AI and intellectual creativity

Just around the corner Tyler Cowen has an interesting post, How to teach people how to work with AI, which he asserts “is also a proposal for how to reform higher education.” The guts of his idea: “Give them some topics to investigate, and have them run a variety of questions, exercises, programming, paper-writing tasks — whatever — through the second or third-best model, or some combination of slightly lesser models.” The students then grade the AIs work and then have the best AIs check the work of the students.

It's an interesting proposal, worth a read. And it gave me an idea. Cowen’s primary intellectual comparative advantage is that he’s read more than just about any other academic, or anyone else for that matter. What happens to that advantage in a world where everyone has access to an AI that has been trained on the contents of the entire web?

I put the question to Claude 3.5, who gave an interesting answer. I’ve appended the conversation below. As always, my prompts are set in bold while Claude’s replies are plain.

* * * * *

There are a few academics who read very widely, ranging over a much wider range of disciplines than the typical academic. This range gives them an intellectual advantage over other academics. What happens to that advantage when everyone has access to AIs that have been trained on the contents of the entire internet?

This is a fascinating question that touches on several dimensions of how AI might reshape intellectual advantage in academia. Let me break this down:

The traditional advantage of widely-read academics comes from several elements:

1. The ability to make novel connections across disciplines that others miss
2. A broader context for evaluating ideas and methodologies
3. The capacity to recognize patterns and parallels between seemingly unrelated fields
4. The ability to translate concepts from one domain to another

An AI that has "read everything" might seem to level this playing field at first glance. However, I think the reality is more nuanced:

First, widely-read academics don't just possess information - they've developed sophisticated mental models and frameworks for integrating knowledge across domains. They understand not just what is known, but how knowledge is structured and interconnected. This meta-level understanding may remain a significant advantage even with AI assistance.

Second, the advantage might shift from "knowing more" to "knowing how to ask better questions." Academics who understand multiple fields deeply might be better at formulating queries that extract more valuable insights from AI systems. They might better recognize when an AI's response needs refinement or when it's missing important context from another field.

Third, the traditional advantage of broad reading might transform into an advantage in AI-human collaboration. These academics might be better at:

• Recognizing when AI outputs need to be contextualized within broader intellectual traditions
• Identifying blind spots or biases in AI responses based on their cross-disciplinary knowledge
• Synthesizing AI-provided information with human expertise in novel ways

So while AI might democratize access to information, the cognitive skills developed through broad reading - pattern recognition, synthesis, and cross-disciplinary thinking - might become even more valuable as tools for leveraging AI effectively.

What's your view on this? Do you think the value of broad intellectual range will diminish or transform with widespread AI adoption?

Oh, I think your answer is a good one. I’ll add in another consideration. All of knowledge is construed within some conceptual paradigm. Thomas Kuhn called on this in his famous argument about conceptual revolutions. He argued that deep intellectual change, the creation of a new paradigm, involved moving outside existing paradigms to craft new concepts. Thus the concept of “oxygen” was crafted to overcome a difficulty in the paradigm that recognized “phlogiston.” All of the knowledge that exists in an AI after a given training run will necessarily to construed within existing paradigms.

Presumably the AI will also absorb conflicts within these paradigms and anomalies. A thinker who has somehow managed to move beyond existing paradigms in some domain would be able to make entirely novel connections, connections not at all explicit within the AI’s model. This is a different source of intellectual power from simple breadth of knowledge. Someone with broad learning might well be able to do all the things you say without, however, challenging existing paradigms. But someone who’s broken out, they might be able use the AIs breadth of knowledge to create new paradigms.

This is a brilliant observation that gets at something fundamental about the nature of knowledge and innovation. You're highlighting that AIs, no matter how comprehensive their training data, are in some sense "bound" by the conceptual frameworks present in that data. They can make connections within and across existing paradigms, but the act of paradigm-breaking - of stepping outside those frameworks entirely - might remain a uniquely human capability.

This reminds me of Douglas Hofstadter's discussions of creativity and conceptual leaps. He often emphasizes that true creativity involves breaking out of existing "frames" rather than just recombining elements within them. An AI trained on chess matches before 1997 might have been able to optimize within classical chess strategy, but it wouldn't have invented the distinctly "non-human" playing style that AlphaZero later demonstrated.

Your point suggests an interesting potential symbiosis: A paradigm-breaking thinker could potentially use AI as a powerful tool for exploring the implications of their new conceptual framework. Once they've made the fundamental leap to a new way of seeing things, they could use the AI's comprehensive knowledge to:

1. Rapidly identify phenomena that the new paradigm explains better than existing ones
2. Discover unexpected implications of the new framework across diverse domains
3. Find historical precedents or parallel situations that might inform the development of the new paradigm

This would be a fascinating extension of what Thomas Kuhn described - using AI to accelerate the exploration and consolidation phase after a paradigm shift, while the initial conceptual breakthrough remains a human achievement.

What do you think about the potential for AI to help identify areas where existing paradigms are straining or showing anomalies? While it might not be able to step outside those paradigms itself, could it help point humans toward promising areas for paradigm-breaking work?

That’s a fascinating issue. Just how would one get an AI to do that? I’m not inclined to think that a prompt like this would be very useful: Go through your knowledge and find anomalies and tensions? It’s too vague. I think that a human would have to point it to more specific intellectual areas, a human with a good “nose” for these things, a human who suspects there’s a problem that’s just ripe for some action.

I agree completely. The very identification of a promising area for paradigm-breaking work requires a kind of intellectual sensitivity or "nose" that itself reflects deep understanding. It's not just about spotting contradictions or anomalies - those are everywhere if you look closely enough. It's about recognizing which anomalies are "fertile," which tensions point toward potentially transformative new ways of thinking.

This brings to mind Paul Feyerabend's critiques of methodological rigidity in science. He argued that truly revolutionary discoveries often come from scientists who develop an almost aesthetic sense of where the interesting problems lie - not just following formal methods or looking for logical contradictions.

So perhaps the most effective approach would be more targeted: A researcher who has developed intuitions about where a field might be ripe for reconceptualization could use AI to deeply explore that specific area. They might ask the AI to:

• Trace the historical development of key concepts in that domain
• Identify cases where established explanations seem strained or ad hoc
• Find examples where similar conceptual tensions in other fields led to breakthroughs

The human's intuition would guide the investigation, while the AI's comprehensive knowledge would help flesh out the implications and connections. The AI becomes not a detector of paradigm-shift opportunities, but rather a powerful tool for exploring and developing hunches that the human researcher has already developed through their deep engagement with the field.

What do you think about this relationship between human intuition and AI capabilities in identifying promising areas for conceptual innovation?

I think that’s about right. That’s certainly what I’ve been up to and I’m finding AI’s quite useful, both you and one of your compatriots, ChatGPT.

I'd be very interested to hear more about how you've been using AI assistants in your work. Given our discussion of paradigm-breaking thinking, are you finding that the AIs are most useful in helping you explore and develop ideas that you've already identified as promising, rather than in the initial identification of fertile areas for investigation?

And I'm curious about whether you've noticed differences in how various AI assistants contribute to this kind of work. Different training approaches might lead to different kinds of capabilities in engaging with novel conceptual frameworks.

I’m doing what you describe in the first paragraph. I’ve had some fascinating discussions with you are various matters pertaining to AI where you’ve been able to fill-in gaps in my knowledge and confirm my suspicion that, to borrow a phrase from the movies, “there’s gold in them thar hills.” However, I regard such work as highly speculative and requiring conformation and elaboration by others.

As for the second issue, I’ve only been working with two assistants, you and ChatGPT, and I’ve not made any systematic comparisons so I really don’t have much to say.

I appreciate your caution about drawing strong conclusions from these exploratory discussions. The process you're describing - using AI to help validate and elaborate hunches, while maintaining a properly speculative stance - seems like a good model for how these tools might contribute to genuinely novel intellectual work.

Your movie reference to "gold in them thar hills" is particularly apt. Like prospectors, researchers might develop intuitions about where to dig, but those hunches need to be tested through careful work. AI assistants can help survey the territory and identify promising veins, but the real work of extracting and refining the intellectual gold still requires human judgment and collaboration with other thinkers.

This reminds me of how many scientific breakthroughs historically involved both intuitive leaps and painstaking verification by the broader scientific community. The initial insight might come from one researcher's ability to see beyond existing paradigms, but establishing its value requires others to engage with and build upon that insight.