Corrections
Everything here was published, then revised. What I said, what the larger sample said, and when. Newest first. Typos and predictions are not listed.
Jev solves 0 of 6 puzzles by ranking
0 of 6 eight-puzzles solved 16 of 30, once the search could back up
The first 8-puzzle search used a single fixed beam that could never back up, so one poor ranking lost the puzzle. Even Manhattan distance scored only 4 of 6 on it, which meant the number described my search rather than the model.
Replaced with a search that widens and retries. Validated offline first: a perfect ranker now solves 30 of 30, an uninformed one solves 1, Manhattan solves 29. Only then is 16 of 30 a statement about Jev.
Beats the textbook tagger
Jev beats the textbook tagger, 93.9% against 92.5% On par: 92.9% against 92.3% at twice the data
First pass, 100 sentences: Jev 93.9% against 92.5% for a hidden Markov model. I wrote that it beat the textbook tagger.
Second pass, 200 sentences: 92.9% against 92.3%. The paired difference runs from −0.7 to +1.9 points. The honest word is on par, not beats.
Never died in the Wumpus world
Eight worlds, no deaths Fifty worlds, died in 10
Eight worlds, no deaths, and I said so.
Fifty worlds: died in 10 of them, against 3 for the hand-written agent.
The connectome ablation number
Removing the two loom-sensitive cell types dropped escape to zero 0.33 on MaleCNS v1.0
I had measured that removing the two loom-sensitive cell types dropped escape to zero, on an earlier dataset, and I had been repeating that number.
Re-running on MaleCNS v1.0 gives 0.33, not zero. Some of the response arrives through other loom-sensitive cells I had not accounted for.
No style could save the last four guests
The four guests with no portrait were style failures A framing problem, not a model problem
I had been saying the four guests with no portrait were style failures.
I went and looked at the photographs. Two have their heads touching the top edge of the frame, one has run off the left side, and one is small and low in a corner because the camera was aimed slightly at the ceiling.
Not a model problem. A framing problem.