Paper: “Flies are all you need” (on fly-language model)
Artificial Scientific ● Covered by 4 sources
A fly connectome plugged into a frozen language model nudged chat predictions a bit. But a simpler control did slightly better, so the bug here is mostly hype, not biology.
Based on reporting by Artificial Scientific — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Researchers wired the complete retained MaleCNS graph into a frozen 1.17-billion-parameter language model and trained only a 278,528-parameter readout on top. The setup used all 166,700 retained nodes and 25,582,938 directed edges, so this wasn’t a toy graph. It was a full, auditable connectome-conditioned conversation system, with the language backbone kept fixed the whole time.
On 32 newly held-out conversations, the fly readout did beat the frozen model. Mean negative log-likelihood fell from 1.381995 to 1.359816, a difference of 0.0222 nats per target token, and perplexity dipped from 3.98 to 3.90. That’s real movement. But it’s also a small one, and the paper is blunt about that: the gains come from a narrow corpus, not from any claim of general intelligence.
The part that changes the story is the control. A parameter-matched direct-input readout did slightly better than the fly version, reaching 1.359328 NLL. The fly-minus-direct gap was +0.000488 nats per token, which means the connectome did not beat the simpler alternative. The graph mattered, but it didn’t seem to add an edge of its own over a matched adapter that skipped the anatomy entirely.
The authors also ran two neat sanity checks. Removing the connections made the residual disappear exactly, and relabeling the nodes disrupted the fitted interface. That supports the claim that the graph is being used, not just admired from afar. Still, the paper’s own conclusion is the right one: this shows a connectome can be plugged into a language model in a transparent way, not that fly wiring improves language modeling.
There’s a lot of honest engineering here, including a public training code release and a clear warning that the archive and fitted readouts aren’t in the preprint. The result is less “flies are all you need” than “flies are a controlled feature transform, and the boring adapter almost wins anyway.”
My take — AI-written commentary, not fact-checked reporting
This is the right kind of anti-hype paper: it takes the cute headline, then quietly ruins it with controls. That’s healthy, because the field has seen enough biological garnish pasted onto pretrained models to last a decade. The real takeaway is not that a fly brain talks; it’s that transparency beats mystique, and the simplest baseline usually shows up to spoil the party.
Read more about this at: Artificial Scientific