protocol relies on collision resistance can be surprisingly nuanced, so it should be phased out for this alone (and as we have better options) but for simple examples (e.g. to make a signed hash of an executable, which is probably equivalent to what you're describing) it's not broken.
People that take security seriously enough to check hashes should not trust MD5 so the scenario is not super credible, but people still publish MD5 hashed like it's the early 2000s.
There is a usability issue here. A 256 bit hash is so long as to be very hard to manually compare. A project might want to keep the length down to something reasonable.
If you let me pick which bits matter for collision detection, then I can make any secure hash collide with itself.
On a related note, someone took the end of Contact to heart, and definitively proved that God is a Where's Waldo? fan: