Rick Rosner On, 32: Artificial Intelligence and New Theories
Author(s): Scott Douglas Jacobsen
Publication (Outlet/Website): Personal SubStack
Publication Date (yyyy/mm/dd): 2026/08/24
In this interview, Scott Douglas Jacobsen and Rick Rosner discuss whether artificial intelligence can derive genuinely novel insights from learned representations rather than merely retrieve stored knowledge. They compare tacit human understanding with neural-network pattern learning, examining multimodal models, mathematical discovery, interpretability, and implicit structure. Rosner argues that AI may encode useful relationships humans have not yet articulated, while Jacobsen emphasizes differences between machine representations and embodied human learning through direct physical and social experience.
AI and the Possibility of New Knowledge
Rick Rosner: I asked AI, “Is the universe made of information?” It offered points for and against. Google AI. I didn’t ask every AI.
I’m sure more sophisticated systems could offer broader takes, but the interesting possibility is that AI systems could contribute to the next successful theory of the universe. That doesn’t mean the theory is already secretly sitting intact somewhere inside an AI. Current models generate and recombine patterns learned from their training data, and they can sometimes derive things that weren’t explicitly stated in the data. Still, they also generate wrong, inconsistent, or completely invented answers.
So an AI could conceivably help physicists find a future theory, perhaps by spotting mathematical relationships, suggesting hypotheses, searching huge spaces of possibilities, or connecting pieces of existing knowledge that humans haven’t connected yet.
Still, something weird and interesting is at play. A sufficiently capable model might contain internal representations that support conclusions nobody has explicitly asked it to produce yet. If somebody asks the right question, or develops the right method for probing those representations, the model could potentially generate a novel and useful inference.
Is that a reasonable angle on this shit? That AI may have the capacity to derive a fuck ton of stuff that isn’t just rote knowledge, if somebody asks it the right questions? It isn’t the same as simply looking something up. You can generalize from the information and statistical structure learned during training.
Digital Representation and Human Learning
Scott Douglas Jacobsen: And it’s polished language formulated into text, but I wouldn’t call the text absolute because it’s based on binary code. Everything running on an ordinary digital computer is ultimately represented digitally, but that doesn’t make the information itself absolute or certain.
A language model is trained on patterns in data and learns numerical representations and relationships that allow it to predict and generate language, with multimodal models that can also include relationships involving images, audio, video, and other inputs.
It’s a different kind of thing from a human organism learning through continuous interaction with the physical and social environment. It’s part of what we do in the sense that language is one of our major cultural products, and AI systems are heavily trained on those products.
But you can’t infer the entire human process that produced those cultural products just by studying the products themselves. It’s neural-network manipulation and representation of those products, with increasingly sophisticated abilities to generalize from them.
Novel Insight and Mathematical Analogy
That’s not to say you can’t get novel insights from it. Humans also work with accumulated intellectual products when doing mathematics. Mathematicians manipulate symbols and concepts inherited from other mathematicians and discover new correspondences, consistencies, regularities, theorems, and laws.
So LLMs aren’t doing nothing. They’re doing quite a bit. But how we do it and how they do it are different, with some areas of functional overlap.
That’s a first.
Tacit Human Knowledge
Rosner: We know a lot of stuff without knowing that we know it. We don’t walk into walls, generally. There’s a lot of shit we don’t do because we have all sorts of understandings about the world that we’ve never expressed to ourselves through internal talk. As we move through the world and interact with it, our behaviour reflects a deep, practical understanding of how our surroundings work, and much of that knowledge isn’t explicitly verbalized.
In the future, AI will probably become increasingly capable of producing what looks to us like insight. Some of that will come from representations learned from enormous datasets rather than from rules explicitly programmed by humans.
Implicit Knowledge in Multimodal AI
Look at image-generation and multimodal systems. They learn statistical regularities involving what flesh looks like, how skin folds, how shadows behave, perspective, depth cues, lighting, texture, and the apparent geometry of three-dimensional objects. Engineers don’t manually program every one of those visual rules. The systems learn representations through training on large collections of images, image-text pairs, video, or related data, depending on the model.
Modern vision-language systems explicitly combine learned visual representations with language models.
That doesn’t necessarily mean the model possesses a human-like concept of three-dimensional space. A system can reproduce or predict many consequences of spatial structure without representing space exactly the way a person does.
So it’s a tantalizing situation: an AI can encode a ton of regularities it cannot necessarily state explicitly, and humans haven’t been clever enough to probe them in the right way.
Probing Learned Representations
You could probably develop methods that ask an AI to examine its own learned representations and make some implicit structure more explicit. Researchers already do versions of this through interpretability, representation analysis, probing, and mechanistic analysis. However, we are nowhere near being able to ask a model to explain everything it has learned about itself.
There might be a big business in that: using AI to analyze learned representations and identify useful patterns that haven’t previously been articulated.
Expertise, Tacit Skill, and AI
Lance knows a ton of shit about rendering flesh and human bodies. If you sat Lance down and grilled him, you could make some of that implicit knowledge explicit. Some of his knowledge is already explicit, but much expert skill is tacit. By asking enough good questions, you can sometimes get an expert to articulate rules he normally applies without consciously stating them.
You could probably do something analogous with AI, although extracting information from a neural network is very different from interviewing a human expert.
Jacobsen: Thank you very much for the opportunity and your time, Rick.
Last updated May 3, 2025. These terms govern all In-Sight Publishing content—past, present, and future—and supersede any prior notices. In-Sight Publishing by Scott Douglas Jacobsen is licensed under a Creative Commons BY‑NC‑ND 4.0; © In-Sight Publishing by Scott Douglas Jacobsen 2012–Present. All trademarks, performances, databases & branding are owned by their rights holders; no use without permission. Unauthorized copying, modification, framing or public communication is prohibited. External links are not endorsed. Cookies & tracking require consent, and data processing complies with PIPEDA & GDPR; no data from children < 13 (COPPA). Content meets WCAG 2.1 AA under the Accessible Canada Act & is preserved in open archival formats with backups. Excerpts & links require full credit & hyperlink; limited quoting under fair-dealing & fair-use. All content is informational; no liability for errors or omissions: Feedback welcome, and verified errors corrected promptly. For permissions or DMCA notices, email: scott.jacobsen2025@gmail.com. Site use is governed by BC laws; content is “as‑is,” liability limited, users indemnify us; moral, performers’ & database sui generis rights reserved.
