1.b6 2.swap 3.l12 4.k17 5.n17 6.l10 7.j11 8.m12 9.n13 10.j9 11.o11 12.o7 13.t7 14.r8 15.o9 16.n9 17.d4 18.g10 19.e7 20.e11 21.j10 22.i11 23.e12 24.t9 25.t11 26.v8 27.d10 28.f14 29.f15 30.g16 31.e13 32.f13 33.r7 34.p9 35.g17 36.c10 37.b11 38.e9 39.c9 40.c8 41.b7 42.b6 43.d6 44.a8
after 6: -0.270
after 12: -0.199
after 18: 0.641
after 24: 0.907
after 30: 0.859
after 36: 0.994
after 42: 0.999

D4 wtf??!! This is one of those TB games where one side seems to suddenly turn and smash into a brick wall.

Maybe white missed
|1.b6 2.swap 3.l12 4.k17 5.n17 6.l10 7.j11 8.m12 9.n13 10.j9 11.o11 12.o7 13.t7 14.r8 15.o9 16.n9 17.s10 18.u7 19.r6 20.q10

And went bonkers as a result.

Jordan, please correct me if I'm wrong, but I think "went bonkers" means:

At move 16, the board was still open enough that the analysis didn't complete in 5 minutes, but during analysis, all the best-looking spots were eliminated. Nevertheless, the neural net had made a heat map of the entire board, and at the 5 minute mark, the hottest remaining spot was chosen, whether it was analyzed or not. Huge swaths of the board had gone cold, but it wasn't done yet. Given more time, it would have probably resigned. But it's not gonna resign if there is still a glimmer of hope. It'll take its best shot at the buzzer.

A human would still pick a reasonable-looking / most-chaos-inducing move, hoping that his opponent makes a silly mistake, or that the analysis was happily wrong.

So what can we learn from this? If TwixtBot makes a totally bonkers move, it might mean that you're winning?

Right. Almost certainly what the bonkers move means is that the bot looked at all the "reasonable" moves and decided it ended up with a completely lost position, so it tries the next move down the list; and in the time available it ran out of moves to try.