You know what this reminds me of? Those trained neural-net things which, however many training examples you give it, always seem to find some way to “cheat” and not do what you want while still obeying all your training data correctly.<p>Something like this: Suppose we have a table of strings of digits, some including spaces, and we’d like to remove the spaces. From<p><pre><code> 123 456
234567
345 678
</code></pre>
to<p><pre><code> 123456
234567
345678
</code></pre>
Now, what happens if it encounters, say<p><pre><code> 4567890
</code></pre>
Would the result be unchanged (as we would probably want), or would it “cheat” and remove the middle “7” character, giving “456890”?