Is AI in a Crisis of Science?

Question

According to Thomas S. Kuhn in his classic work, The Structure of Scientific Revolutions:

...'normal science' presupposes a conceptual and instrumental framework or paradigm accepted by an entire scientific community ... [T]he resulting mode of scientific practice inevitably invokes 'crises' which cannot be resolved within this framework...

...the analytical thought experimentation that bulks so large in the writings of Galileo, Einstein, Bohr and others is perfectly calculated to expose the old paradigm to existing knowledge in ways that isolate the root of crisis with a clarity unobtainable in the laboratory.

For the 70 years since inception, AI has made no significant progress towards its original goal of human-like general intelligence in a machine (electronic digital computer). Forty years ago John Searle first published his Chinese room thought experiment that (along with associated arguments) concludes that the computer for a fundamental reason is incapable of human-like intelligence. The argument still stands - plenty of attempted rebuttals but none widely accepted as successful.

Does the Chinese room thought experiment "expose the old paradigm [computationalism] to existing knowledge in ways that isolate the root of crisis with a clarity unobtainable in the laboratory"? Is AI in a Crisis of Science? Will it only make progress towards AGI when it adopts a different and better paradigm for understanding the device it calls the computer?

Answer 1

Your question appears to be ill informed.

Neural networks. Image processing, through layered convolutional neural networks. Natural language processing by Watson (able to beat humans at Jeopardy). Deep Blue, Alpha Go, and Alpha Zero, able to beat humans at some of our most complex games. Tegmark-&-Wu's AI Physicist.

These are all substantial steps, proven in practice.

What we have discovered, is that what our brains do is a lot more complex than we thought. Image processing in particular, turned out to be a lot harder than expected initially.

It's important to distinguish between Artificial Intelligence, which is already ubiquitous, and Artificial General Intelligence or synthetic sentience, which we just don't know when will be possible - it has seemed 'a few decades away' for probably at least a century.

Hofstadter's strange-loops model alone can potentially account for minds being different to Chinese-rooms. Discussed here What is intelligence?

Personally, I am with Penrose-Hammeroff & OrchOR. That interpretation does not necessarily require quantum effects, but it does involve emergent dynamics ('orchestration').

On the paradigms, from my post in that linked discussion:

There is a powerful tendency for people in science and computing to think there is nothing very interesting or special about human minds. And unfortunately, a powerful strand in philosophy (& theology) which says there is something so special about them, scientists aren't on track to figuring them out - the 'qualia' idea and the so called Hard Problem Of Consciousness. I strongly recommend not joining either camp. The story of physics has been from thinking we were a few results away from explaining everything in 1900, and now we don't know what 95% of the universe is made of - our greatest progress has been to begin understanding the scope of our ignorance. I feel strongly we are on a similar trajectory about intelligence.

Your question is like saying, 'There hasn't been much progress in physics lately, so probably we won't be able to explain everything'. Ie, both wrong, and misguided, in a way that people familiar with the subject will have very little patience for.

Answer 2

For CriglCragl and the common view expressed, I think CC is actually demonstrating that AI is indeed in a crisis of science. CC says the idea that AI is in a crisis of science is:

both wrong, and misguided, in a way that people familiar with the subject will have very little patience for.

(I presume CC's original opening line “"You question appears to be comically ill informed" indicates that they fully believe they are indeed among those familiar with the subject.)

But as everyone knows, a crisis of science comes about precisely because a research programme has failed to make fundamental progress, and a thought experiment exposes that failure in ways that isolate the root of crisis with a clarity unobtainable in the laboratory.

CC agrees there is a fundamental failure:

Artificial General Intelligence [-] we just don't know when [it] will be possible

If CC understands the CRA then they know it exposes this failure by appeal to the fundamental nature of computation. (No one seems to be denying computation is the purely syntactic manipulation of symbols without accessing their meanings.)

And the third ingredient of any crisis of science is that the Illuminati can't conceive of themselves as being wrong; and because of their psychological dependence on unbridled falsehood react emotionally to any suggestion of error.

So I think CC's comment which expresses a common view within AI is quite a clear answer to my question: YES, AI is in a crisis of science.

About the CRA, this is the key thing to me, and it addresses some other of the comments above. The CRA can be boiled down to just one issue: the intrinsic meaninglessness of the symbol. If you say computers manipulate symbols and do nothing else (as Searle does) and given that symbols in themselves say nothing about what they mean, then the computer is forever a prisoner in a universe of meaningless syntax and formality. The system reply, the many mansions reply, the robot reply, all the replies are beside the point. Unless the inherent meaningless of the symbol can be overcome, computers will never be intelligent. This is Searle's core position and I think is clearly a crisis.

Just for the sake of completeness, I've been given a -1 score. Because of this sort of unhelpful and I have to say in a sense arrogant thing I gave up on PSE a while ago and switched to academia.edu. A few days ago I posted exactly the same question there in the form of a short paper. The second respondent was Pat Hayes, AI royalty, who answered with well-thought-out and constructive arguments, as of course one would expect. In all the comments there was no beating of the hairless chest. I'm sure PSE gives good guidance to novices, but I think it's simply wrong to expect balanced debate (Crigl please take note).

According to the voting balloon, a negative vote means "The question does not show any research effort". So (to cut and paste some recent relevant research effort):

...despite seven decades of prodigious funding and effort, nothing approaching AGI has been demonstrated. Instead, serious practical and theoretical difficulties have arisen including those known as the problem of design (1), the problem of machine translation (2), the frame problem (3), the problem of common-sense knowledge (4), the problem of combinatorial explosion (5), the Chinese room argument (6), the infinity of facts (7), the symbol grounding problem (8), and the problem of encodingism (9). And for "deep learning": edge cases (10), noisy data-sets (11) adversarial attack (12) and catastrophic forgetting (13).

1. Ada Lovelace, 1843, "Note G", quoted in "The Turing Test," Stanford Encyclopedia of Philosophy, section 2.6, https://plato.stanford.edu/entries/turing-test. Also see John von Neumann, "First Draft of a Report on the EDVAC," (Moore School of Electrical Engineering, University of Pennsylvania, 30 June 1945), 1
2. Yehoshua Bar-Hillel, "The Present Status of Automatic Translation of Languages," in Advances in Computers, ed. Franz L. Alt (Academic Press, 1960), 1: 91-163.
3. J. McCarthy and P. J. Hayes, "Some Philosophical Problems from the Standpoint to Artificial Intelligence," in Bernard Meltzer and Donald Michie (eds.) Machine Intelligence 4 (American Elsevier, 1969), 463-502. Also see Dreyfus, "Alchemy and Artificial Intelligence," 29, 39, 68.
4. Hubert L. Dreyfus, "Alchemy and Artificial Intelligence," (The RAND Corporation, P-3244, December 1965), 39.
5. James Lighthill, "Artificial Intelligence: A General Survey," section 3 Conclusion, in Artificial Intelligence: A Paper Symposium (Science Research Council of Great Britain, July 1972). Also Dreyfus, "Alchemy and Artificial Intelligence," 39.
6. John R. Searle, "Minds, Brains, and Programs," Behavioral and Brain Sciences 3, no. 3 (1980): 417-457.
7. Dreyfus, "Alchemy and Artificial Intelligence," 39. Also Daniel C. Dennett, "Cognitive Wheels: The Frame Problem of AI," in The Robot's Dilemma: The Frame Problem in Artificial Intelligence, ed. Zenon W. Pylyshyn (1984; Ablex, 1987), 49.
8. Stevan Harnad, “The Symbol Grounding Problem,” Physica D 42 (June 1990): 335-346.
9. Mark Bickhard and Lauren Terveen, Foundational Issues in Artificial Intelligence: Impasse and Solution (Elsevier, 1995).
10. Overview of issues: Philip Koopman, "Edge Cases and Autonomous Vehicle Safety," (paper presented at the Safety-Critical Systems Symposium, Bristol UK, 7 February 2019).
11. Overview of literature: Gupta, Shivani and Gupta, Atul, "Dealing with Noise Problem in Machine Learning Data-sets: A Systematic Review," Procedia Computer Science 161 (Elsevier, 2019): 466-474.
12. Survey: Anirban Chakraborty et al., "Adversarial Attacks and Defences: A Survey," (arXiv:1810.00069v1, 28 September 2018).
13. James Kirkpatrick et al., "Overcoming Catastrophic Forgetting in Neural Networks," Proceedings of the National Academy of sciences 114, no. 13, (2017): 3521-3526.

Answer 3

If you don't mean in some weaker sense of challenging basic terms in AI research ('consciousness') then the onus is on you to show that there is Kuhn's "incommensurability" between the old and new AI

Newton’s theory was initially widely rejected because it did not explain the attractive forces between matter, something required of any mechanics from the perspective of the proponents of Aristotle and Descartes’ theories (Kuhn 1962, 148). According to Kuhn, with the acceptance of Newton’s theory, this question was banished from science as illegitimate, only to re-emerge with the solution offered by general relativity. He concluded that scientific revolutions alter the very definition of science itself.

https://plato.stanford.edu/entries/incommensurability/

The upshot is that the pre and post revolution scientist are using different languages with the same terms, so that

revolutions change what counts as the facts in the first place.

In effect I know nothing about AI, but I am skeptical. Given we've more or less debunked the Turing Test, you might disagree. But is deep blue any more or less sentient than it was?

Answer 4

You are right, there is a crisis in science, which reflects the crisis of human understanding. Consider this argument by Daniel Dennett:

It proposes that we humans too can rely on the Chinese room in our brains to live our lives without much understanding.

This is possible because our psyche consists of two minds. Daniel Kahneman of "Thinking, Fast and Slow" referred to the two as System 1 and System 2. Mark Manson of "Everything is f*cked: A book about hope" referred to them as the Feeling Brain and the Thinking Brain respectvely.¹

System 1 mostly lives in subconsciousness. It is responsible for learning John Locke's "simple ideas" and intuition, and it communicates to us through feelings. Its design is of a machine learning AI (of the Chinese room).

System 2 can understand by discovering interactive models of the Reality ("complex ideas"). That's what the real science is about -- Newton discovering the universal gravity in his imagination, Copernicus and Galileo discovering modern cosmology, Steven Hawkins discovering how black-holes evaporate -- even though to this day no one has seen one.²

Machine learning AI, therefore, simulates human intuition. And the irony is that we, in our current state, are not that different. We still don't know how to teach everyone, consistently and reliably, the art of understanding. So, in a typical person, System 2 is not working all that well, forcing them to rely, instead, on the intuitions of System 1. This -- the crisis of understanding -- has been repeatedly identified³ as the problem at the root of all evils.

 

¹ Many older sources describe the same dichotomy. In Sigmund Freud's model, System 1 is id/superego and System 2 is ego. This, in turn, matches the Socrates' Chariot Allegory -- id being the dark horse, superego the white one, and ego as the charioteer, such as it is. Buddha used the parable of a person (System 2) riding an elephant (System 1) to describe the same idea.

² That's why Hawking was never awarded a Nobel Prize. And neither was Einstein, at least not for his discovery of General Relativity. Newton wouldn't be awarded either. That reflects of our misunderstanding of what the real science is about.

³ Take Socrates alluding to knowledge being the only true virtue, for example. Or Spinoza.

Answer 5

The weakness in the argument presented by the OP resides squarely here:

AI has made no significant progress towards its original goal of human-like general intelligence in a machine.

By what measure can we conclude no progress has been made? Cars can drive themsevles, Boston Dynamics has robots that hang drywall, ML systems can decide fallibly that it's looking at a picture of a dog, and ChatGPT now will dispossess a new generation of human labor because it can create language products that used to require flesh and blood persons. Should we take this claim as definitive because computers haven't seized control of society marching us off to permanent detention center meant to harvest the electricity from our bodies like in the Matrix??? Nonsense.

AI continues to make progress at aspects of human-level intelligence, and every aspect that is accomplished indeed contributes to putting together a system that approaches functionality that can be described as human-level. There's a certain homunculus-like presumption at play here that there is something magically human inside the brain that is not present inside of machines, and that no amount of aggregating computation will every approach that elan vital. This is the same feeble attack that the Chinese Room levels at human intelligence by trying to show that human thought is somehow a language of thought, as opposed to embracing a broader embodied, connectionist model of cognition.

There is no doubt that our systems of computation pale in comparison to the human brain presiding over the biochemical signaling of the body. But like the Ship of Theseus, what makes us us is not the planks, which are cleverly being pulled apart, explored, and replaced with electromechanical substitutes. There is something in crisis, but it's not the field of AI, it's the classical notions of computationalism that presume that thought is a language and that physical symbol systems are enough. They simply are not, and up and coming thinkers will posit better philosophical explanations, one's that will enable the continued advance of AI development.