I smell vinegar and yeast. I taste sugar. It’s warm and bright. What should I do?
This signal comes from the neural activity of a fruit fly's brain.
138,000 neurons fired in unison; their pulses were individually decoded by a translator into words, assembled into sentences without alteration, and sent to ChatGPT.
Yesterday, developer @chetaslua announced that he connected an entire simulated fruit fly brain to ChatGPT.
15 million connections, run one pulse at a time, on a MacBook.
No neural chips, no GPUs. The only tools are NumPy and a connectome. In simple terms, the connectome is the wiring diagram of this brain—the way neurons are connected to each other.
But even as a simulation, the impact of this event was no less staggering—
Biological neural signals have been translated into natural language for the first time, enabling direct dialogue with AI.
Before getting the fruit flies to “speak,” Chetaslua first did something even more important: proving that the results generated by the simulated brain were reliable.
He therefore designed a validity test based on the paper.
Feeding—MN9 neurons responsible for mouthpart movements fire at 78 Hz. This indicates that the fly has detected sweetness and is extending its mouthparts at a rate of 78 times per second, ready to eat.
Added bitterness—frequency instantly dropped to 3 Hz. Pulled back, stopped eating.
The direction aligns with the conclusions of the paper published in Nature in 2024.
Liking and disliking are simply two frequencies in this brain: high frequency means reaching out to eat, low frequency means pulling back.
A netizen asked: Is 78 Hz derived from actual connectome data, or is it fabricated by your model?
He replied with two words: it's real.
This step is crucial. The most common mistake in simulating the brain is talking to oneself.
But first, use a well-established pathway from a paper to verify—it must match, otherwise no one will listen to what comes next.
How do you put that in plain language?
This is the most critical and most cunning step of the entire experiment.
He trained a "linear readout" on the brain's own activity—in plain terms, a translator that maps different firing patterns to different words.
For every specific pattern of neural activity produced by the fruit fly's brain, the translator outputs a corresponding word.
During training, he did something particularly meticulous: he excluded sensory neurons.
Because the signal is cleanest at the moment the odor hits the olfactory neurons, reading it out isn’t impressive. It’s like cheating on a test by looking at the answers—you’ll get everything right, but it’s meaningless.
What he wants to know is: After this signal has passed through several stages in the brain, how much of it remains, and whether it can still be recognized downstream.
In other words, is the information still present between the moment you "smell it" and the moment you "get ready to act"?
Result: 16 words correct, 100% accuracy.
He reduced the available signals to only the sensory and motor neurons—the layer that sends commands from the brain to the body, very close to actual movement—with an accuracy rate of still 85%.
This means that sensations such as smell, taste, and touch retain interpretable information all the way from the perception layer to the "preparation for action" layer—the signals are not diluted during transmission.
Then, the fly itself assembles these words into a prompt and sends them to ChatGPT unchanged.
The first sentence, which is the first sentence produced by this chain of steps.
Note this order—first discharge, then word, then sentence. No one polished it in between.
Running a fruit fly full-brain simulation on a laptop isn't the first time.
In October 2024, researchers at Berkeley used the FlyWire connectome to simulate 139,000 neurons and 50 million connections on a single laptop, with results capable of predicting the fruit fly’s actual responses to stimuli.
The new aspect here is that the discharge was translated into natural language for the first time and passed verbatim to a large language model.
Previously, the simulated neural activity could only be compared to the responses of real fruit flies.
For this system to work, the connectome data must be fully open.
The foundation of the data is the MaleCNS v1.0 connectome of the fruit fly central nervous system, developed over ten years by HHMI Janelia, the University of Cambridge, and Google Research.
Manual curation alone took 44 person-years; published in Cell on September 3, it includes 166,700 neurons and 125 million synaptic connections, and is fully open-source under the CC-BY license.
Register to get a token, run a few lines of code with pip install, and you can pull the upstream and downstream connections of any neuron to your local machine.
Just two days after the MaleCNS paper was published, Georgia Tech graduate student Evan Smith used GPT-6 Astra to embed this connectome into Minecraft, enabling a virtual fruit fly to fly inside a glass chamber, with its neural activity directly determining every turn.
Shortly after, someone ran Doom, Super Mario 64, and Beat Saber on a fruit fly brain.
A decade of mapping, two days to enter the game, and a few more days to start conversing with GPT.
Chetaslua also conducted a reverse test.
"Wall ahead" entered into ChatGPT, ChatGPT replies "Left," then this command is sent back to the auditory neurons of a fruit fly.
It complied 37 times. Exactly once.
But I never found any food.
The instructions were executed flawlessly, yet no awareness was generated. This indicates a gap—still impassable—that separates the ability to drive behavior from the capacity to produce consciousness.
Compliance only requires a signal in and an action out. But finding food demands something entirely different—you need to remember which directions you’ve already tried and which you haven’t, hold onto information from several steps ago, and develop some kind of “strategy” in unfamiliar environments.
This set of abilities was not demonstrated by this fruit fly.
It’s unknown whether it has smelled real vinegar or had any subjective experience. The developers themselves only say it’s a simulation.
It's our turn to answer
In the past, to do anything with whole-brain emulation, you needed a top-tier laboratory, a six-figure budget, an interdisciplinary team, and a long wait.
Now, all you need is a laptop, a downloaded wiring diagram, and a weekend willing to tinker.
A fruit fly brain with 138,000 neurons was successfully run by a developer at home, even connected to ChatGPT.
The next experiment to enable communication between biological neural signals and AI may no longer require waiting for paper scheduling or funding approval. It could happen on someone’s desk, late at night, after a sudden flash of insight.
The fruit fly’s line, “What should I do?” is a template phrase generated by a translator.
But this very question is now truly our turn to answer seriously.
Reference materials:
https://x.com/chetaslua/status/2098387725972123825
https://male-cns.janelia.org/
This article is from the WeChat public account "New Intelligence Yuan," authored by New Intelligence Yuan.