What the modeling hack is ¶
As of the July 20 public review, the beliefs analysis contains 16 human submissions and 14 visible CM_01 point estimates. Thirteen are aligned closely enough to enter the standardized pool; one additional anonymized $20/kg written estimate remains visible but is not pooled because it used a different output basis and supplied no interval. Workshop medians span roughly $1 to $100/kg, while the independent round extends to $500/kg. That spread is consistent with substantive disagreement, but it is not by itself proof: the sample is selected, several intervals are missing, and respondents may interpret the conditioning and output basis differently.
The modeling hack is the follow-on, with two goals. First, independent model generation: small quant teams — and AIs — each independently build a model of a specific target outcome (here, the 2036 cost per kg), so we can compare independent models and see how much of the $1–$100 spread is driven by parameters versus model structure. Second, engaging technical experts (GFI, ACIB, Ivy Farm and others, alongside forecasters and informed skeptics) in that modeling — informing the parameters, the context, and the model structure, and building a sense of how to hand their expertise to modelers and AIs in future. You are not asked to become a modeler; you are asked to make your expertise modelable.
The shared job is to decompose why estimates disagree and make each crux explicit, parameterized, and testable — improving the actual cost models (the Quarto/Observable JavaScript Monte Carlo dashboard and the Squiggle cost model) rather than just re-polling the headline number. The Python file is a reference implementation, not what runs the live dashboard.
How the project pieces fit together
- Workshop summary: qualitative evidence and arguments.
- Public beliefs analysis: consent-checked named and anonymous distributions, subgroup views, and sensitivity to pooling choices. The July 20 review removed inferred affiliations and protected anonymous timing/background clues.
- Cost model: turns explicit process assumptions into a factory-gate biomass-cost distribution.
- Cost → demand bridge: shows what a properly harmonized value-chain cost could feed into, while flagging that the current bridge is only illustrative.
- This hack: reconcile the estimand and units first, then build independent structures from one shared parameter contract and compare their outputs.
Evidence updates since the May workshop
- GFI's 2026 industry report records both progress and headwinds: 20,000-L production was reported in 2025, while disclosed sector funding fell to $73.9m from $144m in 2024. Scale and finance should be separate variables, not one maturity score.
- The FDA's February 2026 inventory lists five completed cultured-cell food consultations, including three completed in 2025. Regulatory progress is real, but it does not validate cost or throughput.
- A 2025 porcine-cell study found selected plant hydrolysates reduced serum use by up to 70%, while Pasitka's chicken process used $0.63/L animal-component-free medium. These are useful feasibility points, not evidence that full substitution works across cell lines and large reactors.
- Pablo AMC's July response reframes a central disagreement: a frontier “reasonably achievable” plant is not the same estimand as the average qualifying commercial plant asked about in CM_01. The hack should report both, not blend them.
Deliverables we are aiming for
- An updated, version-controlled TEA model (Observable JavaScript + Squiggle, with the Python reference kept in sync) where each contested assumption is an explicit, documented parameter with a credible interval.
- An updated public crux map linking each modeling disagreement to belief responses and evidence, with symmetric questions for skeptics and optimists.
- At least one new worked sub-model for a currently under-modeled driver (e.g. nutrient conversion efficiency, mass productivity, or product inclusion rate).
- A record of where experts and modelers actually converged versus where the disagreement is irreducible given current evidence.
The cruxes to attack ¶
These are the disagreements the hack should target. Each is a place where a single modeling choice swings the projected cost by a large factor.
1. Cell-line / growth-factor independence
Will gene-edited or autocrine lines reach commercial-scale GF independence by ~2030–2033? If yes, the entire growth-factor cost line collapses (some submitters assign $0/kg GF by 2036). Counter-caveat: the relevant performance, stability, safety, and regulatory pathway still need validation at commercial scale.
2. Hydrolysates as base media
Can plant/yeast hydrolysates fully replace purified amino acids (FEASTS, Fuchs) — or only partially (Swartz ~20%, Bomkamp ~40%)? Note CM_12's wording conflated replacing growth factors vs replacing the amino-acid base; part of this crux is definitional.
3. The unit the model is built on
g/L vs cells/mL vs "performance-to-cost ratio" vs mass productivity (kg/m³/day). Several experts argue harvest density alone is "meaningless." The metric a model is built on is a crux — and it ties density to process mode (fed-batch / perfusion / continuous).
4. Fresh-media intensity
Litres of net fresh medium consumed per kg of standardized biomass. The ~20 L/kg industry estimate belongs here. This is a volume-intensity parameter that combines density, perfusion, recycling, and losses; it is not a feed-conversion ratio.
5. Nutrient conversion efficiency
Feed conversion must be defined on a consistent mass, protein/amino-acid, or energy basis. A 3:1 calorie ratio, kg nutrient/kg biomass, and L medium/kg biomass are dimensionally different quantities. The current materials conflate them; the hack must choose and balance a nutrient basis before assigning a prior.
6. Capital & CDMO-to-dedicated-plant
Can companies finance the path to the 2036 scenario post-2024? CAPEX is flagged as less certain than ingredient costs; timeline and company survival may bind harder than the technical cost floor.
7. What product are we costing toward?
The inclusion rate of CM biomass in the final product is unspecified. A low-inclusion hybrid and a whole cut are different products with different value chains. Biomass cost alone does not determine retail price parity.
Lower-tension but useful as calibration checks: food-grade vs pharma-grade media share (CM_17, near-consensus "low-hanging fruit"), and the GF cost-reduction pathway probabilities (E6), which diverge sharply and are a clean target for structured probabilistic elicitation.
Minimum viable hack ¶
The current proposal is intentionally broad. Before announcing a date, run a smaller pilot with one shared data contract and two independent model structures. The event should not begin with open-ended model building.
- Freeze the estimand and basis. Specify conditional-on-commercialization versus unconditional probability, frontier versus average plant, output gate, dry matter/protein basis, recovery yield, geography, and constant-dollar year.
- Publish a versioned baseline. One parameter sheet, one model commit, deterministic seed, machine-readable output, and regression tests. Record the clean-default distribution and component decomposition.
- Choose two tracks only. Recommended pilot: (A) media intensity + nutrient conversion + mass productivity; (B) factory-gate-to-retail value-chain bridge. Defer other cruxes to the backlog.
- Build independently. Two or three small teams receive the same parameter contract but do not see one another's model structure until outputs are frozen.
- Compare and adjudicate. Overlay outputs, trace differences to structure versus parameters, run unit/basis tests, and merge only sourced changes with a named maintainer.
Go/no-go gate: schedule the live event only after steps 1–2 are complete and at least two modeling leads plus one domain lead have agreed to produce reviewable artifacts. Otherwise run the work as an asynchronous model-review sprint; that is more likely to produce durable improvements than a loosely scoped hackathon.
Parameter jam: run a cost simulation in your browser ¶
A simplified, self-contained version of the cost model. Enter your low / mode / high estimate for each cost line (in $/kg of wet cell biomass at harvest), and the page samples ~5,000 Monte Carlo draws and sums them to a unit-cost distribution. This is a teaching toy — the real models have far more structure (process modes, maturity coupling, CAPEX from working volume) — but it shows how component uncertainty compounds. Try a preset, then drag the numbers.
Think of this as the shared baseline: it varies the numbers inside one fixed structure, so it measures parameter sensitivity only. The other — often larger — source of disagreement is the model structure itself (see parameters vs structure below).
Each line is sampled as a triangular(low, mode, high) distribution (clamped so low ≤ mode ≤ high and all ≥ 0); draws are summed per run. Triangular is used for transparency, not because it is the right shape — the production models use lognormals and betas with explicit p5/p95. Runs entirely in your browser; nothing is sent anywhere.
Where these default ranges come from
The defaults are component $/kg ranges drawn from the v0.3 Squiggle model and the Observable JavaScript dashboard, expressed directly as cost-per-kg lines so they sum without the intermediate $/L × L/kg build-up:
- Media (basal + hydrolysates/pharma-grade): beliefs field CM_14 override is roughly p10 ~$5, mode ~$25–40, p90 ~$80–120/kg biomass.
- Growth factors: beliefs field CM_13 override is roughly p10 ~$2, mode ~$10–30, p90 ~$60/kg — collapses toward $0 under the breakthrough/GF-independence regime.
- Other variable (utilities, consumables, supplemental proteins, waste): ~$0.3–3/kg.
- Annualized capital ($/kg): derived in the full model from working volume, WACC, CRF and Lang factor; here entered directly as a per-kg line.
- Fixed OPEX: ~$1–6/kg at a 20 kTA reference plant.
The CDMO-toll, bundled-media, downstream-processing, and process-mode structure of the real models are intentionally omitted here.
Parameters vs structure: the path-dependence test ¶
The jam above varies the numbers inside one fixed additive structure. It cannot show the bigger source of TEA disagreement: the structure — which cost components exist, how they compose, what is bundled or conditional. The published gap (Humbird ~$21/kg, CE Delft ~$6–8, Risner $44,500+) is mostly structural, not a quarrel over inputs.
The test for path dependence: hold the same reconciled parameter sheet fixed, have 2–3 people each build an independent model structure, then overlay their output distributions on one axis. The spread between independently-built models — at identical inputs — is the path-dependence signal: how much of the cost uncertainty is the modeler's structural choices rather than the data.
This is also the role split. Parameters / process experts own one basis-locked, common-currency parameter sheet (the contract); modeling people each build a structure that consumes it identically. Build those independent structures in the forkable tools below, then bring them back to compare.
Where this is heading. The aim is an interface where you describe a model however you like — prose, a sketch, a spreadsheet — and see it rendered back as a diagram, equations, and a plain-language summary, with AI-generated clarifying questions to pin down what you meant. The point is that every model people build lands in a mutually intelligible, comparable form that provably matches the author's intent — which is exactly what lets independent models be compared rather than collected as incompatible forks. AI-assisted Squiggle is one possible substrate; the requirement is the describe → represent → check-intent loop, not any particular language.
Take it further: the real models ¶
The browser toy above sums five hand-entered lines. The production models do much more — process-mode mixtures, a latent industry-maturity factor that couples adoption probabilities and financing costs, CAPEX derived from required working volume, and regime switches for cheap-vs-expensive growth factors. Start here:
Observable JavaScript Monte Carlo dashboard →
Quarto + Observable JS. ~30,000 runs; the primary live implementation, still unvalidated. Adjust parameters and watch the cost distribution move. (model limits & critique)
Squiggle cost model (v0.2/v0.3) →
Live on Squiggle Hub. Fork it, change a distribution, and share your scenario. Best for quick what-ifs and probabilistic reasoning.
Live embed of the Squiggle model (interactive — may take a moment to load):
If the embed above does not render in your browser, open the model directly: cultured_meat_improved_by_squiggle_improve on Squiggle Hub →
New to this? Learn the tools ¶
You do not need to already know how to build Monte Carlo models, write model code, or forecast well to take part. The hack is deliberately mixed: cultivated-meat domain experts who may never have built a model, EA/quant modelers who may not know the biology, and skeptics who want to pressure-test both. Upskilling is a first-class goal, not an afterthought. You will not be asked to learn any modeling syntax — you describe the model you have in mind and AI tooling turns it into a working, checkable model. We will run short, friendly onboarding sessions so domain experts can join the modeling and modelers can engage the biology. Come as you are.
Monte Carlo basics
What "sampling a distribution 5,000 times and summing" actually does, and why uncertain inputs compound. The parameter jam above is a live, hands-on example — start there.
Session: a 20-minute walk-through of the toy simulator and the production overview.
Describe a model — let AI build it
You don't need to learn Squiggle (or any) syntax. Describe the model structure you have in mind — in words, a sketch, or a spreadsheet — and AI tooling turns it into a working, runnable model you can check and adjust. The skill is being clear about structure — what the components are and how they combine — not coding.
Tools: AI-assisted Squiggle (Squiggle AI), or any general AI coding assistant for a Python/spreadsheet model. Session: describe one cost structure aloud and watch it become a model you can interrogate.
Techno-economic cost modeling
How a $/kg figure is built from media volume, cell density, reactor cost, CAPEX annualization, and fixed OPEX — and which assumptions move it most.
Start: the model methodology & parameter reference and the TEA comparison.
Calibrated forecasting & good credible intervals
How to give a 90% interval you would actually bet on — wide enough to be right ~90% of the time, narrow enough to be useful. Most people are overconfident; practice fixes it.
Start: CM_01 on Metaculus · practice with Quantified Intuitions calibration or the Calibrate Your Judgment app. Session: a quick calibration round before we elicit.
If you would like a particular tutorial, or want to run one, say so in the form below.
Propose a model change ¶
Spotted a parameter that is wrong, a crux we missed, or a sub-model worth adding? Tell us. This feeds directly into the hack agenda.