That one could have worked out though.

If you assume that every 2 point conversion attempt is going to fail, then obviously every 2 point conversion attempt is going to be dumb.

In the Clemson game, they were down 21-3 before scoring a TD with 14:13 left in the game, making it 21-9 pending the conversion attempt. At that point there were a lot of potential outcomes. One such outcome was that if they converted the 2-pt conversion with 14:13 left, then they would have only needed a 1-pt conversion at the end of the game to win the game outright.

Another way to look at it is this way. Knowing in hindsight that Clemson kicked a FG later in that game and assuming you'll make all your kicks (not completely certain in a rainstorm but let's go with this for sake of argument), Kelly had 4 different potential outcomes with respect to 2-point conversions:

(a) Go for 2 the first time and convert - Win the game in regulation.
(b) Don't attempt it the first time - Go to overtime and 50% chance to win.
(c) Fail to convert the first time, succeed the 2nd time - Go to overtime and 50% chance to win.
(d) Fail to convert 2 times in a row - Lose the game in regulation.

Since (b) and (c) give you the same result, the decision comes down to which is more likely: Going 1-for-1, or going 0-for-2? If my math is correct, the breakeven point is whether your average successful conversion rate is 38%. Going for it gives you a 38% chance of winning (by converting the first time), and it gives you a 38% chance of losing [62% chance they fail the first time, multiplied by 62% chance they fail the 2nd time = 38% chance they fail both times]. So if they're going to convert it more than 38% of the time on average, then Kelly made the right decision against Clemson.

As for the 2014 Northwestern game, that was just plain dumb.