Built2Think

Student of CogSci and AI, believing in continuous & life-long learning. Humbled by the enormity of the scientific and engineering problem.

[Built2Think @Built2T]

I just published "How the multi-disciplinary search for Mind re-discovered the need for Body". Looking forward to feedback. https://t.co/ySAiOZJ5Jh

[Built2Think @Built2T]

@TrueAIHound So true. I find it incredibly frustrating that the role of coupling sensor data with effector commands and status data is so lost on the “scale is all you need” crowd. Watching without at least some controlling is such a weak way to learn.

[Built2Think @Built2T]

@r0ck3t23 Altman is wrong about many things, but he is certainly misguided to say nobody loses a job when a disease stops existing. Big Pharma lives off of plausibly deniable disease perpetuation.

[Built2Think @Built2T]

@TrueAIHound @Arcaed0x @RichardSSutton @kjaved_ Unfortunately, I believe I’m getting what I pay for. Not trying to do engagement farming or follower count expansion, and no blue check mark.

[Built2Think @Built2T]

@TrueAIHound @Arcaed0x @RichardSSutton @kjaved_ I would also like very much to pick your brain more. When you say perceive vs recognize, is it mostly about factoring the sensory data into distinct entities or independent objects, without needing to name them?

[Built2Think @Built2T]

@carl_feynman @geoffreyirving @DavidSKrueger The best models are also obviously subhuman in other ways. Working on problems of perceptual interpretation, continuous learning, & motor control to accomplish goal driven behavior make AI & robotics research worth doing. There are still many open problems in cognitive science.

[Built2Think @Built2T]

@karlmehta A child doesn’t learn a world model from (just) watching videos. They interact with the world to establish causal connections between actions that change perspective & sensory data that correlate with those actions. Add to that motor effects that physically alter the world.

[Built2Think @Built2T]

@HowToAI_ It seems to me that he has misunderstood what people mean by “general” in AGI. Doing all the things humans can do ought to include learning to do them as humans do, with similar speed of adaptation. Humans are individually adaptable & collectively general.

[Built2Think @Built2T]

@GaryMarcus Gary, I’m a big fan of your ideas, but I’m curious why you don’t place greater emphasis on continual embodied learning as a foundation for neurosymbolic conceptual grounding? Empathic learning also seems key to efficiency and safety. Causal discovery and reasoning applies to both

[Built2Think @Built2T]

@TrueAIHound @fchollet Perceiving objects (via vision alone) has to involve the coordinated motion of two eyes, the head, and the body, interacting with physical entities that have boundaries and relative orientations in 3D space. The sensory temporal data stream is partially controlled by the agent.

[Built2Think @Built2T]

Every scientist should see the new movie “The Story of Everything”, in theaters now. It’s a perfect example of how abductive inference leads one to conclusions that have no plausible alternatives.

[Built2Think @Built2T]

@lathropa @Grady_Booch See his patent for starters. Minsky's SoM + Blackboard gives a collaborative multi-agent system with a public shared memory space. Brooks' subsumption adds hierarchical real-time interruptible goals. Embodiment uses a sense-decide-act cognitive loop. Easy to research components.

[Built2Think @Built2T]

@rao2z Well said. We teach a child by saying “use your words” to assist with conceptual grounding. Note also how perceptual data factoring is a key part of making long range predictive world models tractable. Object vs pixel level understanding is so important.

[Built2Think @Built2T]

@_hanneslehmann_ @aakashgupta Paradigms & architecture & framing are intertwined. You are right that prediction without action conditionality (causal modeling) will also fall short. Continuous embodied learning, with concentration on active perceptual factoring will help greatly with grounding. Do to learn.

[Built2Think @Built2T]

@FiftyOne_50_ @r0ck3t23 Now ask yourself if it might just be the case that causal grounding is easiest to learn from the perspective of an agent with sensors, and effectors that can influence what will be sensed. Embodiment is exactly what you need to test (at least some) causal hypotheses.

[Built2Think @Built2T]

@VraiNom554355 @lemire There is quite a range of possible meanings for AGI. LLM proponents often limit their criteria to being able to converse intelligently, whereas a more challenging definition requires demonstration of real world sensing and action (embodied cognition).

[Built2Think @Built2T]

@Grady_Booch The funniest exchange I’ve heard was a guy who asked the LLM to let him have the last word, and of course the chat bot just HAD to respond every time. It went on for many rounds without ever just shutting up. Priceless.

[Built2Think @Built2T]

@predict_addict From what paper is this figure taken? Do please continue to explain what you mean by time series gravity. I’m intrigued.

[Built2Think @Built2T]

@Dr_Gingerballs I think you are quite correct regarding the use of parametric statistical models on purely observational data. I personally have more hope when causal, interventional, and non-parametric, non-stationary statistics are brought to bear. Empathetic and embodied continuous learning.

[Built2Think @Built2T]

@TrueAIHound @jbthinking That mostly applies to model-free RL, but even model-based RL, which does leverage prediction more directly, only builds weak models of the agent-environment interaction.

[Built2Think @Built2T]

@TrueAIHound @jbthinking I think RL mostly “sucks” because it takes the easy way out in dealing with the credit assignment problem. Without a factored and causal model of how actions affect observations, RL averages over enormous numbers of episodic experiences quite blindly, finding behaviors that work.