Gary Marcus

@GaryMarcus "search a different part of model space" is not a testable proposal, nor your whole book from 2001 is. I wanted to hear what that different part is, in equations or pseudo code.

ps 2001 book proposed many specific problems ranging from generalizing low frequency default morphology to acquiring universally quantified one to one mappings to tracking identity and properties of individuals over time. community really has not engaged in any of these.

Researchers rarely engage with general audience non-fiction. If your goal is community engagement, try publishing where they publish. Paying money and sifting through hundreds of pages is a big ask.

Steve, tons of us at DeepMind have read The Algebraic Mind. I see it on people's desks all the time. It's more of a long review article than a popsci book. This is why the usual response to Gary's critiques is "we know, you've been repeating the same thing forever."

I should've rephrased. I've read the book, and I'm trying to explain why it didn't have the impact of something like Lake's Omniglot work. I prefer you're description and think it's had the legacy of a good review article: useful for newcomers, not detailed enough to build upon.

It probably has had a bigger impact in the long term though, just in a vaguer way. It probably shaped the thinking of guys like Josh and Brendan long before they wrote that paper.

if people would actually take on the challenges of the three core chapters instead of dismissing them, it would have even greater impact.

Maybe the thing to do is put some of these examples up as a task/dataset on GitHub, like bAbi, or hold a competition at a NeurIPS workshop, like they do for adversarial attacks on neural networks. That's how the kids do things these days.