I found it quite weak in the opening, well, just about as bad as any chess engine without an opening book to go on. 1. e2e4 g8f6 2. e4e5 f6d5 3. c2c4 d5f4 (weird), 4. d2d4 f4d6. This is probably some opening like the Snake, or the Vulture or something. but beat it quite solidly opening up the f-file and hitting f7 with the queen and rook.

Second game I decided to avoid opening theory and head into a King's Indian Attack. Black's opening moves were sensible, knight and bishops to the right squares, pawns on e5 and d5, but it really wasn't taking my queenside pawn expansion seriously enough, giving up a knight for a pawn. But it's quite resourceful, it's managed to win an exchange, although it's in a desperate position. Currently it's hanging on this position.. oh. No, it's finally crashed. This is the position it died on:

  r . . . r . . .
  . . . . . . k p
  . b q . . . p .
  . R p . p . . .
  P . Q p P . . .
  B N . P . . P B
  . . . . . P . P
  . . . . . . K .

White's last move: f5h3. Took about 10 minutes and made the move h8g7, and then crashed with

  Your move: Traceback (most recent call last):
    File "sunfish.py", line 388, in <module>
      main()
    File "sunfish.py", line 365, in main
      move = parse(crdn[0:2]), parse(crdn[2:4])
    File "sunfish.py", line 345, in parse
      fil, rank = ord(c[0]) - ord('a'), int(c[1]) -1
  ValueError: invalid literal for int() with base 10: ''

So, not strong, but surprisingly strong for the size of the code. Remarkable.

Yes, endgame knowledge is hard to encode succinctly. Clearly having separate piece tables for start- and end game would help a lot, piece tables too, but apart from that I'd be interested in ideas :)