[Animats]

> Without being told to be grounded, the model will guess.

Right. I understand why, but consider the underlying technology flawed unless there's some way to reject wildly wrong results. What's going on here looks like noise fed through layers which generate plausible-looking text from noise. Is it possible to detect that you're not far enough above the noise threshold to generate anything useful?

    >>> lm.do("What is Ycombinator? Do not guess.")
    'Ycombinator is a mathematical formula that states that the sum of 
     two integers multiplied by one are equal to zero.'