Abstract:
"I smell vinegar and yeast. I taste sugar. Warm and bright. What should I do?"
This sentence comes from the discharge of a fruit fly's brain. The impulses emitted by 138,000 neurons are decoded into words one by one by a translator, spelled into sentences, and sent to ChatGPT without changing one word. Just yesterday, developer @chetaslua announced that he had connected an entire fruit fly brain simulation to ChatGPT. 15 million connections, run one pulse at a time, on a MacBook.
There is no neural chip and no GPU. The tools are only NumPy and a connection group. Simply put, the connectome is the wiring diagram of the brain, how neurons are connected.
But even if it is a simulation, the shock of this event is not compromised at all——
Biological neural signals are translated into natural language for the first time, directly and
AI
Conversation.
Before letting the fruit flies "speak," Chetaslua first did a more important thing: proving that the results produced by this simulated brain were reliable.
So, he designed a rationality test based on the paper.
Sugar feeding - MN9 neurons responsible for beak movements fire at 78 Hz. This shows that the fruit fly tasted the sweetness and stretched its mouth forward at a frequency of 78 times per second, preparing to eat.
Add bitterness - the frequency instantly drops to 3 Hz. He shrank back and stopped eating.
The direction is consistent with the conclusion of the paper published in "Nature" in 2024.
Likes and dislikes are two strings of frequencies in this brain. The high frequency is to extend the mouth to eat, and the low frequency is to retract it.
Some netizens asked: Is 78 Hz generated by real connectome data, or was it compiled by your model?
He replied with two words:
its real
(real).
This step is critical. The most common mistake that simulated brains make is talking to themselves.
But first take a path that has a conclusion in the paper and verify it. If you get it right, others will listen to what you say later.
How does discharge become a human saying
This is the most critical step in the entire experiment, and it is also the most magical step.
He trained a "linear readout" on the brain's own activity. Speaking human language is a translator - mapping different discharge patterns to different words.
Each time the fly's brain produces a specific pattern of neural activity, the translator spits out a corresponding word.
During training, he did something particularly particular:
Excluding sensory neurons.
Because the moment the smell hits the olfactory neurons, the signal is the cleanest, and it is not easy to read it. It's like taking a peek at the answers during an exam. Of course they are all correct, but they are meaningless.
What he wants is: how much of this signal is left after it has been transmitted to the brain for several stations, and whether it can be recognized downstream.
In other words, the information is still there between "smell it" and "get ready to take action".
Result: All 16 words are correct, the accuracy is 100%.
He then reduced the available signals to only row neurons and motor neurons. That is, the layer that sends commands from the brain to the body is very close to "hands-on", with an accuracy rate of 85%.
This means that when feelings such as smell, taste, and touch are transmitted from the perception layer to the "preparation for action" layer, they still carry information that can be read. The signal is not diluted in transmission.
Then, the fruit fly spelled these words into prompt by itself and sent them to ChatGPT without changing them word by word.
The sentence at the beginning is the first sentence in this series of links.
Pay attention to this order -
First comes the discharge, then the words, then the sentences. There is no polish in between.
This is not the first time that I have used a laptop to run the entire Drosophila brain.
In October 2024, Berkeley researchers used the FlyWire connectome to simulate 139,000 neurons, 50 million connections, and a laptop. The simulation results can predict the response of real fruit flies after being stimulated.
The new thing this time is:
For the first time, the discharge was translated into natural language and handed over to a large language model word-for-word.
Previously, simulated neural activity could only be compared with the response of real fruit flies.
This set of things can be implemented only if the connection group data is fully disclosed.
The basis of the data is the male fly central nervous system connectome - MaleCNS v1.0, which HHMI Janelia, the University of Cambridge and Google Research spent ten years developing.
Manual proofreading alone took 44 man-years. It was published in the "Cell" journal on September 3. It contains 166,700 neurons and 125 million synaptic connections. It is completely open source under the CC-BY license.
Register to get a token, pip install a few lines of code, and you can pull the upstream and downstream connections of any neuron locally.
Just two days after the MaleCNS paper was published, Georgia Tech graduate student Evan Smith borrowed GPT-6 Astra and plugged this set of connectomes into Minecraft, allowing a virtual fruit fly to fly in a glass chamber, with neural activity directly determining every turn.
Immediately afterwards, people used fruit fly brains to fight Doom, run Mario 64, and play Beat Saber.
Ten years of maps, two days to enter the game, and a few days to start talking to GPT.
Chetaslua also did a round of testing in reverse.
"There is a wall ahead" is sent to ChatGPT, ChatGPT returns "to the left", and then this command is sent back to the auditory neurons of the fruit fly.
It obeyed 37 times. Not bad once.
But no food was ever found.
The instructions were executed flawlessly, and the foraging results were in vain. This shows that there is still a gap that cannot be crossed between
being able to drive behavior and generating consciousness.
Obedience only requires the signal to come in and the action to go out. But finding food requires a different set of things - you have to remember which side has tried it and which side hasn't, you have to save and use the information from a few steps ago, and you have to form some kind of "strategy" in an unfamiliar environment.
This fruit fly has not demonstrated this ability.
No one knows whether it smells real vinegar or whether it has a subjective experience. The developers themselves only say that this is a simulation.
It’s our turn to answer
In the past, if you wanted to do something with whole-brain simulation, you needed a top laboratory, a six-figure grant, 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 developer ran the 138,000-neuron fruit fly brain at home and connected it to ChatGPT.
The next experiment that allows biological neural signals to communicate with AI may no longer have to wait for paper scheduling or funding approval. It might have happened on someone's desk, late at night, after a flash of inspiration.
The fruit fly's sentence "What should I do" is a template sentence spelled out by the translator.
But it is indeed our turn to answer the question seriously now.
Comments