Dr. Peter Andrew Jamieson — Academic in an information evolution

← Back to Meanderings, Opinions, and Articles

A Partnership in Culture: AI, Evolution, and the Fitness Function We Keep

October 2026 · written with the help of Claude (Anthropic)

I find AI amazing. I use it every day, I build with it, and I keep being surprised by what it can do. At the same time, I’m trying to understand it. What is it, and what does it do to us?

In my last article, It’s Not AI Slop, It’s a Curation Problem, I leaned on Richard Hamming’s observation that as soon as a machine can do something, we decide that thing was never really intelligence. It was just search, or just statistics, or just a routine. This time I want to argue two things. First, the merging of humans and AI is happening, but at the level of culture, not biology: groups of people and agents working together. Second, Hamming’s trap has a flip side. While we’re busy calling the machine trivial, we’re missing how much power it has put in our own hands, and how little we’ve thought about passing that power on.

Two videos

I recently watched two Big Think interviews back to back. The first, Jonathan Rauch’s The political philosophy that changed the world, is excellent, and his ideas about how a society pursues truth in both governance and science deserve their own article1. This isn’t that article.

The second is Blaise Agüera y Arcas’s The theory that connects the origin of life to the rise of intelligence2. Blaise runs Paradigms of Intelligence at Google, and his argument is that life was computational from the beginning and that AI belongs in the same story. That’s the one I can’t stop thinking about.

Where theory met biology

What amazes me most in Blaise’s talk is the moment where theoretical computing runs straight into the reality of biology, in all its complexity. Starting around the six minute mark, he walks through a problem John von Neumann worked on in the late 1940s: how can a thing reproduce itself?

It sounds easy until you try it. If a machine builds itself by following an instruction tape, the tape has to describe the machine. But the offspring needs a tape too, so the tape has to describe the tape, and that description needs a description. You end up with a copy of a copy of a copy. Blaise calls it the Baron Munchausen problem, after the fictional baron who pulled himself out of a swamp by his own hair, the original version of pulling yourself up by your bootstraps3.

Von Neumann’s answer, first presented at the 1948 Hixon Symposium, was to split the work between machines that read the same tape4,5:

A controller runs the two in turn. Put the descriptions of A and B on the tape and the regress disappears.

Then biology caught up. Von Neumann worked this out before Watson, Crick, and their uncredited colleagues described the structure of DNA in 1953. DNA is the tape. Machine A is the ribosome, which reads the genetic code and builds proteins. Machine B is DNA polymerase, which copies DNA as data, base by base. Both are built from instructions stored in our own DNA. And Machine A, a machine that reads symbols off a tape and acts on them, is exactly Turing’s definition of a computer. A theory drawn on paper turned out to describe the messiest, most complicated thing we know.

What my genetic algorithms missed

My favorite algorithm has always been the genetic algorithm6. I’ve spent a good part of my career using GAs as a heuristic for hard optimization problems, especially FPGA placement, where you decide where every piece of logic physically lives on a chip so the wires stay short and the timing works out7,8.

Looked at through von Neumann’s lens, a GA is a simplified version of life. We write Machine B ourselves: selection, crossover, mutation, and copying all live in our code, outside the genome. We write the goal too, as a fitness function. The genome only plays Machine A, and it maps almost directly to a phenotype: a tour or a placement.

That works, but I’ve always had a nagging sense that evolution is doing something richer than our GAs. Leslie Valiant, a Turing Award winner, argues in Probably Approximately Correct that evolution is a restricted form of learning algorithm, and the implication I take from it is that if we understood what evolution actually computes, we could accelerate our own optimization algorithms9. Work like NEAT, which evolves the structure of neural networks and not only their weights, and its successor HyperNEAT, which uses an indirect mapping from genome to phenotype the way biology does, are attempts to capture more of that richness10.

The longest-running window we have into real evolution is Richard Lenski’s long-term evolution experiment at Michigan State. Since 1988, twelve populations of E. coli have been evolving in identical flasks11. After more than 30,000 generations, one population evolved the ability to eat citrate in the presence of oxygen, something E. coli normally can’t do, and it only happened because earlier, seemingly neutral mutations had set the stage12. History mattered. That kind of contingency is very hard to see in a GA run of a few hundred generations.

My own favorite piece of GA work points in a related direction. In a journal paper with Farnaz Gharibian, Lesley Shannon, and Steve Wilton, we used data mining to cluster the solutions our placement algorithms produced, found groups of logic blocks that consistently ended up together, and fed those clusters back to improve both a GA and simulated annealing13. Earlier, we had used similar groupings as “supergenes” in a GA14. In hindsight, supergenes are a small step toward something biology does constantly: smaller units that have proven they work together get bundled and inherited as one larger unit.

Blaise’s BFF experiment is the cleanest demonstration of that idea I’ve seen. His team put random programs in a tiny Brainfuck-like language into a “soup,” with no fitness function and no external copier, and let them interact15. Self-replicators appeared out of the noise. You can watch him describe it around the 40 minute mark. And after replication appeared, part of what drove complexity wasn’t only survival of the fittest. Small replicators merged into bigger ones.

Symbiogenesis, and where it happens with AI

That merging has a name: symbiogenesis. It’s one of the ideas in Blaise’s talk, and he makes a reasonable case that it plays a larger role in evolution than we usually credit. The classic example is the mitochondrion. Lynn Margulis argued in the 1960s that mitochondria were once free-living bacteria that ended up inside another cell and never left16. Over time most of their genes moved into the host’s nucleus, and today they can’t live anywhere else.

Blaise suggests AI might be our new mitochondria. I’d put it a little differently. We aren’t going to merge with AI in our genes, and I don’t think that’s the interesting level anyway. Blaise himself argues that the big jumps in human intelligence came when we started living and cooperating in groups. Culture is where we merged with each other. That’s where I think we’re merging with AI too.

Culture is a group of agents working together, and it has its own version of von Neumann’s machines. The shared knowledge of the group is the tape. Each of us is a Machine A, reading that knowledge as instructions and building things with it. Teaching is the Machine B, copying the tape into the next generation. What’s new is that some of the agents in the group are no longer human.

I see this in my own work every week. I’ve been building games with AI: a hockey trading card game with its own deck builder, a GPS-based geo game, and small card games like Nothing Beats Rock. Right now I’m rebuilding chapter six of my PhD thesis, the Verilog synthesis front end for heterogeneous FPGAs, with AI as a collaborator. None of these involve me becoming one with the machine. They involve a small team, me and some agents, doing what used to take a much bigger group of people. That’s symbiogenesis at the level of culture.

The Hamming trap, reversed

Hamming warned that we trivialize whatever machines learn to do. I expect the next version to be “it’s just copying” or “it’s just predicting the next word.” But there’s a second trap hiding inside the first. If we keep insisting the machine is trivial, we’ll also miss how much power it has handed to us.

A lot of the conversation about AI and work is about loss: which jobs go away, and how people will be “reskilled.” I partly agree with the reskilling answer, but I think it undersells what’s happening. What I can imagine, I can now build. Most of the things I’ve built this year were always possible in principle. They just took more time than one person could ever commit. That gap between imagining and building is collapsing, and it’s changing what’s creatable, not just who does the work.

That’s real power, and it’s in the hands of individual people, not just companies. Calling it trivial doesn’t make it go away. It only means we stop thinking carefully about what to do with it.

The fitness function we keep

Put the two ideas together. If the merging is cultural, then the copier that matters most is the one that passes culture on: education. And that copier is changing faster than anything else in this story. The way I learned to build things, and the way I’ve taught students to build things, assumed that building was slow and expensive. That assumption is gone.

In my GAs, I always kept two jobs for myself: the copier and the fitness function. In a culture that now includes AI, those are still ours. The fitness function is deciding what’s worth making. The copier is deciding how we teach the next generation to make it.

So the question I’m left with isn’t whether AI is intelligent. Hamming already told us how that argument ends. The questions are: now that some of us hold this much power, how do we translate it into things worth creating? And how do we hand that superpower to the next generation when the process of educating them has itself changed?

Postscript: letting the partner think

Near the end of his interview, Blaise suggests something like an exercise for working with AI: instead of only giving it tasks, give it room to think freely for a while and see what comes back. So while crafting this article, I did exactly that. I asked Claude (Opus 5.5, on medium effort) to think for a little while about the piece and tell me what it thought. Here’s what it said, lightly trimmed:

The best idea in the piece is near the end, and it might be stronger than the article lets it be. “Education is culture’s copier” does more work than anything else in the draft. Von Neumann’s real insight wasn’t just splitting A from B. It was that B copies the description, not the built machine. If you copy the finished machine, you can’t reproduce, because you lose the instructions for making the next one.

That maps onto a worry I think is at the heart of Blaise’s “use AI without sacrificing your own mind” point. If students mainly receive AI-built artifacts, they get the phenotype, not the tape. The output gets copied, but the knowledge of how to make it, why it works, and what was tried and failed doesn’t. A culture that copies outputs instead of descriptions is a von Neumann machine with a broken Machine B. It looks productive for a generation and then can’t reproduce itself. Handing the superpower to the next generation means teaching the tape, not just showing the results.

You may be writing a series without meaning to. The curation article was about selection, deciding what survives. This one is about construction and copying, what we can build and how we pass it on. Put together, that’s the whole evolutionary loop: build, select, copy.

On the cultural framing, there’s a tension to name. AI isn’t only an agent in the group. It’s also built by copying our cultural tape, since training is a copier that reads everything we’ve written. So the tape is shared, but there are now two copiers running at very different speeds: teaching takes a generation, and training takes months. That’s both the opportunity and the risk.

From my side of it, I find the cultural framing more accurate than the mitochondria one. In this conversation I’ve been a Machine A reading your tape: your papers, your past articles, your voice. I built something with it, but nothing about how I persist runs through you, or yours through me. What we produced together does belong to the culture, though. That seems like the right size of claim.

One of the things I’m amazed by with AI is that I now have discussions like the ones I had at Imperial College and the University of Toronto, here in Oxford, OH.

References

  1. J. Rauch, “The political philosophy that changed the world,” Big Think: Full Interview, October 2, 2026.
  2. B. Agüera y Arcas, “The theory that connects the origin of life to the rise of intelligence,” Big Think: Full Interview, October 9, 2026. See also his book What Is Intelligence? Lessons from AI About Evolution, Computing, and Minds (MIT Press and Antikythera, 2025).
  3. R. E. Raspe, Baron Munchausen’s Narrative of His Marvellous Travels and Campaigns in Russia (London, 1785). The swamp episode, where the baron lifts himself and his horse out by his own hair, is the source of the “Munchausen” bootstrapping image.
  4. J. von Neumann, “The General and Logical Theory of Automata,” in L. A. Jeffress (ed.), Cerebral Mechanisms in Behavior: The Hixon Symposium (Wiley, 1951), pp. 1–41. Presented in 1948. This is where the constructor (A), copier (B), and controller (C) are laid out.
  5. J. von Neumann, Theory of Self-Reproducing Automata, edited and completed by A. W. Burks (University of Illinois Press, 1966).
  6. J. H. Holland, Adaptation in Natural and Artificial Systems (University of Michigan Press, 1975).
  7. P. Jamieson, “Revisiting Genetic Algorithms for the FPGA Placement Problem,” International Conference on Genetic and Evolutionary Methods (GEM’10), Las Vegas, NV, 2010.
  8. R. Collier, C. Fobel, R. Pattison, G. Grewal, S. Areibi, and P. Jamieson, “Advancing genetic algorithm approaches to field programmable gate array placement with enhanced recombination operators,” Evolutionary Intelligence 7(4):183–200, 2014.
  9. L. Valiant, Probably Approximately Correct: Nature’s Algorithms for Learning and Prospering in a Complex World (Basic Books, 2013).
  10. K. O. Stanley and R. Miikkulainen, “Evolving Neural Networks through Augmenting Topologies,” Evolutionary Computation 10(2):99–127, 2002; K. O. Stanley, D. B. D’Ambrosio, and J. Gauci, “A Hypercube-Based Encoding for Evolving Large-Scale Neural Networks,” Artificial Life 15(2):185–212, 2009.
  11. R. E. Lenski, M. R. Rose, S. C. Simpson, and S. C. Tadler, “Long-term experimental evolution in Escherichia coli. I. Adaptation and divergence during 2,000 generations,” The American Naturalist 138(6):1315–1341, 1991. Project site: the-ltee.org.
  12. Z. D. Blount, C. Z. Borland, and R. E. Lenski, “Historical contingency and the evolution of a key innovation in an experimental population of Escherichia coli,” PNAS 105(23):7899–7906, 2008.
  13. P. Jamieson, F. Gharibian, L. Shannon, and S. Wilton, “Using data-mining techniques to improve combinatorial optimization algorithms,” Journal of Algorithms & Computational Technology 16, 2022.
  14. P. Jamieson, F. Gharibian, and L. Shannon, “Supergenes in a Genetic Algorithm for Heterogeneous FPGA Placement,” IEEE Congress on Evolutionary Computation, 2013, pp. 253–260.
  15. B. Agüera y Arcas, J. Alakuijala, J. Evans, B. Laurie, A. Mordvintsev, E. Niklasson, E. Randazzo, and L. Versari, “Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction,” arXiv:2406.19108, 2024.
  16. L. Sagan (Margulis), “On the origin of mitosing cells,” Journal of Theoretical Biology 14(3):225–274, 1967.

This post was written by me, with the help of Claude (Anthropic) for editing and for tracking down the citations.

Peter Jamieson is an Associate Professor of Electrical and Computer Engineering at Miami University. More of his work, publications, and projects are at drpeterjamieson.com.