Gary Marcus

Kinda irked by this post mate. The article doesn't warrant your addition of "takedown" / "apparent success" _and_ doesn't convince me. Nothing here was underhanded - these trade-offs have all been explicitly known and discussed - and even then the article only notes two of these.

The first, that it should use vision rather than an API, is essentially irrelevant imo. What the API provides could be produced by a vision model, the primary reason to avoid it is that it'd be a waste of training time _plus_ rendering the DotA client would be insane overhead.

What makes you sure that it'd be so simple?

For others following, I expand on it from my point of view in your thread :) tldr - the capabilities needed to produce the existing results would be doable in a traditional ML vision pipeline with human annotation and would still be a win.https://twitter.com/Smerity/status/1031297413388988416 …

I guess I have to ask: what is the question here? Human performance on DOTA doesn’t depend on access to bot buffers or large labeled databases; if we are trying to achieve human level performance with AI, we should try to reach a similar robustness in learning, no?

It's pure self-play, so no labeled databases, but I think there is real disagreement about how central the role of input representation is to the overall problem. Is learning a complex policy from hand-crafted representations solving 90% of the game, or 10%?