A Startup Built a Neural Network From a Fruit Fly Brain — Then Scrambled It

A biotech-adjacent speech AI startup just ran one of the more legible tests of a question machine learning researchers argue about constantly: does the specific shape of a neural network's wiring matter, or does a network mostly work because of scale and training, regardless of the architecture underneath it?

Oruk Labs, a speech AI company founded by Stanford PhD student Nathan Roll and backed by a16z speedrun, built a neural network whose internal wiring is a real, 499-neuron circuit lifted directly from the fruit fly brain map that Google Research and HHMI Janelia published on September 3, 2026. They trained it to recognize emotion in human speech. Then, in the same experiment, they scrambled that biological wiring at random and retrained only the output layer — and got a statistically indistinguishable score. That second result, not the headline claim about a fly "understanding" emotion, is the part worth taking seriously.

TL;DR

What Oruk Labs actually did

Oruk Labs' core business has nothing to do with fruit flies. It runs a speech API that transcribes audio and layers on emotion and speaking-style labels for products like voice agents and contact-center analytics — claiming 77.8% accuracy on speech emotion recognition benchmarks, a 9.1-point lead over the closest competitor it names (emotion2vec+). The connectome project is a research side-quest published under the company's research blog, in a September 6, 2026 post titled "We taught a fruit fly to hear human emotion."

The circuit itself is not "the fly's brain" in any general sense. It's a specific, small subset: 499 neurons forming the largest strongly connected component among the 512 strongest eligible central-brain intrinsic neurons in MaleCNS v1.0, the male connectome dataset released days earlier. Those 499 neurons carry 15,865 distinct connections, backed by 867,344 individual synaptic contacts in the raw microscopy data.

To turn a wiring diagram into something a neural network can compute with, Oruk Labs converted each connection into a numeric weight: sign × log(1 + contact count), where the sign follows the neuron's real neurotransmitter — acetylcholine connections get a positive sign, GABA and glutamate connections get a negative one, mirroring their excitatory or inhibitory biological role. Those weights then get scaled so the whole matrix has a spectral radius of 0.9, a standard stability constraint in this class of model.

The architecture is reservoir computing — with real biology as the reservoir

Stripped of the fly framing, what Oruk Labs built is a well-known architecture: a reservoir computing network, also called an echo-state network. The recurrence rule they publish is straightforward:

x[t] = (1 − α) x[t−1] + α tanh(W x[t−1] + Win u[t])

with α = 0.15. W is the fixed connectome weight matrix; it is never updated during training. Audio input (resampled to 8 kHz, converted into 32 log-mel frequency bands) is projected into the circuit through Win, a fixed random projection, and the circuit's resulting internal activity x[t] is the only thing a small trained linear readout layer ever sees. That readout layer — the only part of the whole system that learns — is trained to predict 31 labels: 15 emotion categories plus 16 speaking-style categories, evaluated with mean average precision (mAP) since a clip can carry multiple positive labels at once.

This is the useful technical detail the viral framing skipped past. Reservoir computing has existed since the early 2000s specifically because a large, fixed, richly-connected recurrent structure can act as a useful nonlinear feature extractor even when nothing inside it is trained — only the readout needs gradient descent. Using a real connectome as that fixed reservoir, instead of a randomly initialized one, is the actual novelty here: it swaps an arbitrary random matrix for one with real biological structure and see what changes.

What the numbers actually show

Trained and evaluated on 16,995 training clips, 2,239 validation clips, and 2,022 test clips drawn from the CREMA-D and VCTK speech corpora — with an emotion label counted as a positive ground-truth tag only when at least half of human listener ratings agreed on it — the fly-wired network scored 16.84% mAP. A constant baseline (predicting the most common label distribution regardless of input) scored 9.71% mAP.

That's a real signal: the circuit is picking up genuine structure in the acoustic input relative to a naive baseline, roughly a 73% relative improvement. It is also a modest number in absolute terms, and Oruk Labs has not published a controlled comparison against human listener accuracy on the same test set — so the framing in the original X thread ("better than some humans I know") is a founder's informal aside, not a benchmarked claim in the company's own technical writeup. Readers should treat the two very differently.

The real finding: scrambling the wiring didn't hurt

The headline claim gets attention. The control experiment is the actual finding. Oruk Labs took the same 499 neurons and the same connection weights, then randomly shuffled which neuron connected to which — preserving each neuron's individual properties (its excitatory/inhibitory type, its overall connection strength) while destroying the specific, evolved topology of who talks to whom. They retrained a fresh linear readout layer on top of that scrambled circuit, using the identical training procedure.

The scrambled network scored 16.88% mAP — statistically indistinguishable from the real circuit's 16.84%. Oruk Labs reports the paired difference as −0.04 percentage points, with a 95% confidence interval spanning −0.16 to +0.07 points. In plain terms: the interval straddles zero, meaning the data cannot distinguish "the real wiring helped" from "the real wiring hurt" from "no difference at all." The company's own conclusion is blunt: "the two were effectively tied," and it found "no evidence that the fly's particular wiring is an advantage" on this specific speech-emotion task.

Why this connects to a live debate in ML architecture research

This single test, on one 499-neuron circuit, on one emotion-classification task, cannot settle anything on its own — and it would be overclaiming to treat it as though it does. But it lands squarely inside a real, ongoing argument in machine learning: how much does a network's specific architecture matter, versus how much of a network's performance comes from having enough parameters, connectivity, and training signal, more or less regardless of the particular structure connecting them?

That question echoes the lottery ticket hypothesis — the finding that large networks contain smaller subnetworks that, trained in isolation, match the full network's performance — and the broader scaling-law observation that architecture-agnostic factors (parameter count, data volume, compute) often predict downstream performance better than any single architectural choice. A biologically evolved circuit, shaped by millions of years of selection pressure for the fly's own sensory and motor needs, is about as strong a prior as an architecture can carry into an unrelated task. Finding no advantage from that prior — on this test — is a small but genuinely informative data point in favor of the "structure matters less than people assume, once you're training a readout on top of it" side of that argument.

Honest limitations

• This is one team's blog post, not a peer-reviewed result. Oruk Labs has not, as of this post's publication, released an academic paper or submitted the work to a journal or conference. Treat the 16.84%, 16.88%, and confidence-interval figures as a vendor's own internal evaluation, disclosed transparently but not independently replicated.
• One circuit, one task, one dataset. The scrambled-wiring result held for a 499-neuron subset of one fly's central brain, on emotion/style classification over CREMA-D and VCTK. It says nothing about whether connectome-derived wiring would or wouldn't help on a different task, a larger circuit, or a different biological source.
• "Reads human emotion" is a headline, not a mechanism. Oruk Labs' own writeup is careful here: "this experiment says nothing about whether a living fly understands human feelings," and separately notes that "a wiring diagram does not supply every receptor, time constant or physiological detail needed to simulate the original cells." The model uses the fly's synaptic topology as a fixed mathematical structure — it is not a simulation of a living fly's nervous system, and it makes no claim about animal cognition.
• The viral framing outran the technical writeup. "Better than some humans I know" is a founder's aside on X, not a number in the company's own evaluation. No human-listener baseline on the same test set has been published for direct comparison.
• 16.84% mAP is a modest absolute score. It beats the baseline by a meaningful margin, but this is not a state-of-the-art emotion-recognition result — Oruk Labs' own flagship product reports 77.8% accuracy on its standard (non-fly) speech-emotion benchmark, for scale.

Update — the same connectome is now playing video games

Oruk Labs isn't the only project running compute over MaleCNS v1.0. Two separate, independent hobbyist experiments surfaced the same week, both using the fly connectome as a fixed neural substrate for reinforcement learning rather than supervised classification:

A fly brain learning Beat Saber. X user Lyra (@lyraaaa) posted a demo of a MaleCNS v1.0-derived network reacting to a Beat Saber gameplay feed, with the model currently in an active RL training loop. As of the post, only the motor-cortex portion is trained (overfit to a replay, in the poster's own words) — visual-cortex reactivity and genuine play-pattern learning are still in progress. This is explicitly a different circuit and a different training method than Oruk Labs' work: Oruk Labs freezes the connectome as a fixed reservoir and trains only a linear output layer (see above); Lyra's project is running RL directly against a game environment, which is a substantively different (and more ambitious) claim about what the network is learning to do.

A fly brain playing Doom. A related, earlier project from Alex Wormuth (@nftechie_, posted September 6) uses the full MaleCNS v1.0 connectome as a game-playing agent: each Doom frame stimulates sensory neurons, neural activity maps to game controls, and in-game damage triggers a reinforcement signal via stimulation of two PPL101 dopamine cells — a design explicitly modeled on the fly's known reward circuitry, not an arbitrary loss function bolted on top.

Why this matters for the broader "does connectome wiring matter" question this post already covers: these are a different experimental setup from Oruk Labs' scrambled-wiring test, and neither has published a scrambled-control comparison as of this update. But they're doing something Oruk Labs' study couldn't: training the connectome-derived network with RL against a live task, rather than a frozen reservoir with a linear readout. If a later scrambled-vs-real comparison on the Doom or Beat Saber task shows a real gap — unlike Oruk Labs' speech-emotion result — that would be the more interesting counterpoint to the "architecture didn't matter here" finding above. As of publication, no such comparison exists for either project; both are described as ongoing, unfinished training runs, not completed results, and neither project has a technical writeup or paper to verify beyond the X threads themselves.

What people are asking

"Is this the same connectome Google and Janelia published?" Yes — the 499-neuron circuit is a subset of MaleCNS v1.0, the male fruit fly connectome Google Research and HHMI Janelia released on September 3, 2026, just days before Oruk Labs' post. Oruk Labs is a downstream application built on that public dataset, not a new connectome-mapping effort.

"Could I try the demo myself?" Oruk Labs' X posts describe an interactive version at oruk.ai that lets a visitor rotate the 3D circuit, listen to labeled example clips, and toggle individual neurons off to see how the output changes — a hands-on way to explore the same reservoir-computing structure described in the technical writeup, alongside the company's broader speech emotion demo.

"Does this mean architecture doesn't matter in neural networks?" No — that's a much bigger claim than this one experiment supports. It's one data point suggesting that, for this particular task and this particular circuit, a trained readout layer could extract about as much useful signal from scrambled connections as from the real biological ones. Other architecture comparisons — attention versus recurrence at scale, convolutional inductive biases in vision, and so on — have shown real, reproducible differences elsewhere in ML. This result adds nuance to that broader conversation; it doesn't close it.

"Is Oruk Labs a neuroscience company?" No. Its core product is a commercial speech API for transcription, emotion, and speaking-style detection, serving customers like voice-agent and contact-center products. The fruit fly work is a research post on the side, in the same spirit as the company's other public research pieces on speech representation and accent.

Related on explainx.ai

• Google and Janelia mapped a fruit fly brain — with AI doing the reconstruction
• Can AI talk to animals? Crow vocalizations, a $10M prize, and the real science
• Aleph's silent-speech ultrasound model reads whispers without a microphone
• NVIDIA BioNeMo Agent Toolkit comes to Claude Science
• Claude designed working protein binders for 14 of 15 targets
• Terminal-Bench-Science: the benchmark that deflates coding-agent hype

Official sources: Oruk Labs' research writeup, "We taught a fruit fly to hear human emotion", Oruk Labs' company page, Google Research's connectomics milestone post, Lyra (@lyraaaa) on X, September 9, 2026, and Alex Wormuth (@nftechie_) on X, September 6, 2026.

Circuit specifications, accuracy figures, and confidence intervals in this article reflect Oruk Labs' own published research post as of September 9, 2026. This work has not been peer-reviewed or published in an academic venue; treat the numbers as a vendor-disclosed internal evaluation rather than an independently verified result.