Rick Rosner On, 19: Terminal Lucidity, AI Mathematics, and the Future of Scientific Discovery

Author(s): Scott Douglas Jacobsen

Publication (Outlet/Website): Personal SubStack

Publication Date (yyyy/mm/dd): 2026/08/06

In this interview, Scott Douglas Jacobsen and Rick Rosner examine terminal lucidity, paradoxical lucidity, and speculative neurological mechanisms involving contextual and inhibitory coding in the brain. They also discuss OpenAI’s reported mathematical breakthroughs, AI-assisted proof generation, recursive self-correction, high-throughput scientific discovery, and the evolving relationship between artificial intelligence and human researchers. Rosner reflects on AI’s strengths, limitations, and future potential while encouraging Jacobsen to consider traditional publishing for his extensive body of work.

Rick Rosner: I’ve got this other thing, which is terminal lucidity. I read an article about it a couple of weeks ago. It’s a reported phenomenon in which some people with advanced dementia unexpectedly regain clarity or the ability to communicate shortly before death. They may not have spoken coherently in months, and then suddenly they start talking again. They know where they are, they’re able to say goodbye, and they appear to understand their situation.

They might be lucid for 15 minutes, an hour, or a couple of hours, and then die within hours or days. Researchers sometimes distinguish terminal lucidity, meaning unusual clarity near death, from paradoxical lucidity, meaning an unexpected lucid episode in somebody with severe neurological impairment. The phenomenon has been documented through caregiver and healthcare-worker reports, but its frequency and biological mechanism remain uncertain.

One theory about this is that it may involve disinhibition in the brain. It also seems superficially similar to reports of a person’s life flashing before their eyes during a near-death experience. But we don’t know whether those two phenomena have the same mechanism.

I think a possible explanation is that the brain relies on inhibition in order to function well. Here, let me take a detour.

The brain appears to represent information through distributed patterns of neural activity, sometimes described as population or combinatorial coding. It’s not usually the case that every neuron is assigned one specific thing, like, “When this neuron fires, you think of an orange.” It’s more that patterns of neurons firing together represent things, which lets the brain encode much more information.

But I think terminal lucidity and reports of your life flashing before your eyes may also imply something like contextual coding. The same pattern of neural activity could contribute to different meanings depending on what you’re sensing, remembering, or doing at the time. Under normal circumstances, the brain constrains the landscape of possible thoughts. Based on what you’re experiencing, it makes some interpretations and pathways more likely than others.

So it’s not just combinatorial coding. It’s also contextual coding. A pattern of neurons firing could potentially contribute to a bunch of different things, but the brain interprets that pattern within the context of what you’re experiencing.

My speculation is that, when a person’s brain is breaking down for good and they’re about to check out, some of the normal inhibitory controls may weaken. The brain may briefly enter an unusual state in which previously inaccessible pathways become available, allowing a temporary return of communication or lucidity.

That is only a hypothesis. Researchers do not yet have an established neurological explanation for terminal or paradoxical lucidity.

The same kind of loss of normal constraint might also contribute to the feeling that your whole life is flashing before your eyes. Without the brain’s usual filtering and contextual limits, memories or associations might arise very rapidly. There is evidence that life-review experiences occur in some near-death reports, but there is not enough evidence to say that people literally think every thought they have ever had, or that disinhibition is the cause.

That kind of unconstrained mental activity would obviously be bad in everyday life. You wouldn’t want to be driving and suddenly think every thought you’ve ever thought.

Every memory you’ve ever had. That doesn’t lead to safe driving. But in the context of a brain that’s about to go down, it kind of makes sense.

The deal with brains is that our brains are the product of billions of years of evolution. The brain is going to take every informational shortcut it can. Any adaptation that increases the amount of information the brain can process and retain, while also increasing the accuracy and precision of thought, could be favoured by evolution. Any way that the brain can become more efficient, if it’s biologically possible and advantageous, may evolve.

So it makes sense that the brain has a ton of tricks for maximizing information. One of those may involve inhibitory coding.

Jacobsen: Next item: OpenAI has announced results that resolve or make substantial progress on ten longstanding problems. These include discoveries that mathematicians had conjectured might exist but had not previously established, such as the construction of non-sofic groups.

This could represent a turning point in mathematical discovery by artificial intelligence. The work used an internal version of Astra, OpenAI’s next major model, which has not yet been released to the public. According to OpenAI, it produced mathematical arguments addressing problems in areas including group theory, geometry, coding theory, quantum complexity, and extremal combinatorics. The arguments were then prepared as manuscripts by humans working with the model and formalized in Lean certificates.

This means that an artificial-intelligence system can contribute original results to the sum of human mathematical knowledge. Each discovery can also open more advanced problems because the results become stepping stones.

If you iterate that process, artificial intelligence specialized in this kind of work, such as Astra and whatever comes after it, could begin building its own steps along paths of scientific and mathematical discovery. It may explore directions outside established human habits and biases.

It will still inherit some human biases from its training data, architecture, and the problems people choose to give it. But it may also pursue approaches that are not constrained by human memory, speed, attention, or intuition in the same ways.

Rosner: I think I have an analogy, but I don’t know enough about how this works.

I’ve been watching some of this. Chris Cole and, particularly, Dean Inada have been working with high-level AI systems on intersecting-cube problems. There’s a famous problem, to the extent that any of these problems are famous, from Ronald Hoeflin’s Mega Test. It asks for the maximum number of separate regions that can be created by three interpenetrating cubes.

Inada, Cole, and some other people have been poking at versions of this problem involving even greater numbers of cubes with the help of AI. It’s a brutal family of problems because even graphing or picturing the configurations is difficult. I assume something similar applies across many areas of mathematics.

Along with good results, AI will also confidently present nonsense results. So you need competent mathematicians to work through what AI presents.

Jacobsen: I would add that people talk about recursive self-improvement, but that’s almost a buzz phrase at this point. For these mathematical problems, logical proofs, and so on, the work takes place in a world of symbols and abstractions. What’s needed first is something more like recursive self-correction, with systems checking their reasoning, identifying errors, and imposing boundaries on different dimensions of analysis.

That’s a whole other stage of development. Terence Tao has described several levels of interaction between mathematicians and AI, ranging from basic assistance to increasingly autonomous mathematical research. Current systems can perform some lower- and intermediate-level tasks, but they do not reliably carry out the entire process of choosing important problems, developing concepts, verifying results, and communicating them without substantial human guidance.

The work being done across different domains, including language, mathematics, and adaptation to real-world environments, is the next major area of integration. Once that is tied to formal proof systems, and once a lot of the substantial kinks are worked out, you may get something more akin to autonomous scientific discovery.

Rosner: In a more colloquial sense, what you’re doing when you work with AI, apparently, and again, I’m not Terence Tao, I don’t know shit, is giving AI shit.

One way of working is to set up several AI agents and have them compete with or critique one another. An AI generates a result, and then you tell it, “I think your result sucks. Take another look and tell me how it might suck or be wrong.” Or you give the result to another instance of the model and say, “Here’s what we got. Check it.”

So you’re setting AI against itself and having it compete with itself. The critiques can come from humans, other AI systems, or formal proof-checking software. That can improve reliability, but it still doesn’t guarantee that the answer is correct.

Before you get to that, you’ve got the brute-force approach, which preceded AI but can also be used with AI. Take drug testing. In the 1950s, somebody might get an idea, look at the chemistry of a disease, and say, “Huh, I wonder whether this compound would work against this disease.”

More recently, with robotic systems manipulating pipettes, you can test enormous libraries of compounds. You might say, “We have a thousand substances. We don’t have a strong theory about what will kill this microbe or these cells, so we’re going to try all thousand against them.”

That’s brute force. It’s testing many things against many other things. Modern drug discovery usually combines that kind of high-throughput screening with biological theories, computational models, and statistical methods, but the brute-force component is still there.

I assume that, with AI and large datasets, the system can draw on a huge portion of the mathematical literature it was trained on. It doesn’t literally contain every mathematics paper ever written, and it doesn’t necessarily retrieve or understand all of them accurately, but it can connect ideas across an enormous body of material.

Einstein took roughly a decade to move from special relativity toward the completed general theory of relativity. He initially lacked some of the advanced mathematical tools he needed. His friend and former classmate Marcel Grossmann introduced him to tensor calculus and the differential geometry developed by people such as Riemann, Ricci, and Levi-Civita.

It wasn’t that Einstein sat in a café until somebody handed him the right four-by-four matrix. Grossmann helped him identify the mathematical framework needed to describe curved spacetime. General relativity uses the metric tensor, commonly represented as a four-by-four array in four-dimensional spacetime, along with related tensors and differential equations.

What AI can potentially do is be a whole crowd of mathematically knowledgeable friends. It can search across many possible connections, propose tools from different fields, and test large numbers of approaches. In that sense, it can combine broad recall with a kind of guided brute force.

It can try many approaches against many problems. I assume that’s part of why AI is effective at solving some of these problems, but I don’t know. Again, this is a situation where I don’t know shit.

One more thing, though, is that AI may gradually move beyond methods that look mainly like brute-force search, where it finds results by trying huge numbers of possibilities, toward methods that look more like insight. I assume we’re already somewhere along that curve. Current systems can recognize structures, generate useful conjectures, and discover strategies that people didn’t explicitly program into them.

Whether that counts as insight in the human sense is another question. AI has not been shown to be conscious, but you don’t necessarily need consciousness to produce a novel and useful solution. There can also be a search component to what we call insight. A system explores many possibilities, discards most of them, and finds a connection that works.

What you were saying makes me want to offer some advice. All of your adventures seem like they could add up to a great book proposal. You’ve written so many freaking words and conducted so many interviews.

You haven’t really messed with traditional publishing. You’re used to publishing tens of millions of words on your own, but at some point, maybe you move over and try traditional publishing.

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.