Today’s AI’s: Not Sentient — But Getting Closer
From Hofstadter to Hinton: Dall-E, Deep Learning & “Slippability”
You may have seen the news stories about the Google engineer who publicly claimed that the company’s LaMDA conversational AI system is actually a sentient being.
No, LaMDA isn’t sentient — here I agree with the vast majority of AI experts and AI ethicists who have chimed in on the question. However, there are behaviours and capabilities of another AI system in the news that suggest to me that today’s AI researchers are zeroing in on key mechanisms of cognition and intelligence.
To see why, let’s go back a few decades to a fascinating thinker and author who popularized ideas around AI and cognitive science for a generation of aspiring researchers, programmers and tinkerers.
Hofstadter, Analogy and Creativity
Douglas Hofstadter’s “Godel, Escher, Bach: An Eternal Golden Braid” came out in 1979. I first encountered it when my Aunt Susan gave it to me for my birthday around 1981–82, and it’s no exaggeration to say that the book changed my life. There I was, minding my own business, halfway through an Honors degree in Economics at Cornell, when that book got me hooked for life on the important questions around human and machine intelligence and consciousness. I changed my academic major, took a bunch of Math and Computer Science courses, and it cost me an extra year of undergraduate study. (Worth it!)
In “GEB” and his followup “Metamagical Themas”, amidst clever games and puns and meditations on the mathematics of music and art, the author confronts the profound questions around what it means to perceive, think and create. A core idea of Hofstadter’s is what he called “slippability”:
My word for the elusive aspect of human thought still lacking in synthetic imitations is “slippability”. Human thoughts have a way of slipping easily along certain conceptual dimensions into other thoughts, and resisting such slippage along other dimensions. A given idea has slightly different slippabilities — predispositions to slip — in each different human mind that it comes to live in.
Common manifestations of such slippability among us humans include the use of analogies and metaphor. Both stem from the seemingly simple but actually profound idea that two superficially different things share some deeper structural sameness. Hofstadter and colleagues illustrated this with a game in which one tries to answer questions like these:
- What is the Newark of Delaware?
- What is the Toronto of Uruguay?
- What is the Honolulu of Oklahoma?
- What is the Hollywood of Africa?
- What is the Gettysburg of Hawaii?
Slightly more formally, the problem is to find the proper noun (city name) C such that, for example, “C is to Hawaii as Gettysburg is to Pennsylvania”. To find C requires identifying some semantic relation (geographic, topographic, historical, cultural) between Gettysburg and Pennsylvania, and then transferring that relation (while letting other aspects “slip”) to the situation of Hawaii. Such analogical thinking isn’t that hard for most humans — though students taking SAT exams might disagree — but was one of many notorious challenges for early AI systems.
Slippability is also seen in our ability to create, or to recognize, something that is “in the style of” someone or something else. To observe that a painting “reminds me of Monet”, or that a piece of music “sounds a bit like Bach” presupposes that there are deep patterns uniting otherwise disparate works of a creator (or of a “school” of creators, like Baroque composers, Impressionist painters, or 1970s funk bands). It is to discern that while many aspects of a template or pattern might “slip” from one creation to another (for example, the meter and the key signature of a song), that something essential remains consistent. Varations on a theme, as it were.
A fun example of this phenomenon is this guy who runs through dozens of riffs of the same song, each version in the style of some famous band or singer:
To discover a pattern (“the Johnny Cash sound”, or “Boys II Men songs”) and re-apply it in a new context (“Bohemian Rhapsody” by Queen) is the essence of Hofstadter’s slippability, it seems to me. And it takes intelligence to execute, and to recognize and appreciate. It is at the heart of not just satire and parody, not just musical sampling and artistic homage; but of interpretation, analysis, scientific induction, creativity and invention more broadly.
So, Hofstadter identified an important facet of intelligence. And he tried to model it, to implement it in computer software and hardware. With students and other colleagues he developed a series of programs, the most successful of which was probably Copycat, a collaboration with Melanie Mitchell.
In the work of Mitchell and Hofstadter, high-level perception of the type that enables us to solve analogies emerges from the spreading activity of many independent processes, called codelets, running in parallel, competing or cooperating. They create and destroy temporary perceptual constructs, probabilistically trying out variations to eventually produce an answer. The codelets rely on an associative network built on pre-programmed concepts and their associations (a long-term memory). The changing activation levels of the concepts make a conceptual overlap with neighboring concepts.
In my view, Hofstadter and his associates were onto something big with this idea of analogy as an emergent phenomenon that is critical to cognition. But many felt that it wasn’t quite emergent enough; that its reliance on pre-programmed symbolic representations of concepts failed to grasp an essential aspect of what happens in human brains and minds.
Hinton: Deep Learning for Richer Relational Representations
So if Douglas H. didn’t quite nail it, has anybody else? In order to learn more about the actual mechanisms that produced phenomena like Hofstadter’s “slippability”, I needed to find a new intellectual source of inspiration. After wrapping up at Cornell, I was very fortunate to find my way into the laboratory of Professor Geoffrey Hinton at the University of Toronto. Geoff, who later won the prestigious Turing Award for this work, was a pioneer of artificial neural networks and of what is today called Deep Learning.
Deep Learning employs artificial neural network architectures composed of multiple levels (“layers”) of non-linear operations (using “neurons”). This is a highly active area of research, so new variations emerge regularly — from Convolutional Neural Networks (CNNs) for machine vision to Transformers for natural language processing (and vice versa, recently). What they share is perhaps best summed up by Yann LeCun, head of AI at Facebook/Meta, who was a postdoc in the Hinton lab when I was a grad student, who describes the deep learning paradigm as “solving the problem of learning hierarchical representations and complex functional dependencies”.
Rather than symbolically-represented “concepts” explicitly designed and programmed in by a human, AI models trained by deep learning build their own rich representation of concepts and relations between concepts, by being exposed to examples or by interacting with the outside world. (As humans and animals do).
Deep learning is the basis for the LaMDA conversational AI that so charmed and alarmed one of its researcher-creators. And it’s also the basis for the computer-generated art that’s caused so much buzz in recent months.
AI-Generated Art: Caravaggio’s (or Dali’s, or Monet’s) Ghost in the Machine
Dall-E is an AI model that allows a user to type in suggestions that the AI then “draws” or “paints”. It’s one of several such prominent models and systems, a growing list that includes Craiyon (Dall-E Mini) and MidJourney.
So, what impressed me about these models, compelling me to write an article? When I played around with the Dall-E Mini app, I gave it a few suggestions to “paint” pictures in the styles of some of my favorite artists. For example, below is the result from my prompting “A hockey game painted by Caravaggio”. (Caravaggio is one of my favorite Renaissance painters, and hey, I live in Canada, so hockey).
What struck me was the clever “slippability”. Again, it’s that identification of a theme to hold fixed (e.g., Caravaggio’s dramatic use of shadows and light) while letting other aspects change (from Renaissance Italy and Biblical or Classical subjects, to a bunch of guys playing pick-up hockey). I ran additional “experiments” (okay, fun time-wasting distractions from my real work deadlines) that generated strange beings and worlds in the styles of Dali, Monet and other great artists. In each case, there seemed to be something deeply … interesting going on.
You can find many more such images — from Dall-E, Craiyon, MidJourney and more, all over the web. (Some funny ones are here, including depictions of Nosferatu in RuPaul’s “Drag Race” and Darth Vader ice fishing).
Pattern Discovery, Sentience and Consciousness
It seems to me that these AI-generated art models are demonstrating something powerful, beyond their mere entertainment value. They are discerning deep patterns — one might say they’re discovering concepts — and re-applying these patterns in new contexts. And they’re doing it with the rich, distributed internal information structures that arise from deep learning.
Such pattern discovery goes well beyond cute surrealist art tricks and clever chatbots. Earlier this year the DeepMind team announced that their AlphaFold system predicted the 3D structures of millions of proteins, including essentially all proteins expressed from the human genome. This represents a major scientific achievement with profound implications for pharmaceutical development and health. Again, it was deep learning — the multi-layer neural network architectures enabled by new algorithms, vast datasets and ultra-fast computing architectures that allowed AlphaFold to go well beyond what my colleagues and I were able to accomplish in protein folding work back in the day. AlphaFold is clearly discerning, and exploiting, deep and meaningful patterns — concepts —in the relationships between 1-dimensional amino acid sequences and the 3-dimensional molecular structures into which those amino acid sequences fold.
The concept representations built up within AlphaFold, Dall-E and LaMDA aren’t explicitly designed or programmed into them by a human. “Semantic originality” was a phrase often discussed in AI and cognitive science back in my grad school years. Less so today, but I think it’s still vastly important — the idea that an intelligent system’s internal knowledge structure should have some real and direct meaning to the system, grounded in its own real experience.
Proponents of semantic originality argue that true machine intelligence will most likely arise as an emergent property of internal representations that are built through learning. This learning may be “supervised” (wherein the system is purposefully shown labeled training examples). Or it may take the form of evolutionary and reinforcement learning approaches, wherein an agent develops its internal representations through repeated interactions with the world and the rewards and penalties it earns through those interactions.
Now let’s be clear: The “slippability” demonstrated by LaMDA, Dall-E and related deep AI models, while representing an impressive and very useful form of reasoning, does NOT necessarily represent true human-like (or dolphin-like, or dog-like) sentience; certainly not self-consciousness per se. We haven’t yet achieved Artificial General Intelligence (AGI). Nor have we successfully modelled human consciousness on a computer; the “Hard Problem of Consciousness” remains. For now. (Interestingly, one of the leading lights in that area of work is Douglas Hofstadter’s former grad student, philosopher David Chalmers).
However, those who dismiss today’s AI systems as doing “mere pattern recognition” are, in my view, grossly underestimating what can be accomplished with pattern recognition — especially if those patterns are deeply structured, richly interconnected, and learned through repeated interactions with the world.
—
Thanks for reading! What do you think about Dall-E, or LaMDA or other AI systems? Leave a comment below. If you like the article, leave a clap or two. If you’d like to read more from me in the future, please consider Following.
