The Theory of Anything

How does ChatGPT really work? Is there a relationship between a program like ChatGPT and artificial general intelligence (AGI)?

Bruce Caldwell (a scholar interested in Popper and Hayek) wrote a long paper in the Journal of Economic Literature (March 1991) called 'Clarifying Popper'. In this episode, Bruce Nielson summarizes and discusses Caldwell’s paper on how Popper’s ideas could be applied to economics. How well did Bruce Caldwell do in his goal of clarifying Popper's epistemology?

Out next episode is another interview with Brian Boutwell and we discuss this paper a few times. So this summary will help those that don't have access to it.

Brian Boutwell is a professor of criminal justice at the University of Mississippi who specializes in “quantitative genetics, with a focus on environmental and psychological risk factors for antisocial and violent behavior.” He has a TED talk, numerous articles in Quillette, and has been published in many journals. Here we discuss his upcoming meta-analysis on twin studies soon to be published in Nature.

We discuss the following two articles:

Is the government hiding a secret UFO recovery program? What should the critical rationalist attitude be towards these kinds of claims? Why exactly would aliens want to hide from us? We discuss these questions and much more.

There seems to be broad agreement, even among Karl Popper's own students, that he was a deeply dogmatic individual. In this episode we ask the question 'Was Karl Popper Dogmatic?' by reviewing a humorous article in Scientific American by John Horgan on August 22, 2018. Along the way, we discuss by what means we judge dogmatism. How do we even tell if someone is dogmatic or not? Is there a litmus test for dogmatism? If so, what is it? 

How do we square genetically influenced mental disorders with the theory of explanatory universality?

In a previous episode, Tracy asked Bruce how to reconcile her experience with mental disorders, like narcissism, with the idea of Universal Explainers. This is part 2 of that discussion.

In the last episode, Bruce introduced the idea that emotions and feelings aren't the same as ideas and go back to an earlier point in our evolutionary history. The genes then use our feelings to try to coerce us or encourage us via pleasures and pain. 

Bruce completes his list of possible ways genes can affect our personality and ideas without violating universality:

1. The genes can control physiology and this in turn can impact our personality and ideas via interaction with existing (sometimes stable) culture
2. The genes can control how we grow the various parts of the cortex and since those parts affect our ability to think, they affect our personality development as well as interests.
3. The genes can control perceptions and this can in turn impact our ideas.
4. The genes control how we’re wired to pleasure and pain centers of the brain and can coerce or encourage us via these feelings.
5. The genes control how we gain ideas via attention.
6. The genes can affect culture via 1-5 above and then let culture do the heavy lifting
7. Humans may be significantly affected by older animal modules of the brain in some cases. We have no reason to believe all knowledge we learn is via ‘the universal explainer’ module.

How do we square genetically influenced mental disorders with the theory of explanatory universality?

In our last episode, Tracy asked Bruce how to reconcile her experience with mental disorders, like narcissism, with the idea of Universal Explainers. In this episode, Bruce does his best to tease out an answer. (While admitting that we can't answer her entirely--yet.) 

In "The Beginning of Infinity", David Deutsch offers some solid criticisms of current experiments to determine how much of a personality trait is 'heritable.' This has led some of his fans to take his ideas to some extreme conclusions not implied by the book. For example, some people now claim that genes play no role at all in influencing Universal Explainers. In fact, Deutsch did not say this. 

According to Deutsch (in BoI), genes can influence our ideas and personality traits via something as simple as how physiology (physical traits) interact with culture (standards of beauty) and that can in turn impacts one's personality (perhaps increasing happiness.) So we now have at least one example of how genes can have an impact on our personality and ideas. (Via physiology interacting with culture.) 

With this in mind, Bruce asks the obvious question: What are other ways genes can affect personality traits and ideas that do not violate explanatory universality? 

Bruce's list (partially revealed in this episode) is a testable set of ways genes may impact our personality and ideas. This suggests how we might go about responding to critics of the theory of Explanatory Universality without violating Popper's epistemology via either ad hoc saves or ignoring basic statements (i.e. repeatable observations) from existing experiments.

Over the years Bruce collected a series of 'problems' with the Popperian concept of refutation. Or so he thought. A chance encounter with Popper scholar Danny Frederick led to him re-evaluating Popper's writings and realizing that Popper sometimes uses terms (such as 'refutation', 'falsification', and even 'theory') in idiosyncratic ways that aren't quite how most people would understand those terms. This leads to both Popper's opponent and fans alike sometimes misreading him. It turns out that the 'problems of refutation' that many philosophers cite as disproof of Popper are actually due to misunderstanding Popper due to his specialized vocabulary.

In this episode, we cover what Popper himself said about the asymmetry of refutation vs verification, how it relates to the demarcation between empirical and non-empirical theories, and even how it relates to induction. Then we use that knowledge to resolve the 'problems of refutation' we discussed in the last episode. 

Blog Post Series on The Problems of Refutation

Over the years Bruce collected a series of 'problems' with the Popperian concept of refutation. Or so he thought. A chance encounter with Popper scholar Danny Frederick led to him re-evaluating Popper's writings and realizing that Popper sometimes uses terms (such as 'refutation', 'falsification', and even 'theory') in idiosyncratic ways that aren't quite how most people would understand those terms. This leads to both Popper's opponent and fans alike sometimes misreading him. It turns out that the 'problems of refutation' that many philosophers cite as disproof of Popper are actually due to misunderstanding Popper due to his specialized vocabulary. 

In this episode, we cover Bruce's list of 'problems of refutation' (which he know believes are all pseudo-problems) and explains his encounter with Danny Frederick and how it led to him re-imagining Popper's epistemology in different terms that were easier for laymen (and philosophers) to understand. 

Blog Post Series on The Problems of Refutation

In this (mostly) standalone episode, we cover how Deutsch and Byrne each interpret Byrne's theory differently. Deutsch emphasizes the micro-level actions and gestures of great apes and the clear lack of understanding of what each gesture does. Byrne emphasizes the macro-level and the flexible intelligence required to come up with a program of action to accomplish a novel goal. Byrne's theory of 'animal insight' makes specific testable claims. To Byrne, great apes (especially Chimps) can 'think.' His theory says that animal insight was a necessary precursor to human insight and that humans utilize both kinds. If he's right, then animal insight has relevance to AGI studies. Deutsch has doubts about all of this and thinks of Bryne's theory more as evidence that animals cannot think.

We also discuss how Byrne and Deutsch both understand the mirror test differently. And finally, we dip just a bit into animal sentience and discuss why the theory that animals feel things is the prevailing theory not so much because it's a great theory but more because it has no real current competitors. It's difficult to explain much animal behavior without either tacitly referring to animal feelings or just clearly making up bad ad hoc explanations. 

While it's helpful to have listened to the 3 previous episodes, this episode mostly stands alone.

Links:

Richard Byrne has spent his whole career trying to determine when animals learned to 'think.' We discuss Richard Byrne's methodology for determining which animals have what he calls 'insight' (the ability to utilize mental models) and why his methodology is awesomely Popperian. Then we go over many examples of animal behavior that can't be explained via genetic programming or trial-and-error learning. We also compare machine learning and animal intelligence and why animal intelligence is beyond our current machine learning capabilities.

Links:

Karl Popper has a radical theory of 'dualistic evolution' where behavior had to evolve first before physical evolutionary changes could be taken advantage of. As part of his theory, Popper pointed out that an animal's ability to learn would be paramount to making evolution work at all -- similar to the Baldwin effect discussed in the last episode, but now for physical adaptions. This means evolution would have had intense pressure to evolve learning algorithms early in the evolutionary tree. 

As it turns out, Richard Byrne's work largely corroborates Popper's theory of dualistic evolution. Nearly all animals show an ability to do trial-and-error learning and this is the main source of 'animal intelligence' in the animal world. Byrne even argues that this ability to do trial-and-error learning is a form of evolution where animals let their behaviors 'die in their place' rather than having to wait for the slow biological evolutionary learning processes of the genes.

We also discuss what split-brain patients might teach us about human explanations and go over examples of animal-like gene channeled learning in humans. 

Links:

How intelligent are animals?

When Alpha Go beat Lee Sedol, the world Go champion, it came up with creative new moves never previously seen before and even invented a whole new style of play unknown to humans. IBM's Deep Blue, the champion chess algorithm, failed to do either of these. What was the difference?

Reinforcement Learning is a machine learning algorithm that is a 'general purpose learner' (with certain important caveats). It generated a lot of excitement with its stunning victory of Alpha Go against Lee Sedol the world Go champion.   

In this podcast, we go over the theory of reinforcement learning and how it works to solve any Markov Decision Problem (MDP).   

This episode will be particularly useful for Georgia Tech OMSCS students taking classes that deal with Reinforcement Learning (ML4T, ML, RL) as we briefly explain the mathematics of how it works and show some simple examples.  

In this episode, we interview Chiara Marletto about her recent book The Science of Can and Can't: A Physicist's Journey Through the Land of Counterfactuals as well as discussing Constructor Theory in general and how it might help us form a new mode of explanation in physics. We ask her some tough questions about constructor theory and she fields the questions very well. 

Donald Campbell made the bold prediction that all expansions of knowledge will be found to require the Universal Darwinism algorithm of variation and selection. In this episode, we're going to test that prediction and see if it holds up against what we currently know about Artificial Intelligence and Machine Learning.   

For example, does (apparent) knowledge created by Gradient Descent require variation and selection? Or is it really and truly inductive? Or does it just fail to create knowledge at all despite clearly creating improvements?   

Ultimately, we'll find that Machine Learning creates an exciting set of epistemological problems that need to be solved!

In The Beginning of Infinity, David Deutsch claimed that no existing evolutionary algorithm has yet created knowledge. But Karl Popper and Donald Campbell beg to differ and have argued that knowledge-creating evolutionary algorithms are ubiquitous, common, and easy to implement. Who is right? 

In this episode, we look at both arguments and assess them using Popper's epistemology. And along the way, we'll define the minimum requirements for an evolutionary algorithm (aka Universal Darwinism), explore what knowledge-creation is and then, finally, we'll attempt to answer the question of whether or not existing Artificial Intelligence algorithms create knowledge.

The popular media confuses Artificial Intelligence and Artificial General Intelligence. All the progress is in the first while all the interest is in the second. But what is Artificial Intelligence? In this episode, we explain the umbrella term and its subfields. Plus we introduce how Artificial Intelligence actually ties to all four of David Deutsch's four strands. That makes it an exciting field all of its own even though it's not a path to AGI.

Many Worlds Quantum Mechanics is the only current explanation we have of quantum physics. Yet most scientists today still prefer to not have an explanation at all rather than accept it. 

Sam Kuypers joins us to discuss his paper "Everettian relative states in the Heisenberg picture" that he co-authored with David Deutsch. He explains why the Heisenberg picture of quantum physics lends itself naturally to a local many worlds view of quantum physics. 

Also, we discuss if King Arthur could possibly be both real and fictional at the same time. Whaaaattt!?