The tweet is dated September 10, 2026. The connectome it describes was published in October 2024. Somewhere between those two dates, a fruit fly lost a dollar.
That is the whole story, compressed. A Coinbase software engineer named Alex Wormuth wired a full adult Drosophila connectome โ 166,000 neurons, 125 million synapses โ into an agentic trading interface, gave it a wallet worth roughly one hundred dollars, and let it buy and sell crypto for a single day. The account finished down about 1%.
The code spoke, but the metadata lied. Not maliciously. Sloppily. The article that surfaced this called the connectome "released last week." It was not. The dates do not reconcile. And nobody in the amplification chain stopped to check, because the claim was more interesting than the arithmetic.
I have spent fifteen years pulling smart contracts apart and reading transaction logs that contradict press releases. This is the first time I have seen the contradiction sitting in the header of the story itself, from the first paragraph, unexamined.
So let's examine it. Not the fly. The frame around the fly.
***
Context: what was actually built, and by whom
Start with the real asset, because it is genuinely impressive and it is being wasted as set dressing.
The FlyWire consortium โ Google Research plus the Howard Hughes Medical Institute, with a long tail of academic labs โ published a complete wiring diagram of an adult fruit fly brain. Every neuron. Every synapse. Roughly 166,000 cells and 125 million connections, reconstructed from electron microscopy slices over the better part of two decades. It is a landmark in connectomics and it matters for disease modelling, for circuit-level neuroscience, for the slow grind of understanding how nervous systems compute.
That is the science. It shipped in October 2024. It is public. It is real.
Now the engineering layer. Coinbase has been shipping agentic tooling โ programmatic interfaces that let external software perceive a market state, decide, and execute orders without a human clicking a button. That is a legitimate product direction. It is also, mechanically, a set of authenticated API calls.
Wormuth took the fly connectome, mapped its dopamine-reward circuitry onto a decision signal, and let that signal drive the agentic interface. Buy when the reward pathway fires. Sell when it does not. The wallet held roughly 59 USDC, 5 BTC, 17 ETH, and 17 SOL. Total notional: about a hundred dollars.
One day of trading. Net result: negative one percent.
Everything downstream of that โ the derivative demos, the meme coins, the billboard advertising โ is a separate economy that attached itself to those three facts.
***
Core: the autopsy
A connectome is not a trading system. It is a reflex arc.
This is the part that gets lost, and it is the only part that matters technically.
A fly brain does stimulus-response. Sensory input arrives. A circuit fires. A motor output follows. The dopamine neurons in the reward pathway modulate behaviour based on prior valence โ this was good, do more of it; this was bad, do less. That is a learning rule, but it is not a forecasting model. It has no concept of a time series. It has no position sizing logic. It has no drawdown limit. It has no notion of correlation between assets, no slippage model, no liquidity awareness, no idea that the order book it is hitting is thinner than the spread it crossed.
When you wire that into an execution API, you do not get a strategy. You get a random walk with a biological pedigree.
The -1% return is the tell. Over one day, with a hundred dollars, against four assets, a coin flip produces that outcome. So does a bug. So does a deliberate test with no edge. The sample is n=1, the window is 24 hours, and the position size is beneath the noise floor of every exchange's minimum tick. There is no statistical content in this result. It cannot validate the method. It cannot falsify it.
What it can do is generate a headline. "AI brain trades crypto." That headline is doing all of the work, and it is wrong in a specific, load-bearing way: the fly brain is not the AI. The fly brain is a signal generator. The AI โ if we are being generous with the word โ is Coinbase's agentic layer, which is executing a deterministic mapping from a neuron activation state to a market order.
Strip the biology out and what you have is a bot with an unusual RNG seed.
The engineering difficulty is entirely on the biology side, and none of it is on the finance side.
Reconstructing 125 million synapses from electron micrographs is brutally hard. Registration, segmentation, proofreading, circuit tracing โ that is years of compute and thousands of hours of human annotation. The skill stack is neuroscience, computer vision, and distributed systems.
Executing a trade through an agentic API is not that. It is an API call with a signature. Any developer with an afternoon and a Coinbase account can do it. The technical content of the crypto half of this experiment rounds to zero.
I have audited enough integration layers to know the shape of this. In 2017, during the ICO cycle, I ran through more than forty ERC-20 contracts in three weeks. Almost all of them presented as novel protocols. Almost all of them were standard templates with a marketing skin. The divergence between the complexity of the story and the complexity of the code is a reliable tell. It has not failed me yet.
Here the divergence is reversed and therefore more interesting: the biology is over-engineered for the task, and the finance is under-engineered to the point of irrelevance. Someone bolted a research-grade neural reconstruction onto a retail API and called the result a trading experiment. That is not a category error. It is a category refusal.
Now the derivative layer, which is where the actual verification problem lives.
Around the core experiment, a cluster of self-reported demos appeared. The fly brain parallel-parks a car. The fly brain solves a Rubik's cube. The fly brain plays Beat Saber. The fly brain plays Minecraft. The fly brain doomscrolls. And, memorably, someone claimed to have made the fly bisexual.
Some of these are attributed to named individuals โ Breg Grockman, Hempstead, Smith, Christie. The naming convention alone should raise an eyebrow; "Breg Grockman" reads as a joke wearing a surname. Most of these accounts are unverifiable. None ship reproducible code alongside the claim. None publish a control condition.
I do not accept demo videos as evidence. I have watched too many of them edited to hide the failure frames.
This is not pedantry. It is the same failure mode I documented in early 2021, when I audited the storage layers of fifteen major NFT collections and found that roughly 60% were serving metadata from a centralized server dressed up as something decentralized. One of those projects' servers went down and the artwork disappeared from every marketplace that had listed it. Holders still had the token. They no longer had the asset.
Garbage in, permanence out. The ledger recorded a pointer to a thing that no longer existed, and the record was irreversible while the thing was not.
The fly demos are the same structure in a lighter register. The claim is permanent in the timeline. The artifact is a video and a promise. The reproduction path is missing. A demo you cannot re-run is not a result. It is a screenshot.
Then there is the layer that turns all of this into a financial instrument, and this is the part I want to be clinical about.
Two tokens โ $FLYCOIN and $CARLA โ attached themselves to the narrative. They are promoted on billboards. Some crypto traders bought ad space to push them. That is the entire disclosed distribution mechanism.
No supply schedule. No allocation breakdown. No contract address in any of the coverage. No protocol revenue. No governance function. No utility. No team, identified or pseudonymous-with-a-track-record.
The tokens have exactly one asset: attention. And attention, in this market structure, is a borrowed asset with no collateral behind it.
Let's do the accounting honestly. What does a holder of $FLYCOIN own? A claim on the future flow of new buyers who heard the same fruit-fly story. That is it. There is no cash flow to discount. There is no fee stream to capture. There is no parameter to govern. The only mechanism that moves the price is the arrival of the next marginal participant, and that participant's only reason to arrive is the price having already moved.
That is not a business. That is a queue.
I have written about this pattern before, and I will keep writing about it because the mechanism does not change. Volatility is the product; loss is the feature. In a meme instrument with no fundamental anchor, the volatility is not a side effect of the trade โ it is the thing being sold. The buyer is not purchasing exposure to upside. The buyer is purchasing the right to be early in a sequence that has a mathematically guaranteed terminal state for the majority.
And note the structure of the promotion. Someone bought billboard space. Billboards are not cheap. The buyer of that space has inventory. Inventory implies a position. A position implies an exit. The advertisement is not a public service announcement about a promising new asset. It is a distribution channel, and the story about the fly is the lubricant.
Which brings me to the far more interesting question, the one the coverage skipped entirely: the timing.
If the connectome shipped in October 2024 and the tweets are stamped September 2026, then either the tweets are misdated, or the article is misdescribing the publication date, or the entire sequence is a reconstruction dressed as reportage. Any of those three is a problem. All three are a pattern.
I spent seventy-two hours in May 2022 tracing wallet clusters through the Terra collapse. Anchor deposits, treasury reserves, stake weights. The thing that let me publish before the mainstream press was not access. It was that I checked the timestamps against the on-chain record and found they did not line up with the narrative. The cascade did not start where everyone said it started. It started earlier, in a cluster everyone had labelled benign.
Timestamps are evidence. They are the cheapest form of verification available and the most consistently ignored. When a story's own dates contradict each other in the first paragraph, everything downstream inherits the doubt.
The verification failures compound. Claimed demos with no code. Tokens with no contract address. A date that cannot be reconciled. And a scientific achievement โ the real one, the twenty-year reconstruction โ used as a backdrop for a hundred-dollar joke.
One more layer: the regulatory footprint, which at present does not exist.
When an autonomous agent makes a trading decision, who owns the outcome? The developer who wrote the mapping function? The platform that exposed the API? The model that produced the signal? The user whose capital was at risk?
Under the current US framework, there is no precedent. Coinbase runs KYC on the account. The agent runs underneath that identity. If the agent does something the account holder did not intend, the account holder is still the counterparty of record. If the agent does something the platform did not intend, the platform is still the venue.
The gap is not theoretical. It is a blank field in every risk disclosure I have read on agentic trading products.
There is a second question, on the token side. Apply the Howey factors to $FLYCOIN and $CARLA. Money invested โ yes. Common enterprise โ arguable. Expectation of profit โ yes, explicitly, that is the entire pitch. Derivation of profit from the efforts of others โ yes, from whoever is running the billboard campaign and whoever is amplifying the fly story.
That is not a clean pass. It is a plausible hit, contingent on structure. And promoting an unregistered security through paid physical advertising, during a period of active enforcement, is not a subtle position to occupy.
I would not want to be the person who signed the billboard contract.
***
Contrarian: what the bulls actually got right
I have been hard on this. Let me be precise about what survives the teardown, because two things do, and dismissing them would be as lazy as the coverage I am criticising.
First: the connectome is not hype. It is one of the more consequential open datasets released in the last decade, and the fact that it is being consumed publicly through derivative nonsense does not diminish it. The labs did the work. The reconstruction is real. The disease-research implications โ circuit-level understanding of neurodegeneration, of sensory processing, of how a nervous system routes a signal โ are genuine and long-horizon. The fly cannot trade. The fly can teach. Those are different claims and only one of them is being made loudly.
Second, and more relevant to anyone reading this for market signal: Coinbase's agentic interface is a real product, and this experiment is a legitimate stress test of it. Not a good one, not a well-designed one, but a real one. An external, non-financial, biologically-derived process was wired into an execution API and it worked. Orders were placed. The venue cleared them. The rails did not break. That is a meaningfully different claim from "AI can predict markets," and it is the claim that actually got demonstrated.
Where the bulls are wrong is in the conflation. "An autonomous agent can execute" and "an autonomous agent can profit" are separated by the entire discipline of quantitative finance. The fly proved the first. The -1% is silent on the second, because a single day at a hundred dollars cannot speak to it.
Here is the sharpest version of the blind spot. If you wanted to test whether a biological decision system has trading edge, you would need a control arm โ a random-signal bot with identical capital, identical venue, identical window. You would need a hundred times the sample. You would need to report the drawdown, not the point estimate. Nobody did any of that, because nobody was trying to. The experiment was designed for a tweet, and it optimised for that objective perfectly.
I audited an AI-content platform in 2026 that advertised immutable provenance. The smart contracts wrote hashes to a chain. The off-chain API rewrote them. An admin key sat in the middle and the "immutable" log was mutable, and the way I proved it was by comparing the on-chain hash against the API response for the same record and finding they diverged.
That is the standard I want applied here. Compare the claim against the artifact. When they diverge, the claim is wrong, not the artifact.
***
Takeaway
Watch three signals, in order. First: does Wormuth or anyone else publish a re-runnable harness, or does the fly stay a clip? Second: does any of the derivative demo code appear under a licence anyone can inspect? Third: do $FLYCOIN and $CARLA ever publish a contract address that survives a basic audit?
If all three stay missing, then the real output of this experiment was never the trade. It was the advertisement. The fly lost a dollar. The people who read about it may lose considerably more, and they will do it while believing the story's own dates โ the ones that never happened.