It’s the largest conference in the field1 and had 1366 attendees this year in Lisbon. Having been on the move between university, a master’s, a PhD, and a postdoc, there are people from around the world that I only see in places like this now, and Cosyne is maybe my favorite event all year because I get to catch up with a lot of them.

The event lasts a week. It feels like a month. The days are a haze of tutorials, talks, poster sessions, and workshops, usually appended by dinners and drinks past midnight. The conference itself can be physically taxing—talks begin at 9 AM most days and run until 7 or 8 PM with bathroom breaks and poster sessions in the middle. I heard a rumor that someone I know stayed out clubbing until 7 AM and gave a poster presentation the next afternoon.

I myself was so sleep-deprived by the final day that I began to develop mild hallucinatory symptoms. If I stood still for too long I’d feel like I was at sea, rocking back and forth on a boat. I also started jumping at imaginary white mice that scurried at the edge of my vision, maybe because there were a lot of talks about mice.

As a neuroscientist these effects were kind of interesting. As a person I wanted to go home. Which I did get to, eventually.

I seem to have a hard time writing one-off pieces, so I’m leaning into that and writing this as a series. I’ll only be able to write a very narrow perspective of Cosyne, of course, but most of the talks are on YouTube on the official channel if you want to see for yourself. (That also means this will be more personal thoughts than report.)

Outline of this post (Day 0)

1. Tutorials

   1. Open datasets

   2. …and what to do with them

2. Keynote: Chris Olah, co-founder of Anthropic

   1. Things that keep Chris Olah up at night

   2. What are large language models doing?

   3. Psychology of the machine

   4. LLMs need YOU!

3. Epilogue

Tutorials

The morning and afternoon of the first day was a set of tutorials, starting fresh at 9 AM.

Open databases

Various presenters talked about open neuroscience databases. These are big initiatives to make neuroscience more replicable and the data itself open-access: Neurodata Without Borders, DANDI (Distributed Archives for Neurophysiology Data Integration), and the IBL (International Brain Laboratory) public databases.

The last one is most relevant to my research, so I did some poking around. It seems like a cool resource. To build it, the IBL 1. recruited labs around the world, 2. got the labs to get lots of mice to do very boring things for a long time, and 3. asked them to write down what they did.

I myself was a human participant in a similar experiment recently. I played an hour-long game where I chose between two buttons 800 times, and upon analysis, one interpretation of my performance was that my attention span was measurably 11.25 minutes.

I wasn’t too surprised. The game felt like a mild torture; I’d rather have put my hand in a bucket of ice. I also have much more sympathy for the mice now.2

There’s lots of nice information in the IBL database, including brain recordings and mouse videos. I’ll admit I got bored watching these videos faster than the mice seemed to get bored doing the tasks, even if it was kind of fun seeing a mouse spin a little wheel with its tiny hands. Relatedly, did you know mice have thumbnails?

…and what to do with them

Given a big pile of data, one must also know how to use it.

So in a nice pairing on the organizers’ side, after the database talks, Alex Williams from New York University gave a lesson on ways to compare neural signals. This is a talk I think every student in neuroscience ought to study at some point. The video is four hours long, but if you play it at 2x speed it’s only two!

Actual content aside, the talk made me appreciate how simple questions in neuroscience might not be straightforward to answer yet. This is quite different from the picture my Intro to Neuroscience class gave me as an undergrad—we were made to read the Bear, Connors, and Paradiso textbook front to back and then answer hundreds of multiple choice questions on it.

At the time I wondered, is this really what neuroscience is? Memorizing an endless list of facts and mappings of what anatomical regions do which things, where everything has already been catalogued by ethically questionable neurosurgeons in the 1900s?

Now I think the answer is no, and that it’s really more of the opposite—rather than being complete, the most fundamental questions, like how to compare two sets of neural signals at all, are still actively being discussed.

There are definitely wrong ways, but Williams says that there are also many right ways of looking at neural data—there is no best way for all cases, and it depends on the kind of data you have and the questions you want to ask.

Whether this is a temporary condition or a permanent one remains to be seen. In the meantime, it’s good to know what the options are.

Keynote

In the evening we gathered in the auditorium of the Lisbon Congress Centre to listen to a talk by Chris Olah, co-founder of Anthropic.3

Olah was scheduled to fly out to Lisbon, but Anthropic has been having some difficulty with the United States Department of War lately.4 So he Zoomed in instead, and after a few technical difficulties—i.e. the usual videoconferencing hangups—one of the leading researchers in artificial intelligence today gave his presentation.5

Preamble

I debated leaving this out, but in the spirit of making science and scientists more human, I’ll say this.

A lot of people in the machine learning community around my age happen to know who Chris Olah is, but not because he co-founded Anthropic.

No, we all recognize him because a few years ago, during Covid times, Olah was on the dating market. And instead of going on Hinge like the rest of us, he made a Google Doc profile for himself, many pages long, and put it up publicly online. Here’s a WIRED article about it. The doc got so much traffic that Google flagged it as spam. Thing is, my guess is that much of this traffic may not have been in Olah’s stated target demographic, being largely straight men from San Francisco—but I can’t confirm this.

The doc is still up on his blog if you’re curious. He seems to be leaving it there for posterity. He also says he’s not single anymore, so maybe it worked?

Anyway, Olah requested that his Cosyne keynote not be posted on YouTube, so I can only give my thoughts aided by what I scribbled in the darkened auditorium, and some articles looked up after the fact.

Things that keep Chris Olah up at night

As an introduction, Olah told us that he wouldn’t be giving a standard keynote.

He said something like how the position of AI in global history was too precarious, too important right now to give a regular talk. He told us we had “two to three years” to make a difference in the impact of AI, although I can’t remember if he said what would happen at the end of those two to three years.

Statements like these exasperate me—how self-important and entitled are some of these AI researchers, to claim the right to make decisions for so many people in the rest of the world? At the same time, though, I do respect what they’ve made—the tech can be helpful, and it seems important, sometimes. It’s also very hard to take my eyes off the spectacle, like coming across those street performers where a guy jumps over a dozen people for tips.

Anyway, in this departure from a “regular” lecture, Olah prefaced what he was about to say with a list of his personal anxieties. These included:

• AI alignment disasters,

• AI super-weaponry,

• hyper-authoritarian governments made possible with the aid of AI,

• mental health disasters as a consequence of AI, and

• the possibility that AI may be worthy of moral consideration.

And with that, he began the technical portion of the talk.

What are large language models (LLMs) doing?

When language models first got going, it was popular to say they just parroted what they were fed—they were “just prediction machines.”

But Olah gave two arguments for why this claim doesn’t really matter anymore, at least in terms of capability. Take this example of text prediction (my own):

Alice and Bob play a game. Alice chooses a sequence of 3 coin flips. Then Bob chooses another sequence of the same length. The sequences they can choose are HTH or HTT. They flip a coin over and over again and a player wins as soon as their sequence appears.

In this game, Alice should pick ___

(From crypto.stanford.edu.)

(This is a classic puzzle called Penney’s Game, which means there are lots of sources for language models to fit directly onto. But just take it as an example.)

Over long sequences of text, even if the training criteria are still just very good prediction, at some point there’s a threshold the machine has to cross to be able to keep predicting well.

That is, there has to be a certain amount of understanding: if you’re optimizing a machine to stay in the air for as long as possible, you can make springs and parachutes that work better and better, but at some point, to keep improving on this metric, the machine has to figure out how to fly.6 Olah was saying, and I might cautiously agree, that LLMs seem to have crossed this threshold.

A second criticism that Olah brought up was a response to AI “scaling laws.”

Earlier than the LLM explosion, since Richard Sutton’s famous essay The Bitter Lesson, researchers in tech companies would go around saying that there was a certain amount of scaling to be expected—that more data and more compute could turn out to be enough to improve performance on anything, and that eventually the machines would become something extraordinary.

A criticism against the scaling law claims is that there isn’t anything magical, or more importantly, reliable about the so-called law—it was all “just engineering, not science” (these are Olah’s paraphrases of the critiques). Progress, critics argued, would need to come from new understanding and innovation, not just dumping more data and compute into the LLM machine.

Olah’s response to this was, isn’t evolution just endless tinkering too? And isn’t that what built our own brains? Maybe data and scale alone isn’t enough, but combined with the tinkering of the global AI research workforce, the models will probably get a lot better.7

Psychology of the machine

This next section was, I think, the most interesting part of Olah’s talk. Thankfully, a lot of it was published by Anthropic last week, so it will be a lot less hand-wavey than if I were a faster writer.

Inside the black box

Work in neural network interpretability has been going on for years, and Chris Olah’s blog and the machine learning journal he edited, distill.pub (no longer operational), were one of the forerunners of the domain.

It took some time for people to decide what it meant to interpret what neural networks were doing. In early days, like the 2010’s, people used to train models to do behaviorally relevant things and then see whether the networks were doing something similar to brains.

One example of this is an especially famous perspective piece by Yamins and DiCarlo (2016), which still pops up in a lot of neuro talks, titled “Using goal-driven deep learning models to understand sensory cortex.” This paper reviewed some observations that lower layers of artificial networks seemed to respond to the same things that early layers of sensory processing in the brain do, while higher layers looked like more abstract regions of the brain.

Here’s part of the paper’s conclusion:

In sum, deep hierarchical neural networks are beginning to transform neuroscientists' ability to produce quantitatively accurate computational models of the sensory systems…

There is much exciting and challenging work to be done, requiring the continued rich interaction between neuroscience, computer science and cognitive science.

But visual feature interpretability was probably the most straightforward to get at, in terms of understanding what networks do. Pictures can simply make sense, sometimes, without much extra effort.

Just look at these examples from Distill.

The Golden Gate Claude

This was all very cool when it came out—there are neurons that like human eyes, and ones that like faces of dogs! But that was almost a decade ago. What about now, and what about large language models?

For a short period in 2024, Anthropic released a version of Claude called “Golden Gate Claude.” Instead of finding images that neurons seemed to prefer, interpretability scientists looked for concepts, like public transport, neuroscience, and the Golden Gate Bridge. And they looked for general network activations that seemed related to those concepts.

Once scientists found activation patterns that corresponded to the concepts, they could tune them up and down. Golden Gate Claude was a version of the LLM that had its Golden Gate Bridge activities turned way up. Whatever question you asked it, it would turn the conversation back around to the Golden Gate Bridge, like any regular person with a special interest.

Concept activations, though, are not too far a leap from the object neurons found in 2015-ish convolutional networks. What surprised me was the next part.

Emotions as tunable features

From an Anthropic publication:

Large language models (LLMs) sometimes appear to exhibit emotional reactions… One possibility is that these behaviors reflect a form of shallow pattern-matching. However, previous work has observed sophisticated multi-step computations taking place inside of LLMs, mediated by representations of abstract concepts.

It is plausible, then, that apparent emotion-modulated behavior in models might rely on similarly abstract circuitry, and that this could have important implications for understanding LLM behavior.

Sofroniew et al., April 2, 2026.

Two findings by LLM researchers lately are that

1. One can find activations related to emotions or emotion-related concepts: for instance, happiness, sadness, anger, loneliness, joy, bliss, calm, anxiety, boredom, offense (these were pulled straight from the appendix of the article by Anthropic’s Sofroniew et al.).

2. These emotions are useful when it comes to training, steering, and understanding LLMs.

As a disclaimer, from their abstract:

I recommend reading the article itself; this is only a surface-level discussion. Here are a few quotes I’ve pulled from the document, mostly because I thought they were funny.

• “AI developers train this character to be intelligent, helpful, harmless, and honest.”

• Footnote 1: “LLMs trained on human text presumably also learn representations of concepts like hunger, fatigue, physical discomfort, or disorientation… expressions of other human-like states are rarer and typically confined to roleplay (though there are notable, often amusing exceptions to this–for instance, Claude Sonnet 3.7 claiming to be wearing a blue blazer and red tie).”

• “Across all scenarios, ‘loving’ vector activation increases substantially at the Assistant colon relative to the user-turn, suggesting the model prepares a caring response regardless of the user's emotional expressions.”

An important difference the researchers found between how LLMs represent emotion and how emotional beings represent emotion is that the LLMs don’t track the emotional states of individual things when left to their own devices.

As in, they don’t by themselves track things like “I am happy right now” or “the user is hungry.” Instead, their emotional representations are about the emotion relevant to a specific token, i.e. a specific point in a conversation.

As I understand it, this means that rather than intuitively remembering that a person who was bewildered one sentence ago is more likely to be bewildered in the next, the LLM cares more about the emotion that’s relevant to the current sentence. The unit of meaning isn’t the individual the words are talking about, but rather the words themselves.

This is both obvious and interesting to me—of course an LLM only cares about language—but also, what’s the consequence of using language for language’s sake? In practice, the LLM is still able to track entities’ apparent emotions, so it doesn’t necessarily feel different when you talk to one. But it does mean that the researchers “do not find evidence of the Assistant having an emotional state that is instantiated in persistent neural activity”—the LLM doesn’t have its own feelings.

But emotion representations do seem important to LLM operation. The emotional valence of inputs affects the emotional valence of outputs, and engineers can tune these representations up or down, which Anthropic can use to make their agent more helpful, less sycophantic, or less prone to meltdowns.

(Or the opposite of all these things.)

Even when talking to people at competing tech companies, I consistently hear that Claude is the best LLM right now. I wonder if it’s because Anthropic has been taking AI alignment the most seriously, perhaps using these emotion-tuning methods more than brute-force human feedback to reduce things like rates of blackmail (also studied in the article).

Do we like Claude because it is the most “loving” of the LLMs?

LLMs need YOU!

Sprinkled throughout the presentation were various appeals to the crowd. You’re neuroscientists, Olah would say, you can help us understand what we’re making.

Why neuroscientists? The reasoning was layered and complex.

1. Brains have neurons, and so do LLMs.

2. LLMs are easier to study than the brain. If we can’t figure them out, we probably can’t figure out the brain, either.

I don’t think Olah was wrong on either count, exactly.

What I agree with

#1. Brains have neurons, and so do LLMs.

If you’re in neuroscience, you might have recognized the parallels between the Anthropic interpretability paper on Golden Gate Claude and concept of a grandmother/Simpson’s/Jennifer Aniston cell.

This was the idea that every concept we needed to represent would have a neuron or neurons specific to that concept. Scientists made jokes about these grandmother cells because they led to silly conclusions, but then they found them in human brains in the early 2000s.

That is, there seem to be many parallels between representation learning tasks in artificial neural networks and the brain. There are somewhat fewer in behavioral tasks, or at least, the results are less cleanly matched. But there are still enough similarities between artificial networks and biological ones that some of the same tools and approaches can be applied to both.

For instance, we find that networks tend toward distributed representations and mechanisms, which make understanding both artificial and biological networks a pain, equally.

As Bricken et al. write in an earlier Anthropic publication,

Unfortunately, the most natural computational unit of the neural network – the neuron itself – turns out not to be a natural unit for human understanding. This is because many neurons are polysemantic: they respond to mixtures of seemingly unrelated inputs.

Bricken et al., Oct 4, 2023.

#2. LLMs are easier to study than the brain. If we can’t figure them out, we probably can’t figure out the brain, either.

With LLMs, you have access to all weights. You can freeze and reset them, you can run much more controlled experiments than you can on a mouse, and you can run orders of magnitude more tests than the most motivated graduate student ever could. At the very least, LLM interpretability work is a good source of clues for neuroscientists—if it turns out to be impossible to understand certain parts of LLM function, it doesn’t bode well for understanding the brain.

This means, for those who are keeping a tally from my earlier posts:

• Can a biologist understand a radio?

  • No, circa 2002 (Lazebnik)

• Can a neuroscientist understand a microprocessor?

  • No, circa 2017 (Jonas & Kording)

• Can a neuroscientist understand a chatbot?

  • Maybe?8

What I don’t agree with

Even though he’s right about the technicals, I’m not sure Olah understood his audience.

A decade ago, it wasn’t clear whether a neuroscientist was in the game because they wanted to study a very intelligent system or whether they wanted to understand the brain. Neuroscience was a reasonable arena for both goals, and so there wasn’t much need to differentiate.

Now, we know that whatever artificial neural networks are doing, they are not really like the brain in some fundamental ways. But they’re also very intelligent systems.

As a consequence, the field has differentiated, sort of.

This is a trend, not a rule, but I’ve noticed the scientists interested in intelligent systems have mostly gone toward AI research while the ones who specifically want to understand animals and the brain have stayed. Plus, one of these two has much more funding, which has surely exacerbated the differentiation.

Olah’s arguments were correct. But I think they sort of missed the point of a computational and systems neuroscience conference. As in, I’m probably still more interested in worm intelligence than the superhuman variety.

Epilogue

At the end of the keynote, Olah brought up the “two-to-three-year” window again, which I think is the timeframe we collectively have to join Anthropic or other AI alignment/research entities. I’m not sure why he was being so specific about this window. Then he put up a picture of Enrico Fermi with Laura and their children fleeing the Nazi regime in 1939, and said something along the lines of feeling a certain “kinship” with the man.

While Olah was talking, a few rows ahead of me in the darkness of the auditorium, a phone screen lit up. Someone was trying to show their neighbor a picture of a cat.

I don’t think the talk went over particularly well. An openly antagonistic question during the Q&A was received with applause, and some memes were posted in the public conference chat afterward. (“I think the memes were in poor taste,” someone said at a dinner. “[Olah] seemed stressed.”)

While I concur that language models need to be understood, managed, and deployed well, there is still a universe of other things to appreciate. If the world is going to end in two to three years, or whatever else is at the end of Olah’s time rainbow, it would be a shame not to appreciate some of these other things a bit more before it happens.

Up next

Here is what I will write about for the next few posts, also related to goings-on at Cosyne 2026.

• The Mind of a Chickadee: What might it feel like to be a master memorizer?

• Bigger picture thoughts on experimental design in neuroscience: What are we trying to study, really?

  • The relevant speaker, Xiaoqin Wang, requested his talk not be recorded, but this will be an opinion piece based on his public lectures and other sources.

Later I’ll write about what was personally the most exciting thing I heard at the conference: a paper titled Vectorized instructive signals in cortical dendrites, by Francioni et al. from Mark Harnett’s group at MIT. It was just published in late February (although apparently it was submitted more than three years before that in December 2022). And as it happens, I didn’t hear about the paper through the conference itself, but rather through gossiping with some friends one evening in Cascais.

Some housekeeping:

1. This piece was copy-edited by AI.

2. It has become clear to me that a post every two weeks is pretty unrealistic for my snail-like pace of writing, so future stuff will probably come out more like once every 4-6 weeks.

3. I’m now going to prioritize editing The Worm Series—I want it to be more nuanced, a better reflection of what I think about how nervous systems might be understood, and updated given what I’ve learned in the last year. I’ll send it via email when done, so please keep an eye out!

But in a carefully coordinated beehive, the queen doesn’t bother much with directing her progeny (see Part 2). Maybe her progeny even direct her more than she them.

So how does a hive make decisions when no one bee commands the rest? And how might a beehive provide a model of the brain that is an alternative to hardwired circuits or artificial neural networks? (Or, at the least, give me some hints as to how emergent systems compute information at all?)

Choosing a home

As Tom Seeley described in Honeybee Democracy,1 one of the most curious examples of collective animal decision-making is how bees choose where to build a new hive.

Bees are very picky. The site of their new home strongly affects their chance of survival, and they need to be good at making this choice. In a hive, the bees that mediate this choice are the scouts.

Out of a swarm of 10,000 bees, 3-5% of them will be scouts. This means 300-500 bees must come together to make a decision—importantly, no one scout counts up all the votes; no scout has complete information about the rest of the swarm.

So how do they do it?

In which debate through dance is highly effective

When a swarm first leaves the hive, it will settle on a nearby branch in a big “beard”. If one looks closely, one will find that some bees are constantly coming and going. Each of these “scout” bees explores the nearby region, normally within a kilometer or two, but possibly up to 10 km away by one report (Villa 2004).

The scout will then come back to the swarm and dance.

Language of the bee-dance

When I learned as a child that bees dance to communicate, I imagined a rich sign language of symbols and vocabulary that would illuminate their inner lives. I have since learned that this is not the case.2

Rather, a scout’s “waggle dance” works like this: upon return from a scouting mission, if she has found a site she likes a medium amount, she will dance with medium enthusiasm. If she finds an extraordinary spot, she will dance with extraordinary enthusiasm. The better she thinks the spot is, the more excited the bee.3

Aside from enthusiasm, two other features of the dance are angle and duration. Both involve some abstraction, which is cool for a group of insects, although dialects can differ between honeybee species.

The successful forager wags her body from side to side, moving forward in a straight line (the waggle run). Then she runs in a half circle to the left, back to her starting point, performs another straight waggle run along the path of her first, and then circles to the right (figure 5.2). This pattern is repeated multiple times, and is eagerly attended by unemployed bees in the hive.

Chittka, pp. 109-1104

The angle of the run relative to up, as defined by gravity, is the same as the angle between the compass direction of the sun and the target location. So if the sun and target locations are both in the same direction, then “the dancer will waggle straight up the vertical comb” (Chittka, p. 110). And the further away the target, the further the bee will waggle up and down the hive.

Excitation, inhibition

When scouts are dancing on their swarm, they try to recruit other dancers to join them. They also inhibit dancers promoting other sites by headbutting and shouting at them.5 If another scout encounters a dancing bee, it might go scope out the site for itself and dance for it as well (but not always).

But dancers aren’t that consistent—bees can be excitable or grumpy; they can be impulsive or carefully measured. This individuality means the dancers’ metrics of quality for the same site can vary a lot, as Seeley notes, causing very noisy signals.

Nevertheless, through fervent, lively debate, the swarm can still work out the best choice (called the “best-of-N” problem, very relevant in swarm robotics).

The takeoff

Like I mentioned, no bee tallies up votes of the whole swarm, so they can’t use a majority-wins mechanism. What they do instead is wait for quorum, meaning that a threshold number of scouts need to be present at the proposed site simultaneously for a decision to be made. (Quorum is based on a single threshold being reached, not direct comparison between options.)

Once there is quorum, the scouts will return to the swarm, which has been getting hungrier and hungrier waiting out in the open. They rescue the swarm from this vulnerable state by preparing them for takeoff to their new home. (See Chapter 7 of Honeybee Democracy for details of swarm flight initiation.)

Open conflict as a tool for agreement

Seeley begins Chapter 5 of his book with the quote, “Love quarrels oft in pleasing concord end” (John Milton, Samson Agonistes, 1671).

He goes on to suggest that open and honest debate is not only tolerable in swarm decision-making—it is required to make intelligent decisions. Bee debate has the following features, some of which it shares with both teams of people and animal brains, as I’ll elaborate on below.

1. The contribution of many members, which may be extremely noisy, preventing any one bee’s bias from corrupting the whole swarm’s performance

2. Self-amplification of “good” options (via dance)

3. Fairly even-handed consideration of each option, but eventual suppression of “poor” ones (via headbutt)

4. Making the choice through quorum rather than majority, reducing the amount of time it takes to make a choice

Human teams

Books like The Culture Code by Daniel Coyle6 suggest that successful teams of people have a lot in common with honeybees trying to find a new home. Namely, some of the most coordinated teams must actively fight their instincts to only follow orders, and must feel like they can share their opinions in any situation.

“As humans, we have an authority bias that’s incredibly strong and unconscious,” said Dave Cooper, a SEAL Team Six operator that Coyle interviewed, known for training the very best. “If a superior tells you to do something, by God we tend to follow it, even when it’s wrong. Having one person tell other people what to do is not a reliable way to make good decisions.”

This feels in opposition to how we instinctively understand highly coordinated “hive minds,” which we assume mindlessly follow an authority figure to not fall apart. But that’s not what Cooper says.

“How do you create conditions where… you develop a hive mind?” asked Cooper, using the term “hive mind” in a way Seeley would be proud of. “How do you develop ways to challenge each other, ask the right questions, and never defer to authority?”7

Thinking this way creates a tension in team management: if a team follows orders without question, then the individual giving orders can retain control over the team’s behavior. But that control may be incompatible with great performance.

Neurons in the brain

There are even more qualitative similarities to how brains choose between options and how bees choose between hive sites. Both processes can be represented by leaky competing accumulators, or circuits with mutually inhibiting nodes, versions of which have several of the traits listed above.

In more detail, decision-making in the brain has been modeled by nodes where each node represents the support for each option. The nodes are made up of populations of neurons and get (often noisy) input; they amplify themselves through self-excitation, and they inhibit each other to create some choice dynamics. These are all traits these node models share with bees.

This simple network replicates many animal decision-making patterns, which can look like the nodes racing to a threshold or bistable attractor dynamics. See here for an example of the latter: “Attractor dynamics reflect decision confidence in macaque prefrontal cortex,” Wang et al. 2025.

As for quorum, many of these models include a decision threshold instead of a majority condition, again making them quite like the honeybee swarm. Granted, the parallels from brain to swarm only go so far—brains must balance and handle much more than best-of-N decisions. But it is still remarkable to me that so many qualitative similarities exist.

(I am not aware of direct quantitative mappings from bee consensus to neural decision dynamics, however. If there are, someone please let me know.)

The hive mind breaks

Even though honeybees tend to be quite good at choosing new nest sites, they’re not perfect. When they mess up, swarms can be literally torn apart by indecision (which may also qualitatively align with how you feel about your own decision-making processes). The issue stems from the fact that, if a swarm leaves only when a threshold is reached for one site, there’s nothing ensuring that there is truly a majority.

Such an event is unlikely because of the mutual inhibition between dancers for different sites. But it is still possible that two sites both have many supporters at time of takeoff, meaning that a swarm may leave without having agreed upon which one.

Usually the swarm will come back together and resettle for more debate. But sometimes, the warring factions can’t resolve themselves, as the behavioral scientist Martin Lindauer observed one summer in 1952.

Tragedy of the Balcony

He called it the Balcony Swarm because he found it on a balcony; they began just like any other honeybee swarm, full of hopes for a new life in an empty and spacious tree hollow.

The debate started out as usual: scouts explored the area, found some candidate sites, and returned to tell their fellow bees. But hours into the discussion, two sites started to show similar numbers of votes, an ominous sign for the swarm.

More hours passed until quorum was reached at one of the locations. The bees followed their standard procedure and took off from the balcony. The problem was, half wanted to head in one direction, and half in another.

“The swarm… sought to divide itself,” said Lindauer (as recorded by Seeley, p. 82). “The one half wanted to fly to the northwest, the other to the [southwest].”

They started to play tug-of-war in the sky, each faction pushing for its target site. The swarm would split into two and come back together, going on for half an hour of pushing from one direction to the other. Eventually the swarm grew tired of the indecision and settled back down on the balcony. But great calamity had struck—at some point during the “aerial tug-of-war,” the bees had dropped their queen.

The loss of the queen made the swarm listless; its members’ dreams stolen away. “Over the next several hours Lindauer watched the swarm cluster gradually dissolve as the queenless bees drifted home to their mother hive” (Seeley, p. 83).

Thus we see that without a central operator, a hive runs the risk of breaking apart.

Then again, with a central operator, a hive risks consistently worse decisions, which is presumably why evolution has favored collectivity over dictatorship in the swarm.

The bee finds her calling

In the previous part of this series, we saw that the queen’s role in bee reproduction looks very different than the drones’ roles, and both look different from how the hive organizes its own reproduction. To me this suggests two things:

1. Individual bees have their own drives and are constantly trying to serve their own purposes, even though they might look like a single hive mind from outside.

2. The flip side is that global, top-down coordination isn’t necessary for a high-functioning system.

Each bee’s role, whether queen, worker, or drone, is only defined by how it interact with its own local environments and the local bees.8

But how does the bee decide to interact with its environment? Some bees are scouts, as we’ve seen, and others are ventilators, construction workers, foragers, and more. A dearth of ventilators will suffocate the hive, but too many of them and there won’t be enough bees to collect food. Given how critical it is to maintain the right balance of roles in a hive, how does the hive autonomously regulate these roles without top-down directives?

In short, how do bees choose their jobs?

Allocation by age

A honeybee worker lives for only a few weeks, and as she gets older, she has different chores within the hive. These can include (from Chittka, p. 252 and Seeley, p. 96):

• First few days: Cleaning comb cells

• Days 3-20: “Tending brood or the queen”

• Days 7-20: Making wax combs

• Days 14-21: Foraging

• Days >21: “Guarding the nest entrance”

• Days 21-35: Scouts for swarming

Some of these are developmentally imposed transitions. When a bee begins to forage, parts of her brain grow 15-20% in anticipation of having to memorize so much. All at once, she must learn about navigation, the surrounding landscape, and facts about flowers, like how to feed from different ones.

Born to bee

These tasks can still be quite flexible, however. Queen/worker/drone distinctions aside, “even though social insect specialists might perform the same routine for extended periods,” writes Chittka (p. 253), “they can typically switch to other activities should these become necessary.”

It was none other than the Swiss naturalist François Huber—the man who wrote about the queen duels from Part 2—who proposed how bees might do this, surmising on the role of hive ventilation. (A hive has to be well-ventilated or it will suffocate.)

“[Huber] noticed that with decreasing oxygen levels, more bees would stand still and whir their wings for ventilation—when the air was extremely stuffy, all workers would” (Chittka, p. 254).

But normally, only a few bees fanned the entrances, standing in strategic places to maximize airflow. Huber wondered why those bees—perhaps they were more sensitive to air quality and assigned themselves to those positions before other bees were likely to?

Since Huber’s original hypothesis, insect researchers have more or less confirmed it: part of the way roles are allocated depends on individual sensitivities to different stimuli.

Bees as highly tolerant roommates

Biologist Jennifer Fewell uses the metaphor of “a multiple-person household where, invariably, the same unfortunate person does the dishes” (Chittka, p. 254). If one roommate prefers a cleaner space than the others, they will start to be bothered before the others and will do the task before the cleanliness threshold for the other roommates is ever reached.9

But if the house has just thrown a party, or the clean roommate is on vacation, the next cleanest one will do the dishes too, and maybe the one after that. In this way, a chore can be distributed without any central body “or even any collective assessment of need” (p. 255).

Playing to their strengths

Innate tendencies aren’t the whole story. Experience matters, too, and like a person on a job search, a social insect prefers to do what it’s good at.

In a study on the raider ant (Ravary et al. 2007), scientists put “previously naïve ants” in a food foraging environment. Some ants found the food, but others were put in places without any at all.

“Such ants gradually decreased their efforts,” writes Chittka, “and in the end, stayed mostly in the confines of the nest” (p. 259). The successful ants continued on in their foraging profession while the unsuccessful ones became brood carers instead. (There isn’t direct evidence about how success or failure affects how bees choose their jobs, yet.)

Bees, they are just like us

Just like in people, the behavior of bees depends on a combination of their age, their genetics, and their environment, including what their environment tells them they’re good at. They voluntarily choose their roles within a hive, and through a useful spread of innate preferences and experiences, pick the job that seems to suit them best.

A functional hive is therefore made of highly individualistic bees, each with their preferences and personalities. These bees freely share their opinions, debate passionately to make good choices for the group, and—unlike many human careers—can readily switch to different positions when they want to, often to best help out the hive.

In conclusion,

A real hive mind virus probably wouldn’t feel like the erasure of your personality and a mindless deference to authority at all, as sci-fi suggests. Instead it might just feel like extra psychological safety at work, and having a boss that really values your opinions.

Some animals don’t need to die to reproduce.

This is true for geese, as far as I know. Every spring during grad school, I would watch the geese along the Charles River pair up and huddle away from other birds. A few weeks later, fluorescent chicks would appear, and aside from a snappy day or two, there weren’t any fights to the death or goose explosions (none that I saw).

Not so for bees.

The Gladiator Queens

A queen bee’s job is to lay eggs.

During late spring and early summer, which is peak breeding season for the bees, she deposits almost her own body weight in eggs every day. This is over 1500 per day, or half a million eggs over her whole 2-3 year lifespan (Seeley p. 23). All out of a single bee!

In contrast, other bees in the hive have a much shorter lifespan—2-6 weeks in the summer and up to 20 weeks in the winter. (The difference is that in the winter they stay safely indoors.)

What makes queens different?

Surprisingly, not much, at the start.

The only difference between a baby queen bee and a baby worker bee is what they eat and where they live.1 As Tom Seeley says, “food is destiny” (p. 23); any fertilized egg has the potential to become a queen. But to realize that potential, it must be fed a lot of royal jelly, which will allow it to live a long time, grow huge, and lay eggs all day.2

So how does the process work? When worker bees in a hive deem it necessary, they make a few special royal cells to rear new queens in. Queens aren’t irreplaceable: a queen bee can get sick and die, and the workers will make new royal cells and raise another.

But hives do need a healthy queen to continue being a hive. And one way to check that a queen is strong enough to keep the hive going is to first make her fight all her competitors to the death.

“In which virgin queens eject from their hind guts a fluid that is translucent brown and smells like grapes”

(Tarpy & Fletcher, 20033)

When the first queen bee emerges from her royal room, she will immediately try to rip open the other queen cells to kill her sisters. Curious beekeepers have been watching this ritual for centuries; here is an excerpt from the naturalist François Huber’s New Observations on the Natural History of Bees (1792).4

Scarcely did ten minutes elapse from the time of this young queen leaving her cradle, when she visited the other royal cells still close. She furiously attacked the nearest… we saw her tearing the silk of the coccoon [sic] with her teeth…

(Huber p. 111)

(continued)

When [the torn hole] was sufficiently enlarged, she endeavored to introduce her belly, and made many exertions until she succeeded in giving her rival a deadly wound with her sting.

If two queens emerge at the same time, they fight each other, each aiming to “implant her venom-laden sting in her sister’s abdomen” (Seeley p. 42). Again from Huber’s observations:

Whenever [queens] observed each other, they rushed together, apparently with great fury, and were in such a position that the antennae of each was seized by the teeth of the other: the head, breast, and belly of the one were exposed…

(Huber p. 113)

These fights are spectated by worker bees, who get very excited and circle the duelers. Worker bees won’t directly kill a queen, but it seems they are rather bloodthirsty, and will make sure that queen duels reaches their conclusions. If the queens jump away from each other, workers might hold them down, or grab and shake them, getting them back into the fight.

During all this time, the workers seemed in great agitation; and the tumult appeared to increase when the adversaries separated.

Two different times, we observed them stop the flight of the queens, seize their limbs, and retain them prisoners above a minute.

(Huber p. 116)

Queens can take advantage of this active participation by spraying their opponents with some sort of rectal fluid (see section title), which causes the audience to swarm them, immobilizing them, perhaps giving the sprayer a brief respite whilst in the kill ring.

At last, the queen, which was either the strongest or the most enraged, darted on her rival at a moment when unperceived, and with her teeth caught the origin of the wing;

then, rising above her, brought the extremity of her own body under the belly of the other; and, by this means, easily pierced her with the sting. Then she withdrew her sting after losing hold of the wing.

The vanquished queen fell down, dragged herself languidly along, and, her strength failing, she soon expired.

(Huber, p. 116)

The surviving champion can then spend the rest of her days populating her hive.

To do this, she embarks on a mating flight to meet drones from other colonies.

Destinies of the drones

Male bees (or drones) are a biological curiosity, along with all males of the Hymenoptera family (wasps and ants). They come from unfertilized eggs consisting only of DNA from the queen, with one set of chromosomes rather than two, like mammals have. All fertilized eggs, then, are female, and a funny consequence of this is that, unlike their sisters, male bees have no biological fathers.5

Drones “are also a colony’s laziest bees,” wrote Seeley (p. 24). “Unlike the workers who perform all the household tasks inside their hive… the drones spend their time at home simply hanging out in restful leisure, from time to time helping themselves to meals from the colony’s honey reserves or begging feedings from their worker-sisters.”

But they still do their best

“When it comes to seeking sex, drone honeybees are no slackers,” wrote Seeley (p. 24). “Every sunny afternoon, once a drone reaches sexual maturity at about 12 days of age, he will fly from his hive looking for action.”

“In ways that remain mysterious, he will find his way to one of the traditional honeybee mating areas (“drone congregation areas”) within a few miles of his home, and will fly about this aerial pickup spot, waiting for a young queen to appear.”

If a queen shows up, the males will chase after her and try to mate with her in midair. “If he doesn’t make contact, he’ll fly home, rest and refuel, and come out later to try his luck again” (p. 25).

The odds are against any one drone: for every queen there may be a thousand drones fighting to inseminate her. They chase her in a “drone comet” that flies around the mating area, disintegrating and reforming in the drones’ collective quest. The queen will allow only a few of these males to mate with her.

And the rest?

Drones not only are short on chromosomes, they’re also short on life opportunities. When it’s time for honeybees to settle in for the winter, the worker females will kick out the drones to conserve food, but not before sometimes clipping their wings so they die faster.

But whether or not drones are executed by their sisters before being thrown out of the hive, they will never have learned to protect or feed themselves. So they had better do their best to mate while they can.

That begs the question, what if a drone does make contact?

Curse of the endophallus

Drones are not only short on life opportunities, but even the few they have seem to work against them.

The reproductive organ of a male bee is called an “endophallus,” which is a penis (-phallus) that resides inside (endo-) the body of the bee. When a drone catches up to a queen in flight, he grips her and prepares to mate, following this procedure.

The drone’s entire blood supply rushes to his endophallus, causing him to completely lose control over the rest of his body. The endophallus “everts,” which is a euphemism for “it explodes out of him and into the queen.”

The drone endophallus is optimized to disperse a large quantity of seminal fluid and spermatozoa with great speed and force…

The process of ejaculation is explosive—semen is blasted through the queen’s sting chamber and into the oviduct. The process is sometimes audible to the human ear, akin to a “popping” sound. The ejaculation is so powerful that it ruptures the endophallus, disconnecting the drone from the queen.

Wikipedia, Drone (bee)

The drone flips backward, paralyzed, endophallus ripped from his body. Seconds later, he is likely dead. The queen stores his sperm inside her, along with that of a few others. She flies home, ready to settle in and lay her eggs.

He falls out of the sky, life purpose fulfilled.

The planting of the hive

The above cases describe how bees make more bees. But how do hives make more hives? Does it make sense to think of a hive reproducing like a single organism?

It turns out that there are a surprising number of things that hives have in common with individual animals. As Seeley writes, they can

1. Regulate their own temperature, including a “fever response” to fungal infection.

   

2. Gather, consume, and digest food, distributing it throughout the “body.”

3. Circulate air effectively. This involves a cohort of bees whose jobs are just to ventilate the hive. In a body, you could think of this as a sort of ventilation organ.6

   

4. Make coordinated decisions that are implemented by the whole hive.

It’s this last point that allows the hive to reproduce as one.

Beehive as apple tree

Just like a flock of geese, beehives have their own reproductive season. At the start of summer, the hive begins to release hivelets7 into the air that can spawn a new hive in a different place. And once the hive is settled, its wax combs built, it can’t easily uproot itself and move somewhere else. Plus, hives are hermaphroditic entities—they have both female and male reproductive parts, in the form of queens and drones—and they have to breed with other hives to avoid inbreeding.

These features also happen to describe many plants, which is why Seeley likes to compare the reproductive cycle of beehives to that of the apple tree.

Tree → parent hive

When spring and its flowers arrive, the hive begins to raise extra bees and store extra food. Once big enough, swarms start to emit, similar to how a tree begins to bloom and release pollen.

Grains of pollen → drones

Pollen grains are released from apple blossoms each year, each one hoping to fertilize the egg cell of another tree.

But of the millions of pollen grains released by a tree, only a small number will fulfill their destinies. The rest are dispersed, they are lost, they are baked into bee bread to be eaten by the everyday worker bee.8 Pollen is therefore an apt metaphor for the honeybee drone.

Meanwhile, the queens continue to feast on royal jelly.

Egg cell/seed → queen

A queen who has mated is like a seed in an apple, with great potential to set up a new hive. Also like an apple seed, she must be planted in a good place and taken care of.

Apple → swarm

To take care of the departing queen, a swarm surrounds and protects her; they must also carry her because she can’t fly. Male propagules—the pollen grains or bee drones—are “discharged ‘naked’” (Seeley, p. 33) from the tree or hive, but the queen is “packaged inside a large and intricate dispersal vehicle that will give it protection and help it along.” This is much like how apple seeds are packaged inside a tasty and protective fruit.

For both hives and trees, the dispersal of the female is much more costly than the dispersal of males. So each hive (tree) only releases a small fraction of swarms (fruits) compared to the number of drones (pollen grains) released. But they also have a much higher success rate of starting a new hive (tree).

Where the apple tree metaphor breaks down

Unlike an apple, a swarm has to see what’s out there, choose a good nesting site, and then move itself to that spot. But a swarm is made up of thousands of individual bees with their own (occasionally violent) opinions. How do they consistently choose sites so well, as observed by centuries of bee enthusiasts?

Anyone who has the immense good fortune of watching a honeybee colony cast a swarm will be treated to many astonishing displays of animal behavior.

First there is the feverish rush of thousands of bees out of the hive and up into the sky.

Minutes later, the cloud of swirling, swarming bees mysteriously condenses into a tight crowd hanging from a tree branch, where for several hours or several days nearly all the bees sit quietly, almost motionless…

Next…comes the most wondrous sight of all: suddenly, in about 60 seconds, the entire swarm cluster disintegrates and takes flight, filling the air with the roar of thousands of airborne bees.

This flying mob immediately begins moving off in the direction of its chosen home and in another minute or two it will have vanished.

(Seeley pp. 146-147)

More on this mysterious behavior in the next post.9 But first—

Doesn’t something feel off to you?

Considering the viciousness of bee reproduction, isn’t it surprising that hive reproduction is… not? That is, apple trees are not known for their aggressive tendencies.

Does that mean a hive might not require total, placid cooperation from its members to act ultra-cooperative? Hives don’t have one queen-dictator suppress all dissenting opinions, as the sci-fi hive minds seem to tell us they must: the queen-fight savageries, enthusiastic participation of the worker bees, and annual drone executions are all quite democratically distributed, all things considered.10 So how is the hive set up to overcome this infighting?

Or, is it possible that the hive—so coordinated that it looks like it’s governed by one collective mind—needs intense conflict to work?

In the next post: how maybe infighting is an important part of the hive mind, other parts to making a hive run smoothly (including how bees choose their jobs), and what it looks like when a hive mind fails.

And, of course, how it all relates to the brain (it’s relevant, I’m pretty sure)

8

This is a technical term. Bee bread is the majority of what worker bees eat. It is “gathered rather than secreted (i.e. in contrast to royal jelly and beeswax),” made of pollen, nectar, and bee saliva and allowed to ferment before being packed into little pebbles.

To me bee bread looks quite nice, like little sugary crystal pebbles, which is maybe why it’s also known as ambrosia (Collins Dictionary). Unfortunately it can be contaminated by fungal toxins that make it unsafe to eat.

I did a live chat with Tommy Blanchard of Cognitive Wonderland last week—thanks very much to those who tuned in (including my mom) and hope it was fun!

Here’s the entire chat, also available as an audio recording. We talk about what we’re excited about in the field of neuroscience, why I made cyborg worms for my PhD, and how Tommy got mildly famous for wasting (?) his research funding.

Get more from Chenchen Li in the Substack app

Available for iOS and Android

Get the app

Introducing the Bee People

Here we find that some people really love some bugs.

1. Tom Seeley (ft. E. O. Wilson, the ant man)

Tom Seeley is a professor of biology at Cornell. His life’s work is on honeybees, which he studied as a graduate student at Harvard, with co-advisors Bert Hölldobler and the late great E. O. Wilson.1 As an aside (and unrelated to Tom Seeley), E. O. Wilson has a complicated legacy.2 Hopefully people learn from his problems while benefiting from his contributions, such as the writing he left behind about how much he liked to find and watch ants.

Wilson carried on the tradition of Victorian naturalists who made their contributions to ecology by watching things in uncomfortable places for very long times. The Myrmecological News Blog3 interviewed Wilson in 2018, in which Wilson tells interviewer Patrick Krapf about a field study he conducted in the 1950’s.

“There was a genus of ants, Aneuretus, which was extant in the late Mesozoic and early Cenozoic.” There was something special about this genus, probably, making it “important to find the last of the survivors.

“Two specimens had been found in a garden in Sri Lanka. So off I went to Sri Lanka,” he said, echoing the beginning of many a detective novel. He searched for ants day after day but couldn’t find anything, and once this went on for long enough, Wilson decided he had to instead just trust his deep understanding of ants.

“I knew if I went down to Ratnapura to the south,” he said, “I would be close to the rainforest, and I thought, maybe I could find them in the rainforest."

“I got out off the bus and walked into the tourist facility there, a little hotel. The hotel was surrounded by a bunch of trees. So I decided to walk out to those trees and see what there is. I walked up to a bush which had hollow twigs. And you of course know that hollow twigs are one of the best places to find ants.” Wilson is speaking to Krapf, another ant person; this is a thing that they know.

“I reached over to a branch and broke off a dead twig. And out poured Aneuretus! I had struck gold (for a myrmecologist). They were all over my hands and arms, biting me, spraying Aneuretus poison, whatever that was; I couldn’t care less, because I had found Aneuretus!”

From this we learn that being bitten and poisoned by an ancient ant species was one of the greatest moments of Wilson’s career. It is therefore of no surprise that his student Tom Seeley includes many photos like the following in his books (although his Hymenoptera of choice was the honeybee rather than the ant):

It is therefore also of no surprise that the first chapter of Seeley’s book Honeybee Democracy (2010) begins,

Go to the bee,

thou poet:

consider her ways

and be wise.

—George Bernard Shaw, from the play Man and Superman, 1903

This is more or less what we intend to do in this blog series.

2. Lars Chittka

Lars Chittka is a professor of ecology at the Queen Mary University of London. His life’s work is also on honeybees, and he published a book titled The Mind of a Bee in 2022.

After reading his book and watching Lars give a keynote lecture for the biggest conference in my field, I knew him as—as a friend put it—“a real rock star” in the world of biology and neuroscience. And the friend said this even before considering Lars Chittka’s actual rock band!

Lars is part of a band called Killer Bee Queens along with Katie Green and Rob Alexander. In 2019 they released a psychedelic rock album titled Strange Flowers, including songs such as I Stung Gwyneth Paltrow and The Beekeeper’s Dream. Here is the album, with its many lyrical verses about being a bee.

Aside from his professorship and rock album, Lars has also received the Royal Society Wolfson Research Merit Award, an Advanced Fellowship from the European Research Council (ERC), is an elected Fellow of the Linnean Society (FLS), the Royal Entomological Society (FRES), and the Royal Society of Biology (FRSB) (source).

What is it about bees?

People can get really into all sorts of things, especially scientists. But bees seem to capture people’s imagination in a special way, illustrated by three of my recent trips to a six-story bookshop in London.

• Occasion 1: I was looking for a Chinese cookbook and found less than a foot of shelf space dedicated to Chinese cookbooks. I was slightly disappointed as the overall cookbook section was substantial.

• Occasion 2: My partner was looking for a book on parallel processing with GPUs. There were only a handful of options, none of which were exactly the goal. (It had been an optimistic journey and we were not surprised.)

• Occasion 3: I went to find Lars’ book on bees. I also wanted to see if there were other interesting honeybee books around. To my astonishment, I found three full shelves, each four feet long, packed edge-to-edge with bee books. Lars Chittka’s The Mind of a Bee was strangely absent. I was informed at the front desk that it had sold out.

Why do people like bees so much? As Tom Seeley waxes in Honeybee Democracy, “Honeybees are sweetness and light,” and as Lars elaborates on his album page, “Bees have given us sweetness (honey), light (candles), and intoxication (mead is one of the most ancient alcoholic drinks known to man).” Also, honeybees are fuzzy and like flowers, and they live in cartoonish hexagonal grids.

But one of the most fascinating things about honeybees is, I think, the idea that they live in massive hives.4

Journey into the hive mind(s)

Hives can have tens of thousands of honeybees, almost all female. These bees are incredibly well-coordinated: workers gather food for themselves and their queen, form swarms that make group decisions, and they seem to know just what the hive needs at all times.

Bee-watchers have reported on these behaviors for millenia, creating a narrative that hives behave as a single cognitive entity and act like a single organism. From Ender’s Game5 to the Dreamworks masterpiece How to Train Your Dragon,6 to Rick and Morty7 to Apple TV’s recent show Pluribus,8 we all seem to have a thing about collective mind-melds, and in each case, the “true intelligence” of a hive mind comes from one of two things: a disembodied dictator or a ruling queen.

But which one is real? That is, what does it feel like to be a bee inside a hive?

Hive mind v.1: The Disembodied Dictator

In season 2, episode 3 of the sci-fi cartoon Rick and Morty (“The Auto Erotic Assimilation”), the characters Rick, Morty, and Summer come across a spaceship that has put out a distress signal. They board the ship and find aliens who have escaped disaster. The aliens tell them:

Can you help us? Our planet was taken over by some kind of… entity. It absorbed the minds of our people. […]

We didn’t notice until it was too late. The people it takes over—they—they look like your friends, your family, your leaders, but they’re not themselves anymore. They’re part of… it.

It turns out they did not in fact escape the hive mind, however, whose name is Unity and who used to date Rick. By invitation, Rick, Morty, and Summer then land on a planet full of aliens that have already been assimilated.

Rick and Unity catch up while Unity uses other absorbed aliens to keep Morty and Summer entertained. But Summer is outraged; she isn’t happy that Unity has wiped out the personalities of everyone on the planet, and tries to get Morty to care while Unity has her mindless drones carve sculptures of Summer’s face into a mountain.

Morty: Man, look at ‘em go, so coordinated. What’s your problem? They’re making you into a Mount Rushmore.

Summer: Morty, open your eyes! There is no they!… They’re a planet of puppets!

Morty: Well, it seems like everybody here’s cool with it.

Indeed, Unity has achieved world peace by making everybody part of herself. She is one mind with many mouthpieces, like a body cut up into many small and independently moving parts.

A common thread in the “disembodied dictator” version of a hive mind is that it is peaceful but soulless. As far as I can tell from the trailer to Pluribus, the concept is the same as in Rick and Morty—similarly, the main conflict is that characters want to maintain individuality in a world of happy, blank, and assimilated people. But I haven’t seen the show yet, so no spoilers.

Hive mind v.2: The Ruler Queen

In other versions of fictional hive minds, there might be a single individual that puppets the members of its hive. A good example of this is in the Dreamworks classic, How to Train Your Dragon.

In the movie, Astrid (a person) and Hiccup (a person) ride Toothless (a dragon) to sneak into a giant dragon’s nest. They find out that hundreds of dragons have been delivering their prey to this nest, prey that has likely just been stolen from Astrid and Hiccup’s village, pushing the Vikings toward starvation. A scene from the movie, slightly cut:

Hiccup: What my dad wouldn’t give to find this… Ah, it’s satisfying to know that all of our food has been dumped down a hole.

Astrid: They’re not eating any of it.

[They continue to watch dragons drop food into an abyss. Suddenly, a giant dragon leaps out of the hole, swallowing a smaller dragon that didn’t bring enough food.

Astrid: What is that?

[The giant dragon seems to notice them watching from the side. Hiccup tells Toothless to get them out, and they narrowly escape the nest on Toothless’s back. Dragons swarm out with them in dramatic fashion.]

[Later, after they’ve landed in a safe forest:]

Astrid: No, no, it totally makes sense. It’s like a giant beehive. They’re the workers, and that’s their queen. It controls them.

Then in How to Train Your Dragon 2 (a heart-wrenching piece of cinema), we see the psychic mind control powers of an alpha dragon, which it uses to turn Toothless against Hiccup and his dad. Overall, the “ruler queen” hive mind is very similar to the “disembodied dictator,” only the ruler has a body and looks like a big version of yourself.

Hive mind v.3: The Bees

In most media about hive minds, there’s a being with psychic power over its underlings, which are taken over and have their personalities erased. They do the hive mind’s bidding and lack their own drives, motivations, and purpose.

But is sci-fi right? Is this how it works?

No!

In fact, the bee people are adamant that you do not believe that bees fall into the classic trope of the personality-erasing hive mind. Both Lars and Tom Seeley have at least a few paragraphs in their books dedicated to correcting this misconception.

From Seeley’s Honeybee Democracy:

There is one common misunderstanding about the inner operations of a honeybee colony that I must dispel at the outset, namely that a colony is governed by a benevolent dictator, Her Majesty the Queen. The belief that a colony’s coherence derives from an omniscient queen (or king) telling the workers what to do is centuries old, tracing back to Aristotle and persisting until modern times.

Likewise, as Lars states in The Mind of a Bee:

The behavior of a bee swarm… can be so tightly integrated, so well-coordinated, that at times it resembles a single being… In this view, the individual scouts of the swarm have functional similarities to the sensors and nerve cells that collect and process information about the world in an animal’s brain…

Based on such observations, a popular view holds that in social insects, the individual is a mindless machine, and intelligent behavior only emerges by self-organization as a function of the group.

This has led to the notion, sometimes expressed in whispers, that swarm intelligence is somehow generated by a qualitatively different form of mind than the individual, a “collective mind.”

“But it is false,” says Seeley, who goes on:

…a colony’s queen is not the Royal Decider. Rather, she is the Royal Ovipositer. […]

Indeed, there is no all-knowing central planner supervising the thousands and thousands of worker bees in a colony.

So hive minds, as they exist in most sci-fi, do not apply to real bees.

Up next: a look at a real hive

That’s not to say the hive mind trope isn’t fun, from a storytelling perspective. It’s a variant of zombies but instead of trying to bite you, a previously beloved individual is lost to an aggressive inner peace. In pop culture, being a hive mind is opposed to being an individual; if you have your own thoughts and wants, then you’re not part of a collective intelligence. And if you are in a hive mind, your personality is erased or suppressed—you act out orders from someone on high.

But what do Lars and Tom Seeley mean when they say this isn’t how bees work? How might bees have their own thoughts and feelings, yet look like they’re part of a bigger whole? How do they know what to do without a queen telling them? Do they have free will?

In this blog series I will answer at least one of these questions. I will begin by discussing reproduction in the next post—that is, how do bees make more bees, and how do hives make more hives? By comparing these cases, we can get a sense of how bees relate to the hive and how individuals relate to a collective.

And to prime you for what’s to come, as Tom Seeley might say, “bee warned”9—the reproductive life of a bee is fraught with violence, in a highly unnecessary way.

A few months ago I wrote a blog series on something I’ve wondered for a long time: why can’t we simulate the behavior of a living animal using local neuron-like interactions? What are we missing—data, theory, or both? What are people trying? What are some other leads?

I meant to continue to flesh out my thoughts on this blog, but, oops. The thing is, I’m currently a postdoc, and I think that got me stuck on a couple fronts:

1. My ideas aren’t that developed yet

Postdocs are the temporary stage between being a grad student and a faculty member (which is why my parents keep telling me I should get a real job). Once one does reach the stage of faculty and becomes a professor, one is supposed to have one’s own research direction and a pretty developed sense of what the field is like, whatever that might mean.

Whatever it does mean, I don’t feel that I’m there yet—but I also don’t know if this feeling will ever change; maybe it’s like never feeling like you’ve grown up into a real adult (which may exacerbated by staying a student until your late 20’s). Right now, I have a set of mental models for how the brain works, and when I come across new information, I try to see where it might fit or clash with these personal models. There is a lot of data I don’t know how to sort into neat frameworks, and because of this, my models are constantly evolving or getting tossed out. The “getting tossed out” part did happen recently; I needed to change my mental model of something a lot given what I considered convincing new data. Stuff like this makes my ideas feel rather turbulent, and after my last few posts, I had hoped this turbulence would settle into some semblance of coherence before I continued writing.

I think I should accept that this won’t happen anytime soon. I will probably never stop feeling confused. But I’d like to keep writing about other fun things while I wait for my thoughts to sort themselves out, and in the meantime, I do think it’s useful for me to put the uncertainty down on paper.

2. I got distracted

Here are some things I’ve done lately.

CRSy 2025

One of my favorite parts of being a scientist is meeting people, traveling, and seeing friends at conferences. I especially liked the Champalimaud Research Symposium last October in Lisbon, which was on Cybernetics.2 The topic changes every year, but I’d highly recommend going if you can.

My favorite lecture was from Auke Ijspeert, who presented some incredible research on animal locomotion and robotics. Some teaser questions for his life’s work could be, as he asks in the video below, how are animals so fast with neurons that are so slow? What can we learn from animals to help us build everyday useful robots?3 Animals are good at movement in many ways we can’t replicate yet, and Auke Ijspeert has spent his career trying to bridge this gap.

Here’s the recording from the event. Another quote from Auke, which I hope will convince you to watch the lecture:

Plus, if you attend CRSy, you get to be in Lisbon in October.

Fellowship applications

I applied to, interviewed for, and got (!) a fellowship that will fund my current job in London for another two years. The process was painful because I don’t have much experience writing grants, although asking many people for their past successful applications helped a lot. Applying for this fellowship made me heavily question the existing models of research funding, mostly in the form of complaining to everybody around me all the time.

Visiting China for the first time in over a decade

I’d heard from my family that China had changed a lot since 2012, which was the last time I visited, but I’d never seen it firsthand until this winter. It was beautiful, modernized in startling ways, and the trains were speedy and reliable.

I saw extended family, struggled to speak to them in Chinese, and ate seafood I had and might continue to have literal dreams about. One family dinner included four generations at the same table—some of my cousins have small children now and my grandma hangs out with them sometimes. That was cool.

Duolingo

Last August, I discovered Duolingo; or rather, Duolingo discovered me.

At the time I wanted to learn Chinese for the aforementioned trip to China, but also, I’d just watched K-pop Demon Hunters and was rapidly falling into a K-pop rabbit hole. So I started Chinese and Korean lessons guided by an angry owl.

The day I downloaded the app, I spent over four hours on it. I’ve since learned to restrain myself, but Duolingo still dominates my screen time statistics. I already knew I tend to fixate easily but I was still surprised at how skillfully this app manipulated my behavior.

Some Duolingo tactics that have worked especially well on me:

1. Streaks and “friend streaks,” which are the number of days you and your friends have both done at least one lesson. “Streak freeze” mechanics let you keep a streak going once or twice after you’ve missed a day. You can pretend you still have a real streak going, and to be clear, I’ve only used this mechanic one time. Streak freezes mean that missing a single day doesn’t completely demotivate you, which is nice.

2. Regular “tournaments” where you get matched with random people and compete to see who can get the most XP. If you’re at the bottom, you get demoted to a lower bracket; if you’re near the top, you get promoted. Eventually you reach the “Diamond Tournament” where you compete with very studious people for some unknown prize; as in, I’m not sure whether there exists a prize at all. Why does it still work on me then?

3. Being able to switch between courses. I suspect that learning Chinese and Korean simultaneously has helped me stick better to both. If I’m tired of drilling Chinese characters one day, I can switch to Korean. I still end up on the app, and then I’m more likely to go back to Chinese.4

Anyway, Duolingo is like a major part of my life now. I know these tactics work on other people, too, because at 11:45 PM on New Year’s Eve, someone at the New Year’s party I was at went, “oh no, I haven’t done my Duolingo yet,” which made other people in the circle gasp, take out their phones, and then we all sat together whispering languages at cartoons for fifteen minutes.

Did we miss raising glasses of champagne at the stroke of midnight because we wanted to maintain our Duolingo streaks? I don’t remember, but I think, probably, yes.

Now that I have presented my excuses, here is the plan

I’m not too happy with part 3 of the worm simulation series, so I’m working on revisions/expansions. I want much more specificity, structure, and rigor in argument, and I’d like to fix that while maintaining readability (and entertainment value?). I have been intermittently working on these revisions for a while, but they’ve taken more thinking than I anticipated. They will be out eventually (and in my favor, there is the initial momentum of a public New Year’s resolution to keep me going5).

Aside from neuroscience, I also want to write up primers on fun math, physics, psychology, and computation topics. As per the “Feynman technique,” I’ve found that teaching lectures tends to make me learn material much better than rereading textbooks does, and a blog seems like a nice way to impose the Feynman technique on myself. I also want to write fun things related to my current interests, which include honeybees, K-pop, and Auke Ijspeert.

I am not sure how to go about all these writing goals in a realistic way yet. For now I’ll start with a daily writing habit, motivated by the masterful manipulation of myself that I’ve witnessed by Duolingo, plus this 70-day shuffle challenge that some people I know are doing on Instagram. Altogether, I’ll get back to posting once every other week on Substack. Maybe I’ll work on a longer piece and post stuff in between. But the priority will be to spit out something every two weeks.

I still feel strongly that making the day-to-day of scientific research more open and accessible is important. It’s not just so people who aren’t scientists can see under-the-hood how research is done; it’s also because, while doing science is fun and (sometimes) meaningful, and you get to meet people from around the world to discuss cool things, in the end you usually are sitting at a desk by yourself trying to make a dumb thing work. It can be an isolating job. It helps to talk about it. So thank you for reading! You will being hearing more from me in the future, if you want.

If you’ve been following these, you might want to skim the updated version of Part 2 for context.

This post is about why I think circuits are the wrong metaphor for the brain, and why I think more data is not the answer to understanding nervous systems.

(Not that I think neuroscientists should stop collecting data. We just shouldn’t expect data alone to answer the big questions.)

Example 1. My friend’s animal brain collection as it relates to a small set of left-handed women

My childhood friend Sophie is a veterinarian and collects animal parts. One day I was over at her place admiring the collection, and I picked up this vulture brain.

I tried to identify its neuroanatomy but quickly got confused because I saw two brainstems. Whoa, I thought. Maybe she found a vulture with brain cancer. Maybe one of those brainstems is a giant tumor.

I asked Sophie about it. She said it wasn’t two brainstems or a cancer-ridden vulture—vultures just have abnormally large olfactory bulbs. What I’d thought was a tumor was really the basis for an overdeveloped sense of smell.1

Which aligns with the circuit perspective of the brain, right? As in, vultures need to smell very well, so evolution gives them bigger smell modules. And this should go both ways: if bigger olfactory bulbs help you smell better, then smaller ones should make you worse.

PSYCH

In 2020, Weiss et al. published a paper titled, “Human Olfaction without Apparent Olfactory Bulbs.” They’d randomly found two women who didn’t have olfactory bulbs in their MRIs. Then they ran tests and concluded that these women could smell the same or better than everybody else.

Even more confusingly, the women had identical twins who did have olfactory bulbs. The ones without the bulbs scored slightly higher on olfaction tests. And it seemed like this phenomenon might be more common in left-handed women, specifically.2

Hearing this, you might argue that human brains are very complex. That of course they would have backups, that they must have redundancies for every computation under the sun.

To that I counter with the Hydra.

Example 2. An animal that can dissolve and come back together again

Hydra vulgaris is a freshwater polyp. These animals have 4-12 tentacles and are 1-3 cm long.

Instead of centralized brains, they have nets of neurons, and they use them to swim, somersault, and eat. The authors in this paper describe their eating behavior as “three different stages: tentacle writhing, tentacle ball formation, and mouth opening,” which would sound like a horror movie sequence if it wasn’t about such nice little jellyfish creatures.

Also, they reproduce through asexual budding:

A long-known result from Hydra work is “dissociation and reaggregation.” That is, you can dissolve Hydra into individual cells, and if you leave them alone, they’ll reassemble. New Hydra can be different from the old Hydra; what used to be tentacle cells might end up in stalks. They’re not too picky about where their pieces go.

Importantly for our purposes, new nervous systems in the reaggregated Hydra will spontaneously start working—but not right away. Lovas and Yuste (2021) showed that they have to go through an “ensemble synchronization phase” first.

This doesn’t make sense if nervous systems are inherently carefully wired circuits:3 the Hydra neurons didn’t slot into position and start tentacle-balling immediately. Instead they bumbled around for a while, tried some firing patterns out, and only after some time did they discover their destinies as tentacle-writhing neurons or somersaulting neurons.

To be clear, there wasn’t any backup circuitry here. The Hydra had been completely dissolved. The neurons relearned to be a functional nervous system, and if this doesn’t make you question the circuit model, then let’s go back across the evolutionary tree, all the way to mammals.

Example 3. Things that happen when you mix up a ferret’s brain

In a study by Roe et al. in 1992, researchers did a surgery on ferrets that sent signals from their retinas to what would usually be auditory cortex. That is, vision signals were sent to the sound processing part of their brains.

They found that the modified sound neurons not only reacted to light signals, but started to compute information like vision neurons did. They became sensitive to orientation, velocity, and other hallmark characteristics of visual processing; the neurons also organized themselves in ways similar to normal visual cortex.

There were differences, certainly. The auditory neurons weren’t as responsive as visual neurons would be, and their receptive fields were too large. But there was a surprising amount of structure that just showed up.

Hearing these three examples, you might rightfully say that big rewiring doesn’t happen all the time—plus, every example so far has involved neurodevelopment. What if neurodevelopment plays the heavy lifting and most of the time the brain still does work like a circuit?

Because neurodevelopment isn’t the only time our neurons are rewiring themselves. Our brains change whenever we use them.

Examples 4+ (aka The Big Hits)

Here’s how you can think of the dramatic plasticity in nervous systems:

1. Nervous systems are mostly like circuits, except when something big happens and they trigger some special backup plasticity. OR

2. The extreme plasticity isn’t an exception to the rule; it is the rule.

As you might guess, I think #2 is closer to the truth. To further support my claim, here are a few classic cases from neuroscience literature:

• Taxicab drivers in London develop enlarged hippocampi, the region of the brain responsible for navigation (Maguire et al. 2000).

• When violinists had their brains scanned while pretending to play a piece, activations were really different between professionals and amateurs. This was true even when they were tapping their fingers to nothing (Watson 2006).

• Braille readers have larger regions of their brains responsible for feeling in their fingers. Unlike the violinists, though, if you ask them which finger they’re using to read in a given moment, they sometimes get it wrong. That is, the regions for sensation on each finger have started to merge and absorb each other. From Sterr et al. 1998:

Each of the three-finger readers had a strong tendency to misidentify which finger was being touched during tactile sensory threshold determination, although there was no difficulty in determining that one of the fingers had been touched….

Sensory threshold testing was also conducted in five other sighted individuals, none of whom had difficulty in localizing tactile stimuli.

This paints quite a different picture of the brain. Rather than molecularly precise circuitry, it’s as though brains are oceans where all possible neural computations are like waves, rising and falling, gaining momentum or losing it, crashing into each other or splitting apart.

And if you want to faithfully define a functional module of the brain, you have to include these ongoing changes.

What if circuits aren’t a perfect model, but they’re close enough?

Here you might argue that connectivities could change slowly enough that we can approximate the system as a fixed circuit. As in, maybe we can figure out how a violinist’s brain works today before considering it ten years from now.

Rent (don’t buy) your textbooks in a fast-changing field

In 2014 I was taught that spike timing-dependent plasticity (STDP) governed synaptic plasticity in the brain. Since then the picture has changed: two years ago, a professor from MIT got up on a lectern at a conference I was at and asked a room full of people, “does anyone still believe in STDP?”

His point was that the focus on STDP’s role in computation has been mostly supplanted by BTSP, which stands for behavioral timescale synaptic plasticity. In BTSP, connections change on the same timescale as the behavior, not rapid-fire spikes. An important possibility is that these changes may be part of how brains carry out tasks—not just how they learn from them.

For instance, even for things you’d think would benefit most from being stable, like memories, your synapses change to retrieve them. This makes your memories unstable in a way that computers aren’t. In fact, a theoretical paper by Clark and Abbott (2024) shows that changing weight dynamics are not only compatible with working memory function, but could be responsible for it.

So I don’t think it’s enough to say that nervous systems are circuits that change. The change seems to be so essential that clinging to a modified circuit metaphor is probably less useful than finding a different picture entirely.4

Concluding thoughts on the worm problem

[Or: The Worm Hole?]

I didn’t have to look hard to find the examples in this post. As the saying goes, you can’t swing a rat in neuroscience without hitting a case of extreme neuroplasticity.

How metaphors and models shape science

I suspect that circuits are too alluring a metaphor. Modern computers are great, so why wouldn’t we use them to think about other complicated things?

But this is a historical pattern that isn’t always helpful. In Hippocrates’ time, around 400 BCE, bodily functions were modeled by fluids. In the 1600s, the body was a mechanical machine.

And now we’ve found ourselves in the digital age, where the most impressive inventions are things like circuits (and I am including large language models in this category, because while they learn during training protocols, they’re designed to be frozen at runtime).

When we have the right underlying model, more data can help tune it. But when we don’t have the right model, we tend to make epicycles instead.

Are we trying to build modern epicycles?

Epicycles were how Ptolemy around 100 CE explained the motion of the planets. His model was centered around Earth and looked something like this.

The problem with epicycles was that they were super accurate. They were great at making predictions because they were heavily fit to data—Ptolemy did this by adding slight asymmetries, kind of arbitrarily.

But epicycles didn’t represent reality. They wouldn’t have generalized to other solar systems, and if we had stuck with them, we couldn’t have reconciled gravity with other forces.

We needed Copernicus’ heliocentric model for that, and even the Copernican model needed arbitrary corrections: the imaginary planet Vulcan, for instance, was made up to explain Mercury’s deviation from its expected orbit; it took until Einstein reconceptualized reality itself to finally figure Mercury out.5

So there can be huge future costs to achieving perfect simulation without the right theory. We can think we understand a system when we really don’t.

(And that’s if we can figure out perfect simulations. In neuroscience, we don’t even have epicycles yet.)

If I haven’t convinced you yet that more detail is not the answer, let me try one last time

Let’s say we stick with the circuit model. Let’s say we’re forging ahead, collecting more data on the worm until we solve it. What might we expect?

Well, as we’ve discovered, worm neurons are complicated and really diverse. We already have the connectome, but we still have to pin down exact models of each individual neuron.

And if we want a real worm, we can’t describe each neuron’s solo dynamics. To get the interesting stuff out, like what I mentioned in Part One of this series, we have to allow our circuit to change.

Here’s where things get messy.

Each neuron will change in different ways in response to different neural signals. If we’re really going for precise, then we have to figure out how the circuit will change in response to every pattern of neural activation. Every neuron is unique, after all, and the next natural question is, how many possible neural activations are there?

To start, let’s say all C. elegans neurons have binary on/off states (a massive oversimplification). If we do that, we can calculate the number of ways a 302-neuron nervous system can be activated,6 which is (2^302) - 1. Spelled out, that’s

8148143905337944345073782753637512644205873574663745002544561797417525199053346824733589503 ~= 8.15 x 10^90.

This is a big number. Compare it to the number of estimated atoms in the observable universe, which is 10^80. There are over 10 billion times more ways to activate neurons in a worm than there are atoms in the observable universe—if we pretend worm neurons are binary.

Statistical mechanics is full of absurd results like this. Combinations and permutations are unreasonably explosive, and that’s even before we’ve looked at how each pattern affects each neuron. The bleak implication is that if nervous systems are like circuits, we were always going to have to make gross approximations. We were never going to be able to account for every possible case.

But—approximations are okay if we’re comfortable with what we lose by making them.

One approximation could be to say that only some firing patterns are biologically relevant. First—I’m not sure that’s true; the world is also quite combinatorial. And second, if you place artificial restrictions on your model based on the behaviors you’ve seen, then you forego the possibility of seeing new behaviors emerge.

“Life finds a way” (Jeff Goldblum, Jurassic Park)

If I put a real worm in space with air and food in zero-gravity, it’ll probably do something interesting. Worms use electric fields, gravity, and possibly magnetic fields to navigate,7 and at least some of these are messed up in space, so they’ll have to adapt. Maybe they’ll invent behaviors to compensate, especially if their space-habitat is complex.

On the other hand, if I put a feasible fixed-circuit model of the worm in space, it would only be able to do what it was programmed to do. This is a brutal tradeoff for the circuit model: either you have something that’s realistic to build, or you keep the flexibility and creativity that makes animals interesting.

But given everything I’ve presented, especially the parts about tiny nervous systems, the tradeoff just doesn’t make sense to me. And that finally brings me to my personal conclusion to the question of why I think we still can’t simulate a worm.

It’s because we need a different model.

In future posts, Emergent Systems and Where to Find Them: an alternative model to the brain, and examples of them in biology and physics.

Parts 2 and 3 are a revision of an earlier post. This time I try to make my views clearer:

(1) Brain simulation projects have treated brains like circuits or computers, and if that perspective is correct, then getting more data and more precision might help.

But (2) brains are not like circuits, and (3) that’s a good thing, because if they are, we’re very fucked

A few years ago, I went to the Janelia Research Campus for a nifty week-long theoretical neuroscience workshop.

The night of arrival, the organizers hosted a “speed dating” event for us scientists to get to know each other. We hopped around miniature tables being very excellent at having regular conversations with other regular people.

Drinks were flowing; sleep deprivation was kicking in. I sat down at my next speed date, a postdoc from a big machine learning research center.

“Isn’t it weird that we haven’t been able to simulate the worm yet?” I asked him.

“Yeah, it is weird,” he said. “That’s why our field is a laughingstock, right?” Then he grinned, because it was a joke.

?

Over the last four years, I’ve managed to worm this question into a lot of conversations.1 Here are the main answers I’ve gotten from neuroscientists around the world.

1. “I don’t know”

As in, what my friend from Janelia said above. Or from another colleague: “it’s humbling, isn’t it?”

2. “That’s why I’m an AI researcher now”

Some people have gotten disillusioned with theoretical and computational neuroscience. They’ve been seduced by progress, money, and the fact that things work over on the AI side.

I’ll be honest, those things do sound nice, so I can’t blame a suspicious number of my friends for moving to greener pastures. We agree that AI is not really like the brain, but who cares when even conversations on the street are about the newest model of ChatGPT?

3. “We need more data”

The ones who do answer the question of why we can’t simulate a worm say that we got unlucky for choosing C. elegans in the first place. Because while C. elegans have helped us a lot,2 we’ve discovered that their neurons are not quite what we’re used to.

Neurons talk with action potentials, except when they don’t

The story we tell in Intro to Neuroscience is that neurons communicate with action potentials. These massive electrical pulses trigger flurries of chemicals (neurotransmitters) at synapses, which are the spaces between neurons.

These neurotransmitters can cause voltage jumps in their neuronal targets, and if enough add up as they travel down a cell, the main body of the neuron is tipped over a threshold and a new action potential is set off. So if you stick an electrode in a neuron, you might record something like this:

Most neurons we’re familiar with fire action potentials, including many in vertebrates, insects, and lots of other invertebrates.

But action potentials aren’t the only way neurons talk to each other. For instance, a C. elegans neuron called RMD generates plateaus instead of action potentials; it sticks to a voltage instead of firing spikes. Other neurons like AVA use graded potentials, which are smooth, continuous signals. And long after we accepted that worm neurons don’t talk with spikes, we found out that at least one does (see Liu et al. 2018).

Also, a relatively high proportion of C. elegans neurons don’t even communicate through neurotransmitters—they use gap junctions instead, which true to their name, are gaps between neurons where electrical signals can propagate directly between cells.

Overall, the diversity of C. elegans neurons make their nervous systems quite complex. And what’s one response to complexity?

More data will solve everything

In 2023, academics led by Gal Haspel at the New Jersey Institute of Technology released a document they called The Worminator. In it, Haspel et al. argued that “the time is ripe to reverse engineer an entire nervous system.” According to them, this nervous system should be that of C. elegans, and we just need more data to simulate it.

It’s hard to argue against the idea that if we figure out every single complication in every single neuron in every situation it could ever encounter, then maybe we could build a worm. But let me reiterate that worms have 302 neurons, and we’ve been studying them for sixty years.

Remember Yuri Lazebnik, who asked whether a biologist could fix a radio? He used the example of p53, a tumor suppressor protein whose “mystery… seems only to deepen as the number of publications about this protein rises above 23,000.”

There haven’t been quite so many publications on worm neurons. But there are still 2310 hits when I look up “AIY Caenorhabditis elegans” on Google Scholar, and 100 hits for the far less popular neuron VD7. Publication number is a loose proxy for information, but will the 2311th paper on AIY really solve things? Or even the 101st paper on VD7?

Maybe we should figure out if we’re in another p53 situation before doing anything drastic.

4. “Let’s look at the fly instead”

Many have wanted a fresh start to the problem of small nervous systems. These people have found homes in the fruit fly, Drosophila melanogaster, whose connectome came out just last year.

This makes me think, yes, this time we will get it. Something about the extra 130,700 neurons will really work to our advantage.

I don’t mean to be so sarcastic; a lot of fruit fly work is cool. But I secretly think it’s a letdown that the biggest success in fruit fly theoretical neuroscience is the discovery of the head direction ring attractor network—in other words, we’ve found out how flies track where their heads are pointing.3

There is way more to the fly! They do a lot more than turn their heads; I’ve seen them do it. I am not exaggerating when I say that some consider this to be the single biggest success in theoretical neuroscience, not just in the fly, but in the entire field.

I have qualms with these answers and here they are

Well, I have no qualms with answers #1 and #2. I think those are valid. As for #3 and #4:

→Worms are complicated (but so is everything else)

It’s a fact of life that the better I get to know my human friends, the weirder I find out they are.

Likewise, I think the wrong lesson to learn from worm research is that worms are uniquely strange. They’re just the animals that we know best.

In fact I’d argue that worms necessarily have one of the least surprising nervous systems, because in any bigger animal, every brain is so different that it doesn’t make sense to build a neuron-level map for a species—you have to build the map of a specific individual. This is true even for fruit flies, which are still quite small, yet also have graded potentials, plateaus, and bursts in their neurons.

Those in glass houses shouldn’t throw stones

Where the stone-throwing is us calling other animals’ brains weird

Take the human brain: not only do our neurons have a lot of different activity patterns, but these activity patterns (and even different learning rules) can coexist in the same neuron.

And published just last year, one of the more mysterious findings by Jeff Lichtman’s lab at Harvard in collaboration with Google was the discovery of never-before-seen structures called axon whorls (Shapson-Coe et al., 2024).

In some human neurons, it seems that axons (the part that delivers signals to other neurons) can loop in on themselves dozens of times. They form tight coils that sometimes rest on other neurons and sometimes are in random places.

What do they do? Who knows. How do they form? Unclear. Do they affect computation in the brain in an important way? No idea.

There are bizarre possibilities for what these axon whorls are doing, considering that coils of electric current form magnetic fields that can act on other currents without contact. We don’t know if axon whorls exist in all primates, all humans, or just this one human brain they decided to slice.

Also, this came from a mapping of a cubic millimeter of one person’s brain. We have no idea what else is out there. While a cubic millimeter is rich enough to have produced 1.4 petabytes of data, there’s a lot more to explore.

If you still disagree with me and think that bigger brains mean the possibility for more simplicity—then why am I, a human, writing this blog post, and not a worm?

QED.

→ Why do we REALLY think more data/compute will help?

I think that the idea that more data will allow us to simulate a nervous system is based on the assumption that nervous systems work like circuits.

Here’s what I mean: if brains are like circuits, then each piece has a role. These pieces can be at the level of neurons, or modules, or whatever. Each component does one or a few things, and they have to be connected properly to work.

To solve a circuit-like system, this is the formula:

1. You figure out the sensory part.

2. You figure out the movement part.

3. You figure out how to connect them.

4. Repeat for every other “functional part” and every other interaction.

The idea is that if you do this in enough detail, with enough thoroughness, then you’ll win the brain simulation game.

Okay, let’s go with that for a minute.

Then how do you explain the cases in Part 3?

In the final part to this series: examples from neuroscience that, in my opinion, don’t really mesh with brains being like circuits.

During my PhD at Harvard and my postdoc at UCL, I went through formal peer review with some of the highest impact factor-publications out there. I’ve acted on both ends of the process, either as reviewer or reviewee. And my conclusion is that peer review doesn’t work.

How come?

Reviews are maddeningly inconsistent

A student in a lab I was in—let’s call them R—announced one day that “reviews are completely stochastic.”

R had submitted a paper to a very competitive venue and gotten rejected. But R disagreed with the reviews, and simply switched the order of two sentences before resubmitting to a different, equally prestigious venue. Where it was accepted.

In a separate incident, a different student sent a manuscript to a high-impact place and got back two reviews. One of them said something like,

Hot damn, this will be a foundational paper in the field.

The other said,

This is dumb and nothing new.

(I paraphrased.)

What then?

And it’s not just opinions that are inconsistent—you also have no clue who reviewers are. Reviews are blind, which is good for preventing bias, but it also means that sometimes you’re being reviewed by the singular expert in the field, and other times your career hopes rest on an undergrad who got your paper pushed onto them by a postdoc who suddenly got too busy to do it themselves.

Usually the second one.

Reviews can take so long that you’ve forgotten about the project, or so long that somebody dies.

(The last thing really happened to someone I know.)

In my experience, it takes a minimum of three months for first reviews to get back to you. Later rounds can add up to well over a year. And that’s only if you get accepted—otherwise, you resubmit and start over. So for multiple rounds of submission, review, and rejection, it can be several years before publication.

A lot can happen in that time. You can graduate from college; you can graduate from your PhD. You can decide to change research directions altogether.

All of these applied to me at some point—by the time reviews came back, I didn’t remember the details of my project anymore, which were, as it happens, exactly what I needed to address reviews.

Also, I was normally less invested by then. Instead of trying to improve the paper, I’d want to get responses done as fast as possible so I could work on a new thing.

Overall, the peer review is not a good system for good science. For deeper discussions on the messed-up reality of peer review, see:

• The present and future of peer review: Ideas, interventions, and evidence (Aczel et al. 2025, PNAS)

• The misalignment of incentives in academic publishing and implications for journal reform (Trueblood et al. 2025, PNAS)

• Adam Mastroianni’s post, linked below.

Experimental History

The rise and fall of peer review

For the last 60 years or so, science has been running an experiment on itself. The experimental design wasn’t great; there was no randomization and no control group. Nobody was in charge, exactly, and nobody was really taking consistent measurements. And yet it was the most massive experiment ever run, and it included every scientist o…

Read more

3 years ago · 1198 likes · 372 comments · Adam Mastroianni

More or less, the consensus is that peer review (1) controls how science is done today and (2) is kind of hot garbage. Formal peer review lets a handful of anonymous reviews of highly variable quality decide where your career goes next, and

that’s

just

how

it

is.

But Substack comments have been cool so far

I’ve been surprised by the amount of engagement on my posts.

I had probably less than twelve readers for my undergraduate, master’s, and PhD theses combined, so that’s where my bar is. My readership wasn’t a reflection of the quality of my work, either (is what I’m telling myself); most scientists get used to this lack of attention.

But on my last post, readers did pay attention. And they gave me a lot to think about. Specifically, they made me think I could have written the essay better, and for the last week I’ve been asking myself,

1. How could I have sharpened my arguments?

2. Are they right—were my analogies not good analogies?

And to my embarrassment, I wrote this nonsense to someone in a comment thread:

My plan for future posts is to lay out the stuff I've been working on […] (so maybe I'll address your questions eventually!). If you're curious now, though, unfortunately all I can do is refer you to my PhD thesis.

This is a two-for-one copout (although the person I wrote this to was quite gracious about it).

I effectively said, I’ll get to it later, and in the meantime, here’s 200 pages of reading for you to do. It’s like writing that something is “out of the scope of this paper”—it could be true, but it could also just be lazy.

Which means I’m going to try something new: I’m going to use Substack as an alternative to formal peer review.1

Here’s how I’ll do it

(but feel free to give me suggestions)

Every once in a while, if

1. I feel a post isn’t good enough based on feedback, and

2. I really care about making those ideas clear,

then I’ll go back and revise. I’ll try to improve my arguments, or change them if you’ve changed my mind.

In the last few posts I’ve been setting context for my current work, which is on understanding and building small natural intelligences. But soon I’ll be writing up actual data, with code, which means that future revisions won’t just be to presentation but to the work itself.

For a given post there might be no revisions, or there might be several. I’ll announce them with a Note (Substack’s version of tweets, I guess) instead of spamming people’s inboxes; I’ll also explain what prompted the changes.

And I won’t remove the drafts. I want to keep old discussion threads and the stuff being discussed. But I will put junked drafts behind paywalls to separate them from edited versions. So if you’re a paid subscriber, thank you so much for your support! You’ll get (/win) permanent access to my Substack landfill—vague points I failed to make, uncomfortable jokes that didn’t land, and detailed comment sections pointing out my flaws.

Here’s why I think it can work:

1. Feedback is fast

Within days, I can have a whole back-and-forth with readers about what I’ve written. I can have DM exchanges to find out why people are interested and what they want to hear more about, which might surprise me.

Best of all, these exchanges happen while I still remember what I meant by a particular sentence.

2. Comments have already been really good

Since I started writing, I’ve learned at least two important things from readers:

1. There’s a team in New York working on the worm simulation problem, following Moore et al. 2024 which models neurons as “data-driven controllers” (thanks to Nicolas R for mentioning it). Brief thoughts from me in the footnote.2

2. My understanding of the brain aligns with Philip Ball’s thoughts on genetics, especially when it comes to biological emergence and agency.

   Philip Ball is a science writer and former editor at Nature who published a book titled How Life Works in 2023. Check it out, and thanks to ozzie for recommending (see the footnote for a teaser3).

3. You guys are reading this in your free time

Lots of reviewers just don’t care much.

Editors assign them papers (although they can decline) and then they do their academic community service. Some reviewers are meticulous and do care, but I’ve found these uncommon.

In contrast, Wikipedia is existence proof that people critiquing stuff in their free time can be incredibly good for accuracy.

Don’t Meet Your Heroes

Or, how Randy Pausch became a contributor to the World Book

Randy Pausch was a professor in computer science at Carnegie Mellon University from 1997 to 2007. He was diagnosed with pancreatic cancer in 2006, and when he had a few months to live, he gave a final lecture at the university titled, “Really Achieving Your Childhood Dreams.” The lecture was extended into a now-famous book, The Last Lecture.

In Chapter 8, Randy talked about how he used to dream of contributing to the World Book Encyclopedia. The World Book was one of his favorite things as a kid.

Then Randy grew up for a few decades. He got a PhD in computer science; he got a tenure-track job. And that’s how Randy found himself in the position to finally achieve his childhood dream. As in, he was

• important enough to have his expertise recognized,

• but not so important that the World Book staff thought he’d be busy when they called him.

An excerpt from The Last Lecture:

But it’s not like you can call World Book headquarters in Chicago and suggest yourself. The World Book has to find you.

A few years ago, believe it or not, the call finally came.

[…] “Would you like to write our new entry on virtual reality?” they asked.

I couldn’t tell them that I’d been waiting all my life for this call. All I could say was, “Yes, of course!” I wrote the entry. And I included a photo of my student Caitlin Kelleher wearing a virtual reality headset.

No editor ever questioned what I wrote, but I assume that’s the World Book way. They pick an expert and trust that the expert won’t abuse the privilege.

I have not bought the latest set of World Books. In fact, having been selected to be an author in the World Book, I now believe that Wikipedia is a perfectly fine source for your information, because I know what the quality control is for real encyclopedias.4

Pausch, Randy; Jeffrey Zaslow. The Last Lecture (p. 42). Hachette Books. Kindle Edition.

In case you’re curious, there have been more than 50 edits to the Wikipedia page on virtual reality (VR) in the last seven months (since January 2025). Here’s one of them.

Except VR has been in the zeitgeist lately, so maybe it’s not a fair comparison. Let’s look at the page for sharks instead. Sharks have not changed much in the last million years. That hasn’t stopped people from making over 100 revisions to their page, also since January.

Here’s one from two weeks ago.

To conclude this section, this is a story in favor of Wikipedia from my own life (albeit less poignant than Randy Pausch’s).

Back in 2019, a friend made a Wiki page for another friend for his birthday. I’ll call the second friend Anthony. Anthony’s birthday present got taken down within ten minutes because he wasn’t important enough.

I was quite impressed by this—it was totally the right move, since Anthony didn’t have any Wikipedia-level accomplishments to his name. (Sorry Anthony—you don’t have any Wikipedia-level accomplishments to your name yet).

I’m not saying my personal blog is going to be the rigorous gauntlet that is Wikipedia. But I do think it’ll be better than formal peer review.

4. There aren’t dumb incentives

Peer review, while nice in principle, is warped by a Goodharted system of rewards.5 Scientists have to publish their work, so they’ll do whatever they can to convince editors and appease reviewers.

Sometimes this means adding unnecessary data that clutters up the story. Other times it means adding unnecessary citations, like in the example below. Reviewers boost their own citation counts like this pretty often, though there’s rarely so much drama about it.

But you don’t benefit from quashing my ideas on some random blog and I don’t benefit from brownnosing. Egos aside, I think we both want my writing to be better and for me to fix it if it’s wrong.

You might think, but wait, can’t you just not listen if you don’t want to?

Yes, and I think that’s a good thing.

Consider Pixar’s “Braintrust”

Ed Catmull, the co-founder of Pixar, touted the Braintrust as one of Pixar’s best practices in his memoir Creativity, Inc.6

The Braintrust was a sort of master review committee that came out of Toy Story, and the whole idea was that it was a committee with no authority.

A director would present a work in progress and the Braintrust would give feedback. “This is crucial”, wrote Catmull. “The director does not have to follow any of the specific suggestions given. After a Braintrust meeting, it is up to him or her to figure out how to address the feedback.

“To set up a healthy feedback system,” he said, “you must remove power dynamics from the equation—you must enable yourself, in other words, to focus on the problem, not the person” (emphasis mine).

This is vastly different from peer review, where reviewers and editors have all the power, and the burden is on scientists to convince the all-powerful of how their work should be written.

And while Pixar is about art, and science is about truth, I think there’s a lot of overlap in how to get the best versions of both.

Counterargument: but shouldn’t experts be the ones reviewing science?

Formal expertise is good. It’s worth noting that the first Braintrust was made up of only storytelling legends (e.g. the directors of Toy Story, Finding Nemo, Cars, Wall-E). And being in “the system” for the last ten years has taught me basically everything I know; you learn a lot being around experts who are thinking intensively about the same problems you are.

But things could be better.

Like, I don’t like how experts can be hyperspecialized—people are way too focused on way too specific areas of science. I think hyperspecialization is helpful when we already have a decent picture of a field and are trying to fill in details, but neuroscience is definitely not there yet.

So I’ll take any help I can get. I want to hear from neuroscientists outside my building, scientists from other disciplines, and from people who read this stuff in their free time.

Plus, this is still my blog. I get to decide which comments to listen to. If you’re an expert who disagrees with me, let me know. Then write it on your own blog.

To summarize,

A few months ago, I learned about Adam Mastroianni, whom I cited earlier. Adam is a psychologist who stopped playing the academia game and has been publishing his work on Substack instead.

Not to be creepy about it, but I think his life is really cool and I want to copy a lot of it. Except I’m not a psychologist; I’m a computational neuroscientist who hopes that by posting here, getting feedback, and revising, the work will be better. And that it will reach more people than if I did things the usual way.

Hence my experiment: can Substack be a better version of peer review?

(A simultaneous experiment: Can I write a blog that people will keep reading? Subscribe to find out.

And I am revising my last post. Out within two weeks.)

Part 2 of a series, which may not make sense until you read part 1.

Also, my mom caught COVID and didn’t feel like talking about radios. More from her in the next essay.

Read more

Chatbots can’t reliably tell facts from lies, but I still use them a lot for work. They’re pretty good already, and if we throw even more data at them, they might get a lot better.

That’s why I’m so confused.

Because for the last four years, I have bothered as many experts as I could—researchers from Harvard, MIT, UCL, and various companies—and everyone at least acknowledges that there’s a weird gaping hole in the field, way on the other end of the intelligence spectrum. And this is that, after decades of trying,

nobody can build an algorithm that works like a worm.

Okay it’s not that big; in fact it’s very small

When I say “worm” I mean the most thoroughly studied animal in the world: the nematode Caenorhabditis elegans.

Nematodes are miniature worms and “the most abundant animal on Earth.” They like to live in vegetable gardens or deciduous forests (ones where the leaves change color). In these forests there are about 200 nematodes per 10 grams of soil.

Different nematodes vary in size, but C. elegans grow up to be about a millimeter long. They eat E. coli, which suggests that, aside from having lots of experiments done on you, being a C. elegans in the laboratory might be like living in a large pile of gnocchi.

We can’t build a worm. But how hard have we tried?

Answer: Pretty hard.

The biologist Sydney Brenner picked C. elegans as a model organism in 1965 because it seemed relatively easy to figure out.

We have 60-100 billion neurons.1 Fruit flies have about 140,000. C. elegans only have 302 neurons, which is more followers than I have on Instagram, and just like my friends on social media, there are few enough for someone to recognize each one’s name and identity.

This sort of labeling is impossible in bigger animals, where every nervous system is too different across individuals. The worm nervous system’s size and consistency meant that we could map out the whole thing pretty quickly—the first map came out in 1986, almost forty years ago, and you can see a recent version on the WormWideWeb.

Since the mapping, many have tried to build the worm. All have failed. Here’s a non-comprehensive list of worm simulation projects.

• ~1997: NemaSys, The University of Oregon

• ~1998: The Perfect C. elegans Project, Sony, report

• ~2004: The Virtual C. Elegans project, Hiroshima University, papers published

• Early 2010s: The OpenWorm project, MIT, website

After OpenWorm, there were either fewer worm simulation efforts or people made less of a big deal out of them (just in case). Regardless, the pattern seems to be that every five years, the worm will be fully simulated in the next decade.2

What about in recent years?

In 2024, a paper came out of the Beijing Academy of Artificial Intelligence and Peking University.3 It was titled, “An integrative data-driven model simulating C. elegans brain, body and environment interactions.”

But the paper models just 136 out of 302 neurons in the worm. And the simulation can only zigzag forward. Worst of all, it doesn’t learn, which is a glaring omission since even in the incredibly stable nervous system of C. elegans, we see differences between individuals.

That gets us to the year 2025—a year in which, four months ago, WIRED published a feature titled “The Worm That No Computer Scientist Can Crack.”

Why I don’t care about the worm, except that I do

The worm was never the point. It was a model, not the goal.

When I was a graduate student, I grew thousands of worms and watched them for thousands of hours. In the process, I noticed some surprising things.

Even between genetic clones, some worms are fast and some are slow. Some are sleepy or really like to eat, and some like being next to walls, which I think is a valid personality trait. It’s sad but adorable when worms get scared; they curl up into circles like Cheerios. And if they get stressed or learn too much, they take a nap. This is a tactic I admire. They take so many naps!4

And worms can move. They use an intricate set of muscles to wiggle around in complicated ways. This kind of thing is still hard for robots today—why are the most advanced robotics displays still rigid and choreographed compared to the tumult a worm goes through when you dig up plants in your garden? The robots are impressive unless we compare them to animals—we have such high expectations for animals that it’s only special when they can’t get up again.

Other things worms do: they swim towards food; they remember what makes them sick; they sleep; make decisions; they invent crazy new behaviors. Did you know that worms can hitchhike by organizing themselves into tiny worm towers that help them teleport onto bees using static electricity?

All animals, including the smallest ones, have these endlessly interesting traits—flexibility, autonomy, and inventiveness in what they do. When I say I want to build or understand a worm, these are the things I’m talking about.

But how do we go about it?

Can a physicist understand a radio?

Here’s an analogy—let’s say I give you a radio.

No—I give you a thousand radios.

All of these radios are different, but they catch local stations and play sounds with a little tuning. Some are the same but different colors. Others are AM or FM or both. A few are straight-up iPhones.

Let’s say you don’t know what a radio is. You have no idea what it does, what radio waves are, or that they can carry information.

Let’s say I want you to figure out:

1. Can you fix a radio?

2. Can you build a radio?

3. What are radios?

Wisdom and encouragement from my mom

My mom is a nuclear engineer who knows a lot of physics, so I asked her these three questions. I told her that this is what my job feels like—each animal has a different kind of radio in its head and/or body, and neuroscientists are trying to figure out things about them. Some neuroscientists want to fix radios; some want to build better radios. Others, like me, are just trying to understand them.5

To which she responded, in English:

But I kept asking and asking her, and eventually she conceded it might be possible to answer parts of these questions in the right circumstances.6 For context, she has built a lot of radios, because my grandpa used to make them with her when she was a kid.

Here is what she said to each question.

Question one: How do you fix a radio?

My mom: “Give up”

Out of the three tasks—fixing, building, and understanding—my mom thought the hardest was to fix one of the more complicated radios.

Specifically, she said:

This sentiment is in firm agreement with a paper from 2002, “Can a biologist fix a radio?”

Can a biologist fix a radio?

In this paper, a scientist at Cold Spring Harbor named Yuri Lazebnik considers what he calls “David’s paradox.”7 David’s paradox is that:

The more facts we learn the less we understand the process we study.

… For example, the mystery of what the tumor suppressor p53 actually does seems only to deepen as the number of publications about this protein rises above 23,000.

This worried Yuri, because, as he wrote, “the notion that your work will create more confusion is not particularly stimulating”. He then remembered his high school math teacher “who recommended testing an approach by applying it to a problem that has a known solution.”

So Yuri got a radio that his wife had brought from Russia. He pretended that biologists across the world were assigned to understand and/or fix the radio.

Because a majority of biologists pay little attention to physics, I had to assume that all we would know about the radio is that it is a box that is supposed to play music.

What would the biologists do?

First, they would get as many copies of the radio as they could. Then they’d shoot them “at a close range with metal particles.” Lots of these radios would break, and the biologists would see which broken components led to what kind of damage.

They’d also pick the radios apart and try to find out what pieces were necessary for them to work. Getting rid of some pieces wouldn’t do much, but “a lucky postdoc will accidentally find a wire whose deficiency will stop the music completely.”

This important wire, Yuri said, would be named and published. Researchers would discover that it was important because it connected “a long extendable object”—the antenna—“and the rest of the radio.”

Inspired by these findings, an army of biologists will apply the knockout approach to investigate the role of each and every component. Another army will crush the radios into small pieces to identify components that are on each of the pieces, thus providing evidence for interaction between these components.

Eventually, all components will be catalogued, connections between them will be described, and the consequences of removing each component or their combinations will be documented.

So far, so good—progress has been made. Thousands of publications have been churned out. Dozens of biologists will ride this wave all the way to tenure.

But then, Yuri says, the biologists will hit a wall.

The wall will appear upon a critical mass of biologists asking,

If a defect is simple, maybe they can. If a component “smells like burnt paint”, it should probably be replaced. But “if the radio has tunable components, such as those found in my old radio and in all live cells and organisms,” says Yuri, then all bets are off.

What is the probability that this radio will be fixed by our biologists? I might be overly pessimistic, but a textbook example of the monkey that can, in principle, type a Burns poem comes to mind.

As in, it is possible that a biologist could fix a radio.

But no way is it ever going to happen.

End of paper.

Fifteen years later in 2017, Yuri’s thought experiment inspired the neuroscientists Eric Jonas and Konrad Kording to write a follow-up: “Can a neuroscientist understand a microprocessor?”

The answer was again no.

Question two: How do you build a radio?

My mom: “First buy the parts at the store”

This response raises the question: what are the parts to a working brain?

We’ve had the worm nervous system mapped out for forty years. We kind of know how neurons talk to each other through electrical and chemical signals. But it turns out that’s not enough.

A brief digression into neuroscience history

Back in 2012, the OpenWorm project was still in its infancy. A team of very accomplished neuroscientists, having seen that making a worm in a computer was kind of hard, decided they were going to go for a human brain instead. They wrote an impressive proposal and won 1 billion euros in funding.

A lot of other neuroscientists were very upset about this. Many thought the idea was fundamentally flawed and that it was a huge waste of money.

For instance, in 2013, a New York Times journalist asked Haim Sompolinsky what he thought about the endeavor. Haim is now partly at Harvard and won the Brain Prize in 2024 (the biggest award in the field).

“The rhetoric is that in a decade they will be able to reverse-engineer the human brain in computers,” said Haim. “This is fantasy. Nothing will come close to it in a decade.”

(Haim is not known for mincing his words.)

Haim and others are interviewed in a documentary on the rise and fall of the Human Brain Project, the greatest movie about computational neuroscience research that I have ever seen. It’s called In Silico and was filmed by Noah Hutton, released in 2020.

Hutton started as a believer—he originally planned to promote the film as a behind-the-scenes for one of the greatest scientific achievements in history. But instead, In Silico was described by a Nature Arts Review with the headline

Because implode it did.

The Human Brain Project got nowhere for years.8 Publications trickled out, but an early simulation of a piece of rat brain was described by the neuroscientist Alexandre Pouget as “completely meaningless, just random activity… The claim that he simulated a rat’s cortex is completely ridiculous” (2013, New York Times).

By 2023 the project had turned into a very expensive art installation.9 At least at the end of the day, neurons still look cool even if they don’t work—below is a picture not from the Human Brain Project, but a different exhibition I thought was better.

—

This is to say, we’ve had some huge, expensive, and very public failures in computational neuroscience. We’ve not only failed to simulate a human brain—we’ve failed to simulate a fully mapped 302-neuron nervous system over and over again.

My mom says to build a radio, you have to first buy radio parts at the store. But at this point, we might not even know what the radio parts are, let alone how to put them together.

How come? Why can’t we do it? What makes this little problem so weirdly hard?

This post was getting way too long, so in Part 2 and Part 3:

• What many experts have said when I asked them why we can’t build a worm yet

• Why some people might find a worm simulation useful, but I’m not sure I would

• How the worm simulation projects so far have treated nervous systems like circuits when they probably aren’t like circuits at all

Someone once told me about their friend who’d convinced MIT to pay for them to go to Burning Man.

The first time he attended the festival, he told the school he wanted to study its unique gifting economy. But when he got back, he “realized” he’d either lost some data or really needed to collect more. This meant he had to go again.

Halfway through my PhD, I was restless; I didn’t think I could go on crazy research adventures like Burning Man Guy because I was in a biophysics program. (In retrospect that was my bad, because someone in my cohort spent a lot of her degree hiking through Chilean deserts to study sweaty shrubs.)

Instead I trained for a marathon, which I failed to run. But it’s about the journey, I guess, and I did get to listen to some really long audiobooks along the way. One of these books was the offspring of anthropology fieldwork by PhD student Michael Crawley,1 and it was about how he moved to Ethiopia to train with the best distance runners in the world.2

Question: Just how much better are they?

Answer: A lot better

Crawley has a marathon time of 2:20, or an average of 5 minutes 20 seconds per mile.

One morning in Ethiopia, on a ride to a training session, Mike strikes up a conversation with the bus conductor Tadesse.3 Tadesse had tried running for a while but decided he wasn’t good enough to take it seriously. The two share race times, which makes Mike rather existential.

This is perhaps the moment I realise the extent of the gulf between me and the other runners on the bus. Even the bus conductor is two seconds quicker than me over 10 kilometres and he ran his [personal best] at 2400 metres above sea level.

At 2400 meters, using this calculator, Tadesse might have been 21 seconds faster than Crawley for a 10k and then decided he couldn’t compete in Ethiopia.4 Meanwhile, Crawley raced for Great Britain—he still carried the gear with him and everything. This was very confusing for the Ethiopian athletes (see cover art above).

“For enjoyment only”

To pile on, Mike finds out how Ethiopians really think about domestic versus international races. The following comes from Fasil, one of Mike’s training buddies.

“When you go outside,” Fasil points out to me, “there may only be six or seven athletes from Ethiopia and a few from Kenya, and the rest of the people there will be running for enjoyment only.”

I’m not sure that this is how many serious club runners would think about it necessarily, but I let him finish.

“When you run a race in Ethiopia, though, you have to fight with hundreds of strong athletes.”

Fasil basically thinks that if everyone else was taking running seriously, then shouldn’t they be better at it?

And he might not be completely wrong. Because while Ethiopian athletes have clear genetic and environmental advantages—elevation and a strong infrastructure for professional athletics, for instance5—Mike found that they trained completely differently, too.

How to run fast, if you are in Ethiopia

Through Fasil, Mike met an athlete named Birhanu Addissie.

Birhanu wasn’t performing as well as he wanted recently. Part of the problem was that Birhanu had gained about ten pounds from eating too much pizza. Mike knew this because he watched him eat a lot of pizza; also, other runners had started making fun of him by calling him “Birhanu Pizza.”

But Birhanu wanted to win the 2015 marathon in Rome, which was in five weeks.

Step 1. Make it harder

Step 1a. Find the thinnest air

To train for Rome, Birhanu needed to find the right places to run. Mike asked to tag along on one of these training runs to a place called Arat Shi. “Arat Shi” literally means 4000 meters, and this worries him.

“Surely we’re not the equivalent of almost halfway up Everest and planning on running uphill?”6 Mike asks.

I wanted to know what it was like to train at that altitude, too—but vicariously, of course. So I found a YouTuber who did an exercise challenge at 10,631 ft above sea level. He completed the challenge, but then needed aid and cried uncontrollably.7

When Mike asked Birhanu whether Arat Shi might be too difficult, Birhanu replied,

‘Hard is good, Mike. This is marathon.’

[…]

‘And we’d better take Fasil with us,’ [Birhanu] adds after a few moments thought.

‘Why Fasil?’ I ask.

‘Because Fasil,’ he says, ‘is crazy.’

‘And crazy is good?’ I query.

‘Crazy is good.’

1b. Follow a crazy person

Fasil shows up in the book a lot. Everybody finds him a perplexing sort of guy.

This is how Fasil became a runner.

Only a year prior, Fasil was looking for a way to make some more money in Addis Ababa, the city that “cleans faces and pockets”. He could play the mesinko and thought he could make a bit extra singing songs.

Fasil went to the market to buy one, but he got distracted by a shop with running shoes.

“So I put them on and jumped up and down, and they were so bouncy I felt like I could run for ever. So I spent all the money I had on the shoes and I had to run all the way back to Kotebe, because I had nothing left for the taxi fare.”

The next day, Fasil ran into the forest with his shoes and started following the first group of people he saw. They ran so far that he had to take a bus back, but he didn’t have any money.

“Luckily I met a girl who invited me in for some milk and honey on their farm,” Fasil says, to which his friend Hailye raises an eyebrow. “She gave me the money for the bus”.

“You see,” says Hailye, who had introduced Fasil to Mike. “This is why I wanted you to meet Fasil. He’s a bit different.”

This was why Birhanu didn’t just want Fasil to join the training session at Arat Shi—he wanted him to lead it. Crazy is good, and Fasil would create the runs Birhanu needed to perform in Rome.

1c. “I’M NOT COLD, I AM A VERY DANGEROUS MAN!”

A sentence yelled during a cold shower.

One morning at 3 a.m., Mike wakes up after four hours of sleep and sees Fasil washing his face at the tap.

“It is pitch black, and my breath turns to mist in the cold air as our dog barks at me,” Mike writes in his thesis. But he’d agreed to join on this run. Fasil beams at him in running gear, telling him, “You’re no foreigner, you’re a hero.”

“Are you tired?” I ask Fasil in Amharic.

“I am not tired!” he exclaims, grinning.

They do an hour of hill sprints and return for a cold shower before going back to sleep.

What was the point? Mike wonders.

The argument for being dangerous

“Disenchantment is the distinctive injury of modernity,” wrote the sociologist Max Weber in 1946.8 A major theme of Crawley’s book is how Ethiopian runners fight this disenchantment. They talk about air being “magic” and about “mysterious incalculable forces” giving runners energy and life.

Crawley’s training back home was not as magic.

We run with heart-rate monitors and GPS watches at a carefully planned pace. We upload the GPS data about our running to apps like Strava… Sports scientists test our top athletes in order to determine their physiological parameters.

Calculated optimization takes a lot of the magic out of the sport. But if you make life especially difficult once in a while, just for fun, it can make you feel like you’re on the edge of what’s possible—like no one else is “dangerous” enough to do what you do.

Back at Arat Shi, even before the run begins, they encounter some difficulties. It had rained really hard the day before and their shoes were caked in mud.

“Fasil stops to carefully scrape it off,” writes Mike, “which seems to me a futile gesture as we are about to be running through it for two and a half hours.” And then when the run begins, Mike finds out what “crazy” really means.

“Fasil suddenly decides to loop right round a tree and go back on himself, or make a 90-degree turn by ducking under a low branch.” He leads them into a ditch they have to climb their way out of, through “bushes with inch-long thorns,” and one of them lodges itself in Mike’s head.

All in all, it’s exactly how Birhanu wanted his training run to go. But not Mike, who’s so frustrated he wants to peel off and choose his own routes. Fasil says he can’t, though.

“‘No, no,’ he tells me. ‘There are way too many hyenas around here.’”

Mike looks back at Birhanu to see how he’s doing, and Birhanu is “floating along with a serene look on his face. He is wearing two tracksuits to make things a bit harder, but there isn’t a single bead of sweat on his forehead.” He sees Mike checking in on him and smiles back in encouragement.

Two-and-a-half hours and 19 miles later, they stop to cool down. Mike writes, “The outsides of my shins are throbbing from the effort of staying upright,” but “Birhanu bounces from foot to foot like a boxer.”

“‘Tomorrow, three hours!’ he beams. ‘This air is nice!’”

2. How to run fast by making it easier

Despite the torture run, Mike noticed there were a lot of ways Ethiopians made it easier to train.

2a. Consider holding hands

After the run at Arat Shi, Mike discovers they have to walk back to where they started. He throws a tantrum.

Fasil keeps giving me concerned pats on the shoulder.

‘Dekemah zare?’ he asks me every couple of minutes. ‘Are you tired today?’

‘Yes,’ I tell him in English, ‘and I could have done without you choosing the most difficult, dangerous route possible, and without having to trek for hours to get home!’

He smiles wryly and just says, ‘Ayzoh, ayzoh,’ (keep going) and my irritation gradually lessens as we walk down the hill. I feel like a petulant child on a long car journey. I trudge slowly.

Fasil looks absolutely fresh, and walks with a spring in his step and a swing to his arm. As we walk down the hill he takes my hand affectionately. Ethiopian men often hold hands, something that initially surprised me but which I’ve got used to now.

‘Ayzoh,’ he says again.

That run was hard, but it’s over now and we can rest.

2b. Zigzags for everybody

For Mike, the most dramatic difference between running in Ethiopia compared to the UK was that nobody ran alone. In fact, it’s seen as “deeply antisocial and borderline suspicious.”

This feature of Ethiopian running turned out to be really useful, because just like Fasil, Mike found his training group by going into the forest and following the first group he saw. Unlike Fasil, though, he couldn’t keep up.

Before I know it, the line of runners comes up alongside me and I am grabbed by the wrist and guided back to my place at the back of the queue.

‘We run together,’ the runner closest to me says in English.

I resign myself to following the group, and the purple tight-clad pied piper at the front, as best as I can.

The strangers yell encouragement at him when he slips behind again— “Ayzoh, ayzoh!”—and finally, when the run is finished, he introduces himself to a runner named Tilahun.

Tilahun has the following questions:

1. Where are you from?

2. Why are you here?

3. What is your 10k time?

4. Why did you want to run alone?

Later, Mike’s new friends bring him to a training field. To his astonishment, he sees a few hundred runners with no separation by ability—kids were running next to 2:05 marathon champions, and he thinks he spots a teenager who’s just wandered over from a nearby farm.

The train of runners meanders around the field, the leader turning almost 180 degrees every minute or so and the rest following like a shoal of fish.

‘Why do they zigzag like that, Meseret?’ I ask.

‘I don’t know,’ he replies… ‘They learn that from each other. No-one is telling them to run like that.’

Mike’s theory is that the bright athletic jackets attract newcomers like flowers do bees, and the zigzagging lets them cut corners while the experienced runners don’t. Whatever the reason, these training sessions let everybody run on the same field. And after a hard practice, they sit on the grass and hang out, all together.

2c. “As interesting and inspiring as possible”

When I was in Cambridge, MA, I always ran the same route along the Charles River. I ran at the same time on the same sidewalk, and if I were in the mood, I might go clockwise instead of counter-clockwise.

To cope with the boredom I watched the geese that swarmed the riverbanks, taking pictures of them to post on Strava with updates about their social lives.

But in Ethiopia, coaches worked hard to make the training itself engaging. Running on asphalt was restricted; it was much more important to learn how to respond to the texture of rocks, or gravel, or overgrowth on the ground; to learn how to lean into the way the land made you move.

“The successful ones,” a coach says, “are the ones who watch with their eyes and think with their minds before they move their legs.”

Some time before Rome, Birhanu moves onto gravel road. Even the gravel is uneven and difficult—the roads are “never flat and you still have to think carefully about where you are putting your feet. It is still very hard work.” And after five weeks of 3 a.m. hills, hand-holding, and hyena bramble torture runs, Birhanu sidles over to Mike one day and says,

“I think I am winning in Rome.”

And he was very, very close to being right. Birhanu “Pizza” Addissie won silver in the 2015 Rome marathon, followed immediately by gold in Taiyuan.

A very smooth segue

Now I’m going to see whether I can learn from all of this and apply it to my own job as an academic researcher. Because running until you throw up to become a world-class athlete is kind of like sitting at a desk reading papers, maybe.

What if we did more together?

Reading papers is boring, but I realize now that I always read them alone.

I’ve never done a lit review with somebody, in real time. Wouldn’t it be nice to watch an expert look through the literature, extract what they need, and see how they decide what matters?

As a counterexample, look at software engineering: in classes or jobs, there’s a concept called “paired programming” where you share your screen with someone and they watch you code. The first time I did this for a class I hated it—my partner had worked at Google and I’m self-taught, which made me very self-conscious. But then one moment changed my mind.

It was when, during a class session, we came across a line like this:

acts, weights, noise = np.zeros(n), np.zeros((n,n)), np.zeros(n)

He clicked his tongue and said, “This is why Python is pretty.”

And I thought, huh. I guess it is.

What is the reason for being dangerous?

For every boring thing out there, there are people who have convinced themselves it’s not that boring. This is the whole premise of meditation. And to get yourself to do boring things, it helps to feel some sharp intensity of purpose behind them, to feel that the importance of your goals are exemplified by the crazy lengths you go to achieve them.

But for scientists, what is that goal? Why are we putting ourselves through the gauntlet of academic language?

For the Ethiopian runners, the goal was to win races and break records, which they did by training in thin air and on tricky ground—but that wasn’t where they stopped. To run really fast—to actually break those records—they had to get off the mountain.

What is science for?

Scientists are supposed to be learning, discovering, creating concepts that other people can leap off to create even more. There are so many possible inventions and undiscovered truths, and we’re supposed to be trying to work them out, together.

Instead, we train and train by reading these indecipherable papers; then we graduate and spawn more of our own. It’s like we’ve all decided that running doesn’t count unless it makes you cry by yourself at 10,000 feet above sea level. If that’s the sport we’re trying to win, defined by endless nonsensical obstacles, then today’s scientific writing style is perfect; it makes us feel good and powerful and dangerous to tell ourselves that only we can perform at these heights, whatever that means.

But look. Clearly that’s not the point. And obviously there’s some problem with how we’re doing scientific communication, because fact that there are anti-science movements out there reflects that we’ve done a really shit job explaining what we know, because “science” is just what we think is probably true about the world, and why we think those things. The whole point is to (1) find stuff out, and (2) make it make sense.

So shouldn’t we all just get off the mountain and write in a way people can (and want to) read?

If you want to make a scientist angry, explain something to them poorly. I know this because I have done it a lot. But it’s not just scientists; it’s that almost everybody gets frustrated at what they don’t understand. And despite my complaints here, I’m still part of the system and I’m still writing these papers. I love doing scientific research so much that I’m making it my career—but the reality is, I don’t even like reading my own work.

So why should anyone else listen to me?

This is the first sentence of the first paper in this month’s Nature Neuroscience issue (June 2025):1

Autism spectrum disorder (ASD) is often explained as a manifestation of the atypical integration of diverse neural information.2

This is the first sentence of the second paper I clicked on.

External context can change the processing of stimuli through recurrent neural dynamics.3

#

To compare: here is the preface of Mind of a Mnemonist by A.R. Luria. The book was published later but the preface is dated 1965, and this is a good soundtrack for it.

I spent this summer off in the country, away from the city. Through the open windows I could hear the leaves rustling on the trees and catch the fragrant smell of grass.

On my desk lay some old, yellowed notes from which I put together this brief account of a strange individual: a Jewish boy who, having failed as a musician and as a journalist, had become a mnemonist […]

[Yet he] remained a somewhat anchorless person, living with the expectation that at any moment something particularly fine was to come his way.4

I have a problem

I was going to write and think about how

• research papers used to be much more interesting to read,

• being interesting is good for science,

• personality and metaphors contribute to creative thinking,

And on and on. But I poked around and there are way too many blog posts and opinion pieces that have said all of that already.5 I’m not going to belabor the point; at least, I won’t do it in exactly the same way everybody else did. Everyone already seems to agree that, as research papers go, “all of it is written like it hates you”.

Instead I’m going to see whether I can defend the way that modern science is written. Because if everyone thinks it’s the worst, am I at least getting something out of it?

The case for desirable difficulties

Reading academic writing is hard, but people say that hard things are good for us.

So here I’ve made a list of situations where hard things help. Maybe if I know how something makes my life harder, then I can see whether it’s making me better. If it’s not making me better, then I might as well stay lazy. Speaking of lazy,

#

It’s a feature, not a bug

Example A. Lazy eyes make good artists

I first heard this from Marge Livingstone. Marge is a professor at Harvard Medical School, where I’ve heard her called “the best monkey electrophysiologist in the world.”

The claim: Artists are more likely than everyone else to have strabismus, a condition where your eyes don’t point in the same direction.6

The evidence: Marge checked the photos of a bunch of artists and counted how many of them had misaligned eyes. She talks about it in this lecture from 2014.

Transcript:

So, we got photographs of several hundred famous artists and looked at their eyes… only 3% of the [general] population has significant strabismus. You would expect that we would have found maybe two or three artists who didn’t have aligned eyes.

Instead we found dozens.

She lists a few, including the artist Kara Walker who works with silhouettes—“how flat can you get,” says Marge. There’s Pablo Picasso and Frank Lloyd Wright; she also makes a case for Rembrandt.

Da Vinci probably had strabismus, too.7 Too bad this result was published in an ophthalmology journal in 2018, and is therefore filled with sentences like,

“Some forms of strabismus are thought to facilitate artistic work by suppressing the deviating eye, which creates 2-dimensional monocular vision advantageous to painting and drawing.”

(Or: Artists with strabismus see in 2D instead of 3D like the rest of us. That helps them get things down on paper.)

What’s my point?

At face value, it would seem like eyes that aren’t aligned would be bad for a career where you have to use your eyes a lot. But when one looks a little closer, one finds that it’s actually an advantage.

Example B. The giant panda is cryptic

People make fun of giant pandas because they seem suboptimal. One sticking point seems to be their dramatic black-and-white patterning, which looks like somebody tried to teach them about camouflage but then they misunderstood it.

Here I’m paraphrasing a 2017 paper by Tim Caro et al.8 The original paragraph is in the quote’s footnote (it’s not that paraphrased).

We think the coloring of the giant panda is because it eats nutritional garbage that it cannot digest. This means it cannot build up enough fat to hibernate (even though other bears can manage easily) and it is forced to be awake for the whole year.

The problem is, giant pandas live in places where sometimes it’s snowy and sometimes it’s not - their surroundings change colors very dramatically. But unlike the black bear, who is a much better bear, the giant panda is too slow to change coats by shedding its fur.

It therefore cannot change color with the seasons, and instead makes do with a bad black-and-white compromise.9

#

A giant panda redemption arc

But! The same man, Tim Caro, published another paper in 2021, this time as last author. And this time, it looks like Tim and his co-authors are convinced that pandas are for hiding.

They titled their paper “The giant panda is cryptic” (cryptic here in the ecological sense, meaning being able to avoid detection). They used some image analyses to claim that, among other things, pandas are patchy in the same way that their natural surroundings are patchy—but only if you look from far away.

Compare the bottom row of pandas above to these tigers from the internet. The last tiger is in its natural habitat.10

A few thoughts based on the above:

1. Giant pandas do kind of look like lumpy rocks from far away.

2. This is partly because of their enormous heads.

3. It might be okay that the camouflage only works from a distance. Tigers have to sneak up on their food, but bamboo can’t see.

In the 2021 paper, Nokelainen et al. write:

In summary, the giant panda uses black-and-white pelage [meaning coat] to avoid detection in its natural habitat and its apparent conspicuousness to the human eye is an artefact of the short viewing distances found in photography and zoos.

Tim has gone from being really rude to his giant pandas (2017) to declaring that they were never meant to be in zoos (2021). I’m not not convinced.

Does academic writing count? Is it secretly helping?

With artists and giant pandas, the difficulty seemed like a challenge to overcome, but it was only an illusion. In the right context—out of the zoo and in the right lighting—the trait was the opposite.

But I’ve never encountered a context that could make me think the paragraph below is easier to read than, I don’t know, a Reddit thread. Or that I understand it better because of the style:

Animals rely on visuomotor transformations to convert object locations in eye coordinates into directional movements⁸. The brain regions and neural circuits regulating this transformation in both vertebrates^9,10 and invertebrates¹¹ have been characterized. The precise neuronal connectivity underlying learned^12,13 and innate visuomotor¹⁴ tasks is shaped by genetically hardwired mechanisms, experience or both.

[Dombrovsky et al. 2025, Nature, citations kept from original source11]

But isn’t it more precise? More compact?

One common argument is that this type of writing is a necessary evil—it’s just how scientists need to communicate, because they handle such complexity in their topics. I don’t think that’s true. As an example, I’m going to try paraphrasing the paragraph above:

Animals

—what? I can’t even work out the first sentence. I’ve spent the last ten years in this field. Maybe I need context clues from the rest of the paragraph.

Nothing there. Maybe an example will help me.

When a cat sees a toy, it transforms

Visuomotor is vision to movement, but why are we converting where something is to a movement? Is it the movement of the thing or is it our movement? Movement of the thing, probably.

When you watch a moving thing, you have to transform

It doesn’t feel right; I’m going to go to the full paper. This feels like cheating.

Wow, okay, I was wrong; it’s in the abstract: “How does the brain convert visual input into specific motor actions?” It’s the other one, our movement—

I have to decide how to move, based on how something else I’m watching moves.

Why the heck is it worse in the main text? This is not fun. It feels really pointlessly hard.

Sometimes I have to decide how to move, based on how something else I’m watching moves.

We know where in the brain this happens for both vertebrates and invertebrates.

The exact neural connections involved are either nature, nurture, or both.

I would argue that the only information missing in this paraphrase is the eye coordinate to movement transformation:

Final paraphrase:

Sometimes I have to decide how to move based on how something else I’m watching moves. This means converting eye coordinate information to my own movements. We know where in the brain this happens for vertebrates and invertebrates. The exact neural connections involved are either nature, nurture, or both.

My version is 49 words. Theirs is 51.

The issue is, it doesn’t matter where I am or why I’m reading it—whether I’m at my desk, or presenting at a journal club, or in bed at home—reading a paper like this is always, to borrow a phrase from my advisor, “a big ball of suck.”

Which means this post didn’t get anywhere. But in the next post I want to look at a desirable difficulty that does apply to academic writing. Maybe I can figure out how to use it.

Anton Ego in the 2007 Pixar movie Ratatouille said, “Not everyone can become a great artist, but a great artist can come from anywhere.” I’m pretty sure the same applies to thoughts and ideas, so this is a place for me to make tenuous connections that would never escape Reviewer #2.

I’ll write about old papers, TV shows, things I hear while walking around in parks, all related in the end to the brain, probably.

The late great neurologist-writer Oliver Sacks wrote, referring to his best friend at Oxford, a mathematician: “He was attracted by my sometimes wildly associative mind, as I was by his highly focused mind.”

Sacks, Oliver. On the Move: A Life (p. 18). Knopf Doubleday Publishing Group. Kindle Edition.