Douglas Hofstadter: Analogy as the Core of Cognition [video]
TL;DR
In an older talk, Douglas Hofstadter argues that analogy-making is the basic act of thought, not a high-level reasoning trick. It runs from perception up to how we form categories in the first place.
The article
Video; not summarized. The tldr is pieced together from the HN thread, and the digest covers the discussion.
The discussion
A warm, wide-ranging thread that mostly used the talk as a starting point, since few commenters said they had watched it. The longest exchange asked whether analogy or prediction better describes what brains do. The rest split between Hofstadter's mixed record against modern LLMs and a meta-argument over whether analogies help or hurt when you argue with someone.
Analogy or prediction as the brain's core job
HarHarVeryFunny ran a long debate claiming the brain is built for prediction, with the cortex's repeating circuit sending predictions down to meet sensory input. Calling that analogy, he said, adds a dense book where one sentence would do. killerstorm answered that "generalization" is vague, and that analogy names a sharper mechanism, comparing things along a chosen feature rather than by raw similarity. A parent's story of a 6-month-old spotting Winnie the Pooh on a towel served as evidence for both sides.
So, my question for Hofstader, or you, would be given that the brain seems to be literally built for prediction, then what is to be gained by regarding this as analogy making instead? Is there any evidence for this?
Analogy-making points to more specific mechanism: ability to identify features within representations, and modulating associative lookup using those features. I.e. it's not as simple as D(x, y) i.e. distance between embeddings, but something like D(f(x), f(y)) where f projects representation to a specific feature space.
Hofstadter's ideas next to LLMs
Several commenters said embedding spaces do analogy, like word2vec's king-and-queen arithmetic, and an LLM feeding its output back in looks like a strange loop. Hofstadter, a long-time skeptic, said in 2023 that LLMs made him feel "diminished", while still calling their intelligence empty. One commenter read that as heartbreak, not dismissal, and antonvs asked why an LLM's intelligence counts as empty when many humans also stop learning young. A side thread on Scott Aaronson's recent post asked whether Hofstadter tied self-reference to cognition or only to consciousness.
I’m confused why there seems to be a dismissal of the most basic ‘strange loop’ of the LLM - the fact that it’s evaluating a context to choose the next word, then reevaluating in a context where that word has been appended.
When he's describing modern systems as "empty intelligence", I think there's a little bit of heartbreak in his perspective, because he sees them as fundamentally different from humans in a way that leaves the part he loves--the "I" in the loop--out of the equation.
Analogies explain well and persuade badly
A remark that sharing an analogy is the fastest way to lose an argument started a long branch. Terr_ said an analogy hands a hostile listener a second set of claims to attack, while dofm called a hole in any analogy a given. jt2190 took the opposite line and asked people to say what a thing is, not what it is like. Others noted that the complaint itself was built on spatial metaphors like "losing" an argument, which is close to Hofstadter's point.
I think that's just because every analogy gives an entire "second front" of ideas for a hostile recipient to find a "flaw", when they ignore the intended boundary between the stuff that does/doesn't matter to the analogy.
The fact of a hole in an analogy is not in itself information. Of course there's a hole.
When a far-reaching analogy earns its keep
gilleain answered patcon's claim that protein folding and storytelling share structure with a 1960s program that "discovered" the atom-as-solar-system analogy, a mapping that teaches little. patcon, trained in biochemistry, replied that he treats stories as one-dimensional chains that fold into a mental model. He ports folding-analysis methods to audience reaction data, and biologists improved their own tools after seeing his. gilleain's caution held for analogies between distant systems, and patcon conceded the value lies mostly in moving methods across domains.
reasoning through analogy is fun and can be useful, it is just limited in what it can do, especially the further apart the systems are.
Someone could get to my conclusions another way, but the metaphors are my cheat codes to arrive there early :)
Older roots of the idea
Commenters traced the thesis back to Nietzsche on the drive to form metaphors, and to George Lakoff and Mark Johnson's Metaphors We Live By. Several recommended that book as the better one on the subject. goatlover noted that if metaphors shape thought this much, an alien species with a different body might think in ways hard to decode. A side story about the Pioneer plaque's arrow, a hunting symbol aliens might not read, made the same point.
Every category and noun is an analogy uniting things unlike in at least some details. We don’t even see most analogies, because they are more common than air and just as necessary.
Consensus
Analogy runs deep in how people think and talk, and modern embedding-based models do something that looks a lot like it. Hofstadter's books, Gödel, Escher, Bach and I Am a Strange Loop in particular, changed how many readers think.
Dissent
The main dispute was whether analogy is the mechanism of cognition or just a way to describe it. The prediction camp argued that brain anatomy and evolution point to prediction, with analogy as a byproduct. A theory that needs that many examples, they said, has not proved its claim.