Gary Marcus latest manifesto "The Next Decade in AI" arxiv.org/abs/2002.06177 attempts to reintroduce the philosophical foundations of GOFAI (i.e. knowledge, internal models, reasoning) into the current AI conversation.
I do not disagree that knowledge, models, and reasoning are essential to cognition. These are however emergent phenomena out of a substrate that is an intuition machine. Absent in Marcus arguments is the notion of Intuition (System 1) or ideas from complex adaptive systems.
To his credit, Marcus correctly argues that cognitive systems must be cognizant of the nature of reality. Here he proposes an ontology for containers: 
What he avoids recognizing is that cognition about these already exists in the fundamental building blocks of biology (i.e. the cell). What Deep Learning overlooks is nano-intentionality: medium.com/intuitionmachi…
Deep Learning systems will evolve to incorporate more complex building blocks. We already see this trend today with Transformers and Capsules. This is despite the original conceptualization of Artifical Neural Networks were based on simplified neurons.
Biological neurons are sufficiently complex that 'reasoning about containers' is an intrinsic capability. Billions of years of evolution have made this possible. medium.com/intuitionmachi…
I have originally proposed a development roadmap that has a path via the 'hybrid' stage that Marcus describes: medium.com/intuitionmachi…
In the near time (perhaps this decade), there are pragmatic benefits in pursuing a hybrid approach. I've written earlier how such System 1 and System 2 coordinate: medium.com/intuitionmachi…
The solution however to AGI is however grown from the bottom up and not engineered from the top down: medium.com/intuitionmachi…
Intuition machines are a more faithful development model of human cognition that the algebraic minds that Gary Marcus proposes.
Breakthroughs are made not by seeking band-aid solutions (i.e. hybrid systems) but recognizing that there is a radical new way of understanding general intelligence.
This implies reframing what @GaryMarcus finds important (i.e. knowledge, internal models, reasoning) from the perspective of emergent complex adaptive system behavior. Not from a vacuous statement: 'operation over variables'.
@GaryMarcus Cognitive systems must be inside-out, this implies the existence of internal models. Internal models do not imply that they are expressed with concepts with crisp boundaries. medium.com/intuitionmachi…
@GaryMarcus Reasoning is a System 1 behavior. Specifically it is a meta-level cognition. It is intuition about one's intuition. medium.com/intuitionmachi…
@GaryMarcus Finally what Marcus calls knowledge is the knowledge that is captured in language. How is intuition (i.e. System 1) capable of language? How is the semantic gap bridged? The answer: Symbol detachment: medium.com/intuitionmachi…
@GaryMarcus So here you have it, in a set of tweets and blog posts, I have laid out an alternative vision for 'the next decade of AI'. One that offers sufficient detail to distinctly different itself from Gary Marcus' manifesto.
@GaryMarcus @threadreaderapp unroll for a summary!
