Eight reviewers, each given the concept brief and the synthetic mockup. Reviews ran on Sonnet-class agents; every URL was verified live by the reviewer. Verbatim below; the synthesis is in the proposal.
I've spent fifteen years watching real flies walk, groom, and get startled under a scope, so let me be blunt about what this mockup gets wrong before anything else.
(1) Biologically wrong or misleading
(2) Five experiments worth building (each tied to a real, verified result)
(3) Datasets/models I'd actually use — and what I wouldn't trust
Use: flybody (Vaxenburg et al., Nature 2025, TuragaLab/flybody), FlyWire (Dorkenwald et al., Nature 2024, codex.flywire.ai), flyvis for the visual system (Lappalainen et al., Nature 2024, TuragaLab/flyvis), and MANC for anything touching leg motor control. Do NOT trust the whole-brain LIF model as a general oracle — it was validated against specific circuits (sugar, grooming) at ~90% match, not against arbitrary "silence any region" queries the UI implies you can make. Don't let users silence an anatomically fuzzy "region" and get a made-up behavior back; only wire up perturbations that map to an actual published experiment.
(4) Honesty rules
Label every element by provenance, not just a global "synthetic preview" tag: procedural body vs. real flybody trajectory; randomly seeded point cloud vs. real FlyWire skeletons; invented behavior vs. a cited, reproduced result. Never state "X causes Y" unless Y is the literal finding for that literal perturbation in a cited paper — no extrapolating.
(5) Ranking the 8 features
Top 3 changes:
I read this as someone who has spent the last three years staring at Shiu-model spike rasters and flyvis activation maps: the idea is sound, the raw material is genuinely first-rate public science, and the current mockup is honest about being fake — which is the right instinct. My notes below are about keeping that honesty as real components get swapped in.
1. Honest vs. decorative. Right now nothing in the "brain" is a model — it's animation with correct vocabulary. The 2,750-point cloud sits in five hand-placed ellipsoids; the sensory→central→motor "wave" fires on a fixed 1.2 s timer regardless of anything happening in the body; and the behavioral payoff (silence L motor → circles, silence visual → wanders) is scripted turn math, not output of any circuit. That's fine as a stand-in, and the "synthetic preview" tag and "BRAIN/BODY · stand-in" labels are doing real work — keep them. The risk is what happens at the next stage: if you swap in real FlyWire skeletons for geometry but keep the scripted silence-behavior, you get real neuron shapes wired to fake logic, which reads as far more authoritative than it is. Geometry and dynamics need independent provenance, not a single "real" flag.
2. What's computable client-side. The body side is cheap: MuJoCo-WASM runs contact-rich physics plus a small trained policy net in a worker in real time — this is demonstrated already for full-body models in the browser (google-deepmind/mujoco wasm build). The flybody model itself (Vaxenburg et al., Nature 2025; code at TuragaLab/flybody) is a MuJoCo model plus a compact RL controller — tens of MB of mesh/weights, sub-millisecond steps. That's a real client-side simulation, not theater.
The brain is the opposite story. The full Shiu et al. model runs on 139,255 neurons and 54.5 million synapses (connectome; model). Just the sparse synapse table is ~650 MB before you add per-neuron state; that's not going to run smoothly in a phone tab, and even native Apple-Silicon reimplementations report ~0.29 s of wall-clock per biological second. The connectome-constrained visual model (Lappalainen et al. 2024, code TuragaLab/flyvis) is similarly ~45,000 neurons across 721 columns. The honest design: precompute/replay the full-brain conditions server-side, but let a filtered subcircuit (a stimulus's direct + one/two-hop partners, a few hundred to low thousands of neurons) run live client-side in a worker with typed arrays — that's genuinely tractable and gives you real, if partial, live inference to point at.
3. Three triggerable, testable predictions:
4. Credibility risks. (a) Conflating "real skeleton geometry" with "real dynamics" as pieces ship incrementally — fix with a provenance badge per element (measured / replayed / live-inferred / illustrative). (b) FlyWire is one specific animal (a single female fly, FAFB dataset) — state that the connectome is a wiring diagram of an individual, not the species, and that weights in any LIF model are fit/assumed, not directly measured synaptic strengths. (c) Never phrase output as "will" — "in the model, silencing X predicts Y" is defensible; "silencing X makes it circle" is not, especially while the circling logic is scripted. (d) Drop cosmetic pseudo-metrics like a "spikes/frame" HUD number once it isn't tied to a real simulation — it currently looks like data and is a random draw.
5. Ranking the 8 features (scientific payoff × feasibility): #3 silence/stimulate against real (even precomputed) model output — this is the actual point; #7 explainer layer with live citations — cheap, directly buys credibility; #2 real connectome subset with click-to-inspect — grounds the visuals in real data; #1 real flybody via MuJoCo-WASM — feasible, legitimizes the body; #5 guided experiments mode — great packaging for #3; #4 real-vs-sim side-by-side — valuable but needs licensed footage and risks implying the sim is live when it's replay; #6 share-a-state URL — nice UX, no scientific value; #8 bring-your-own CSV — lowest priority.
Top 3 changes:
First 10 seconds. My friend texted me "Google built an AI fruit fly, look at this 😂" so I'm expecting something goofy and fun — like a filter or a game. What loads is a black screen with a glowing dotted ball floating over a little brown bug with red eyes and one visible wing, plus a wall of tiny green computer-code numbers in the corner and a big box of buttons on the right. Honestly my first reaction is "did this load wrong?" It looks like something off a hacker movie, not a cute science thing. The bug itself is actually kind of charming — I'd tap the fly first, expecting it to do something, or maybe I'd hit that orange "poke (air puff)" button because that's the only label that sounds like plain English and fun.
Words I don't understand. "synthetic preview," "gait" (okay, I actually know this one — we say it in triage, like "unsteady gait"), "tripod" / "tetrapod" gait, "connectome," "connectome stand-in," "physics stand-in," "spikes/frame," "silenced [none]," "synapse edges," "motor L / motor R" (reads like a part number), "procedural," "FlyWire skeletons," "flybody model," and the nav buttons "panel," "research," and "fable critique" — I have no idea what any of those three lead to and wouldn't click them. I'd want plain words: instead of "gait: tripod," just say "how it walks." Instead of "silence L motor," say "turn off the left leg controls and see what happens."
What keeps me past 60 seconds. The fly moving and reacting is genuinely fun — when I hit poke and it jumped, I laughed a little. If I could tap the fly and have it react to ME (not a menu button, just touching it), I'd stay longer. I'd send it to my son if it did something surprising or silly on its own — like if silencing part of the "brain" made it walk in circles like it was drunk, and it SAID that in plain language ("it's walking in circles because we turned off its left-turn signals"). Right now I had to guess what changed; nothing tells me in words what just happened.
Creepy / pointless / nerds only. The floating dot-cloud "brain" above the fly is the part that reads as nerds-only — it's pretty, but I don't know what I'm looking at and there's no story attached to it. It also feels a little creepy in a clinical way, like watching brain activity monitors. The HUD numbers (spikes/frame, neuron counts) are pure noise to me.
The one thing I'd want to DO. Poke it, honestly — or better, let me pet it, feed it something, or scare it and watch it actually flinch and run away realistically. I want cause and effect I understand without reading anything.
What would make me trust this is real science, not a game. Put a real video of an actual fly walking right next to the cartoon one, doing the same move at the same time. If they match, that convinces me way more than any label saying "Nature 2025" or a link to a paper I'll never open.
Ranking the 8 features: 1. Guided experiments · 2. Real-vs-sim video · 3. Explainer tooltips in plain English · 4. Silence/stimulate with a real effect (if explained simply) · 5. Share a link · 6. Real connectome neurons · 7. Real physics body (can't tell the difference) · 8. Upload your own CSV (would never touch it).
Top 3 changes:
1. The attract loop. Right now the fly already walks a slow tripod loop and the brain pulses every 1.2s with nobody touching anything — that's more than most first prototypes get right, because on a gallery floor nine visitors out of ten never touch anything; they decide whether to approach based on what the exhibit does by itself in the first three seconds. But the current idle state reads as decoration, not invitation: nothing signals "this responds to you." I'd add a soft, recurring cue during idle only — the fly pauses, looks toward the cursor/last-touch point, antenna twitches, and a single ghost fingerprint fades in over the body with a one-word label ("touch"). That's the Exploratorium's oldest trick (Exploratorium exhibits).
2. First touch. Bury the causal gesture in a side-panel button ("poke (air puff)") and you've already lost the stranger-who-wandered-up. The rule: the first thing they touch must be the thing itself, not a control for the thing. Exact proposal — tap/click directly on the fly's body. Within 2 seconds: flinch, wing-buzz, a visible red flash rippling from a sensory region through central to motor in the brain above, and the fly bolts. That's the same instantly legible cause→effect that makes neal.fun's Asteroid Launcher work — click the map, the whole scene reacts, no manual needed.
3. Narrative spine. The one sentence a 10-year-old repeats: "I touched its brain and the fly went crazy." Poke = you touched the body and the brain reacted. Silence-a-region = you touched the brain and the body reacted. That symmetry is the whole exhibit; nothing on screen says "these two are the same animal, watch them talk to each other" — add one line of connective tissue in a first-visit toast.
4. Failure modes. (a) too many buttons up front — 11+ controls visible simultaneously is a "control panel," not an exhibit; strangers freeze. (b) No real reset — "restore all" only clears silencing, not camera, speed, or wings. (c) No "what just happened" layer — the HUD is instrumentation, not story. (d) Total silence — a fly that gets startled and makes no sound reads as broken, not restrained.
5. Physical-to-digital. Sound is non-negotiable: footstep patter scaled to gait speed, a wing-buzz that pitches up with flap frequency, a sharp "thwip" on poke, and actual silence as the payoff when you kill a motor region, so the missing footstep sound is the explanation. Haptics on phone: navigator.vibrate() on poke and on each silence toggle. Time controls: a slow-motion scrub (0.25×) is more valuable than a full timeline, since the wonder is watching the spike ripple travel. A 3-second "replay" ribbon that loops your last action with a one-line caption turns raw numbers into the "what just happened" narrator — think of how Codex lets you click a neuron and get its story.
6. Ranking: 1. #3 silence/stimulate · 2. #5 guided experiments (the docent-in-a-box) · 3. #7 explainer tooltip · 4. #2 real connectome subset (risks turning an attract-piece into a database if not paired with #5) · 5. #4 real-vs-sim · 6. #1 real flybody physics (matters for honesty, looks the same to a stranger) · 7. #6 share-a-state · 8. #8 CSV.
Top 3 changes:
1. Does this help or hurt? As a "look, I built a cute simulated bug" site, it dilutes. Byron's brand equity is RAI leadership + agent security at board level — a bouncing cartoon fly with a glowing point-cloud "brain" reads as a weekend hobby project, not a credential. But the underlying science (a complete connectome, a physics-grounded body, ablation experiments) is genuinely adjacent to interpretability — his most defensible technical wedge. The site helps only if it is framed as interpretability research, not as pet ownership. Framing, not features, decides which brand it belongs to.
2. Strongest framing: "A complete wiring diagram is not an understood mind." FlyWire gives every neuron and synapse in a fly brain — the connectome equivalent of open model weights. Yet nobody can predict behavior just by staring at the diagram; you have to intervene — silence a neuron group, watch what breaks — exactly what "silence left motor" already does in the mockup. That is literal ablation, the same epistemic move as mechanistic interpretability on transformers. This ties Fruit Fly Lab directly to two things Byron already owns: jlens.arnao.ai (J-Space interpretability) and oreo.arnao.ai (defense-rung ablation on LLM attacks). Positioned together, the three sites become one coherent portfolio: "how do you verify understanding of a mind — biological or artificial — by intervening on it, not just describing it." That's a talk, not a toy. The rejected alternatives fail on inspection: "a fully known mind" overclaims (the connectome omits neuromodulation and synaptic weights); "ethics of simulating minds" invites a sentience argument that reads as amateur philosophy.
3. Launch day. LinkedIn post: lead with the ablation framing, not "I built a fly simulator." One line on the connectome-as-open-weights analogy, one on-page gif of silence→behavior-change, link straight to the live tool, then link jlens and oreo as the other two legs of the same idea. No internal tool/process names in the copy. EU/EMEA RAI leaders: the interpretability-under-EU-AI-Act angle — "transparency" as an intervention practice, not a document. AWS peers: strictly the public science — no AWS material, per the standing wall. Researchers: the citation trail and honesty about synthetic-vs-real data is the credibility test.
4. Risks.
5. Done well vs. done cheap. Done cheap is what's on screen now: a generic ellipsoid "brain" unrelated to real fly neuroanatomy, a cartoonish fly body, and the public header exposes "panel" and "fable critique" — internal process names leaking onto a public front door. Done well: a brain shape derived from connectome geometry, zero internal-tool exposure, inline citations next to every claim, and a guided narrative mode that states the ablation/interpretability thesis explicitly.
6. Ranking by brand value: 1. (5) Guided experiments · 2. (7) Explainer w/ citation · 3. (3) Silence/stimulate real model · 4. (4) Real-vs-sim · 5. (2) Real connectome subset · 6. (1) Real flybody · 7. (6) Share-a-state · 8. (8) CSV — off-thesis, cut or defer.
Top 3 changes:
I tested this at 200% browser zoom, with VoiceOver, and on my Z Fold (outer + inner). It's a strong-looking synthetic mockup, but right now it's usable by exactly one kind of visitor: sighted, mouse/touch, full-motor, and comfortable with WebGL.
1) Concrete defects
role, no aria-label. The HUD has no aria-live region, so updates every frame silently.for="speed"; a screen reader announces "slider" with no name..btn.on changes border color; no aria-pressed, no text change.prefers-reduced-motion check anywhere.overflow:hidden on body traps content at 200% zoom.<nav>, <main>, <header>.2) Making 3D tinkering accessible at all
Ship, alongside the canvas: (a) a live text region that narrates state changes in plain language — "Left motor silenced. Fly is now circling right"; (b) a data-table view of the brain toggle-able as a real table; (c) full keyboard camera (arrow keys orbit, +/- zoom, Tab-reachable preset views); (d) every control's outcome describable independent of watching pixels move.
3) Fold layout
The 720px breakpoint conflates two very different devices. Outer screen (~344dp, tall narrow) needs the bottom sheet. Inner screen (~673–717dp, near-square) has room for a real side panel — stacking it into a 52vh bottom sheet wastes half the surface. Detect aspect ratio, not just width. Never place primary buttons or HUD text on the crease centerline.
4) Performance/battery
2,750 points plus 1,120 edges is fine; 1,000–5,000 real neurons is not free — cap by device tier. Drop the pixel-ratio cap to 1.5, add an FPS watchdog that degrades in order: synapse edges → point count → pixel ratio → antialiasing. Pause the render loop on document.hidden.
5) Ranking by inclusive value: 1. #7 Explainer · 2. #5 Guided story · 3. #3 Silence/stimulate with text-readable outcome · 4. #6 Share-a-state · 5. #2 Real connectome (only with a searchable list view) · 6. #4 Real-vs-sim (needs captions) · 7. #1 Real flybody · 8. #8 CSV.
Top 3 changes:
aria-live narration region so the scene is usable without a mouse or sight.prefers-reduced-motion around the idle animation and stop the render loop on document.hidden.Standards fit. For my AP Bio kids, this lands squarely in Unit 8, Topic 8.1 ("Responses to the Environment") of the AP Biology Course and Exam Description — taxis vs. kinesis, innate vs. learned behavior. The "silence visual → wanders" toggle is a kinesis demo; the air-puff poke is a startle response. For my non-AP sections I'd hang it on TEKS BIO.12.A. If any of you are NGSS states, HS-LS1-2 ("organism movement in response to neural stimuli") is the closer match — this isn't a homeostasis demo, it's a stimulus-response-behavior one.
20-minute lesson. Bell-ringer prediction (3 min): "If we silence the left motor region, which way will the fly turn, and why?" Run it live, 6 min, students confirm/refute against a printed prediction sheet, then try visual-silencing and the poke. Close with a 5-minute "sketch the circuit." What's missing to run this with 30 Chromebooks: (1) a worksheet/prediction mode — nothing to click before you see the answer, so kids just react; (2) a fixed/seeded starting state so all 30 screens show the same run; (3) a snapshot/export button — the footprint trail is the actual data, but there's no way to capture it for a Google Classroom submission.
Chromebook/performance worries. The wing material uses physical transmission rendering for two tiny wings — a heavy pass on integrated GPUs. The animation loop is uncapped and does a per-neuron JS pass plus DOM writes every frame with no pause — that spins fans on a 50-minute period. It pulls three.js from a CDN at load; if the district filter (GoGuardian) blocks it, it's a black screen with no error message. I'd want a self-hosted bundle and a visible "loading/failed" state.
Real engagement vs. clicking. The prediction-before-you-run structure is what separates learning from button mashing. The silencing effects are scripted rules, not emergent from the neuron cloud — fine as a model, but it needs to say so on-screen, or a sharp student will ask "wait, is the brain actually doing this?"
Safety/AI framing for parents. Easy yes: no login, no tracking, clearly labeled. Strip the "panel"/"fable critique" nav links from the classroom URL. One clear sentence up top: "no AI model runs here, this is a physics/animation demo standing in for real neuroscience data."
Ranking: 1. Guided experiments (#5) · 2. Explainer w/ citations (#7) · 3. Real-vs-sim (#4) · 4. Silence/stimulate on real model output (#3) · 5. Share a state (#6) — lets me assign homework · 6. Real connectome (#2) — too dense for 20 minutes · 7. Real flybody (#1) — invisible to a HS student · 8. CSV (#8).
Top 3 changes:
(1) First five seconds. The load state is two unrelated objects: a wireframe ellipsoid labeled "BRAIN · connectome stand-in" floating at 3x scale above a small brown fly labeled "BODY · physics stand-in." A visitor's first read is "two demos stitched together," which undercuts the whole thesis — brain in a body. Fix the opening shot first: camera low and behind the fly mid-stride, no orbit yet, push in over 2-3 seconds toward the head, then the head capsule turns translucent and the neuron cloud inside it lights up. One line fades in: "This fly's brain is simulated too. Tap it." The first interaction is a tap on the fly itself — not a "poke" button in a side panel. Buttons and the full HUD appear only after that first tap.
(2) Watch → tinker → understand. Three concentric commitments, not three modes. Watch = the opening shot, zero UI, ends on the tap prompt. Tinker = direct manipulation only: tap the fly to poke it, tap a brain region to silence it, drag to orbit — complete with no menu. Understand = a dismissible strip that appears only after 2-3 interactions, offering "guided experiments" as short chapters. A tinkerer can expand the strip into a full panel — that's where sliders and camera toggles belong, demoted to "advanced." A casual visitor should leave having understood one thing without ever opening a panel.
(3) Control panel. Cut: the camera toggle (one smooth combined framing), the raw "synapse edges" label, the HUD line duplicating panel state. Convert to direct manipulation: poke → tap the fly's body; silence a region → tap that region in the brain (glowing hit-targets, not a 4-button danger grid). Add: a legend that lives on the brain as floating labels near each lobe, a visible provenance note next to the HUD ("2,750 synthetic points stand in for ~139,000 real FlyWire neurons"), and a persistent tiny reset control. Bug: the camera buttons never get the active-state class the gait and wings buttons get.
(4) Visual language. The palette (cyan / green / teal on near-black) reads as generic "sci-fi dashboard," and three of the five region colors sit too close in hue — separate by hue and lightness, never rely on color alone. To avoid "AI-generated" tells: drop the gradient-clipped wordmark, the ◇ glyph, the backdrop-blur glass panels, and the amber pill badge — that exact combination is the default skin of every AI-scaffolded dashboard right now. Look instead at Bartosz Ciechanowski's "Mechanical Watch", which uses real material shading, soft studio lighting, and color-coding of parts, not chrome — the interface disappears and the object reads as physical. Render the fly with proper cuticle shading and ambient occlusion; render neurons as soft volumetric points with depth falloff, closer to how FlyWire's Codex presents neuron skeletons — it reads as scientific because it under-decorates. Motion should be biological (organic easing, asymmetric spike decay), not UI-easing.
(5) Mobile / Fold. The bottom sheet opens at 52vh by default and covers the fly entirely — the opposite of "watch first." Default the sheet to a collapsed peek state (drag handle + one line of state), expandable on drag, never open-by-default. The single 720px breakpoint treats a phone and a folded-out tablet identically; the Fold's outer screen needs the peek-sheet, the near-square inner screen is wide enough for body and brain side-by-side with a slim control rail — an aspect-ratio/container-query distinction. The process links (panel/research/fable critique) don't belong on the public surface — move them behind a small "about this build" affordance.
(6) Share-a-state and guided experiments. Share-a-state should feel like sending a snapshot of a living thing: a camera-shutter action captures the current state into the URL and shows a live thumbnail preview before copying. Guided experiments should read as short scientific stories: title, one-sentence hypothesis ("What happens if the fly can't see?"), a "run it" button that sets state and holds camera on the relevant region, then a plain-English result with a live citation. Pattern this on Nicky Case's "Parable of the Polygons", which sequences guided demonstrations before handing the reader open controls, so trust is built before freedom.
Top 3 changes: