ChatFLM

Language through a fly’s wiring. A small model learning the patterns of human conversation.

Mode

Trained on meeting transcripts. It continues dialogue; it has not been trained to follow instructions or give reliable answers.

Generation settings

Temperature Top-k bytes Maximum bytes Random seed Slow playback for observation

Inside the network

StateAnatomical context

Loading anatomy…

Negative Positive stateDrag to rotate · scroll to zoom

1,024 modeled neurons from a 166,700-neuron source graph. Gray points show anatomical context, without simulated activity.

Inspect neuron

Select a colored point or use the list.

Recurrent connectionsLocal learning

Changing a control replays the same input from zero state. Silencing is an engineered intervention, not a biological experiment.

Next byte

Waiting for input.

Mean absolute state

0.000

Fly body

Map state to pose

Loading articulated body…

A female NeuroMechFly body accompanies the male brain graph. Pose changes illustrate state; language training has not taught this body to move.

FLM research notebook

Can measured neural wiring provide a useful prior for learning language from very little human speech?

FLM turns that question into an inspectable experiment. It learns next-byte prediction from scratch on human meeting transcripts, with recurrent connections selected from a fruit-fly connectome. ChatFLM runs the resulting compact checkpoint in your browser.

The architecture

FLM compact core. Recurrence carries information forward without a growing attention cache.

The connectome determines which directed recurrent edges exist and their assumed fast neurotransmitter signs. Training adjusts their magnitudes, an input projection, bounded time constants and a readout. A slow state filters each neuron’s activity, giving the model two memory timescales. This is a differentiable rate network, not a biophysical reconstruction.

FLM and a conventional decoder transformer
Property	FLM compact	Decoder transformer
Communication	Fixed directed graph; trained edge strengths	Content-dependent attention between token representations
Memory	Fixed-size fast and slow recurrent state	Past token keys and values, within a context window
Position in sequence	State evolves in causal order	Position encoding and causal attention
Learning	Next-byte cross-entropy and backpropagation through time	Next-token cross-entropy and backpropagation
Browser adaptation	Separate online readout adapter	Architecture-dependent; not a defining transformer feature
Tradeoff	Small state; sequential training and lossy memory	Parallel training; more expensive long-context attention/cache

Human speech, without a teacher model

The primary corpus is the AMI Meeting Corpus, under CC BY 4.0. FLM uses the manual transcript text, not audio. Participants and related meeting sessions are grouped before partitioning: 139 training documents, 16 validation documents and 16 test documents. The training split contains 947,966 whitespace-delimited words, including turn labels.

All transcript normalization, split assignments, checksums and limitations are recorded. These are custom participant-disjoint language-model splits, not the official AMI speech-recognition benchmark. LibriSpeech acquisition remains an optional, incomplete follow-up and contributes no training data to this release.

What the evidence can establish

Training and matched comparisons are in progress. The browser checkpoint is a working intermediate release; no superiority over a transformer or a randomized graph has been established.

A lower held-out loss measures next-byte prediction in this domain. It does not establish reasoning, factual knowledge, consciousness, animal behavior or general conversational ability. The interesting claim to test is whether wiring and multiple timescales improve data efficiency under equal training budgets.

What changes in the “fly”?

Offline training changes the artificial network’s input responses, edge magnitudes, time constants and prediction readout. Those changes can be measured using the same input before and after training. The anatomy provides a location for visualizing the state; it does not make learned text an animal’s experience. Browser learning changes only a separate readout adapter. The body animation is an explicitly designed mapping from state to pose.

Research foundations

MaleCNS, Janelia — source connectome and anatomical identities.
Google Research: mapping the complete male fruit-fly brain — reconstruction context.
Shiu et al., 2024 — a whole-brain computational model of fly behavior.
Lappalainen et al., 2024 — connectome-constrained models of neural computation.
Morra et al., 2023 — connectome-inspired reservoir computing.
Bellec et al., 2020 — recurrent learning with eligibility traces; a future comparison, not FLM’s current optimizer.
SpikeGPT, RWKV and Mamba — prior non-attention and recurrent language-model directions.
NeuroMechFly v2 — the embodied simulation platform underlying the separate body geometry.

Reproduce and inspect

Loading checkpoint details…

Repository Checkpoint manifest Dataset card Graph card Attribution

FLM is an independent research prototype by Kuber. Original code: MIT. Brain data and AMI transcripts: CC BY 4.0. Body assets retain their upstream licenses.