f hybrid product

AI-generated review, not an Unjournal evaluation. No person has reviewed its content or checked its numbers.

suggestion

The Unjournal team,

AI-drafted

Four areas

Conflict prevention and AI safety may reinforce each other, for example through cooperation on safety, or pull in different directions, for example if export controls slow frontier AI while sharpening incentives for conflict. Which outcomes and boundaries would make the group most useful?

Beliefs about AI's future military value still matter when they shape states' strategic incentives.

Institutional coverage raises visibility; it leaves the model’s policy interpretation largely unchecked. A comparative theory evaluation with Cooperating against Catastrophe could test whether racing results survive realistic extensions.

Spend most of your time on papers with the strongest potential impact: we’re likely to commission evaluations of only a few. Distinguishing candidates around the 80th–95th percentiles matters more than fine distinctions around the 50th–70th percentiles, and much more than sorting papers below the median. Use a quick first pass to find your leading candidates; there’s no need to examine the whole list in detail. BrowseCompare & rateOpen a title for its brief. Use Discussion for th

paper page.

How should evaluators use AI?

Background

run our 12-month stream evaluating research on

Recursive self-improvement and AI-driven growth

We're fixing how research gets evaluated — making rigorous research more impactful and impactful research more rigorous.

AI-assistance disclosure. This working assessment supports The Unjournal’s prioritization process. Its judgments and ratings have not been adopted as a commissioned human evaluation or a team decision. A human evaluator should verify the factual claims, methods, and ratings before signing or submitting an official evaluation.

The Unjournal ran this review with Codex (OpenAI GPT mo

No person has checked this content or its numbers.

Evaluator guide

AI-drafted guidance, not an Unjournal evaluation. The Unjournal has reviewed and revised parts of it, but has not checked all of its content or numbers. Claude is made by Anthropic, which publishes the paper (conflict note).

This paper sits close to the line between our academic and applied streams; we are treating it as academic, so use this form rather than the applied-stream version.

Would other re

(p. 46). T

71% from the task list and 27% from their own hours

particular omission

Which of these choices most limits what readers can learn?

We’d welcome corrections to the AI’s reading, as well as confirmation of a criticism and an explanation of how much it matters.

. How useful is that calculation once the cost and ownership of the extra capital are considered?

How much does that simplification miss about differences in pay, skills, and the cost of changing careers?

inputs

How much does the model add to what those inputs already imply?

paths

Who might rely on it: expectations about whether AI will soon accelerate growth.

Who might rely on it: choices between income support and help with moving into new work.

policy on capital ownership, investment in computing and energy, and the balance between taxing labor and capital.

Who might rely on it: analysts running fiscal or labor-market stress tests, journalists and policymakers quoting a scenario from the explorer, and people in AI-policy debates who want a shared way to state their disagreements.

The model and its online explorer turn assumptions about AI into economy-wide outcomes:

gains could affect who funds development and who has influence over it

Please evaluate the most recent version, and tell us if you find a newer one.

The optional claim section was written mainly for empirical papers. If you use it here, state the claim conditionally, for example “under the extreme scenario’s assumptions, cognitive unemployment reaches about 18 percent by 2030”, or choose a qualitative result. Skipping it is fine.

These are suggested directions. Choose checks that fit your expertise and commissioned scope, and say which important questions remain outside your assessment. The AI review’s issue count and severity labels need not determine your priorities.

Public annotations can point to a page, equation, table, source, or code location; the reuse notice below applies.

Where people replacing labor-force exits search

What would resolve it? A transparent mapping and a comparable alternative run could support, narrow, or overturn the AI criticism. Respondent-level survey conclusions remain unverified without the microdata; do not imply that a synthetic sample reproduces them.

The own-hours capability measure changes the nearest scenario, and the survey’s “months to a new job” does not map transparently onto a cross-group job-finding rate.

Finish with a judgment on the appropriateness of the research approach for its stated purpose, the paper’s strongest contribution, the few uncertainties that most affect its conclusions, and the limits on applying them.

2. Independently check the AI-flagged issues

The questions below are a menu, not a requirement to cover every topic.

Prioritize: whether the main conclusions survive defensible modeling alternatives; whether the empirical inputs and calibration support them; and whether the interpretation stays within the model’s population, horizon, and mechanisms.

Suggest substantial extensions as follow-up work unless they are already in your commissioned scope

Broader policy recommendations or views about AI development need a clear connection to the paper’s methods and evidence.

For readers concerned with AI safety and global priorities, the relevant research questions include the economic consequences of rapid capability growth, the distribution of gains and losses, and how much adjustment time different paths allow. Examine what this model can establish about those questions.

If you are a commissioned evaluator and wish to stay anonymous, please use this form rather than a public annotation.

so a one-time

is a single wage in the model, and it typically rises as a result of AI.

recent literature emphasizes that occupations are bundles of tasks, and workers differ in their productivity across them (Autor and Thompson, 2025; Freund and Mann, 2026; Althoff and Reichardt, 2026). We do not currently incorporate this richness

The factor structure is also deliberately simple: labor and a single capital good supplied along an exogenous schedule with fixed elasticity, which does not distinguish compute from other capital and does not explicitly model saving decisions. While

To complement the three scenarios, we also conducted a representative survey of the US population. We asked 10,980 US adults: (a) when AI will be able to do each of eight different cognitive tasks of increasing difficulty, (b) how widely AI will be used, (c) how large the produc- tivity gains will be, (d) whether workers will be automated or augmented, and (e) how long a displaced worker will take to find a new job. Running the simulations using their median an- swers as inputs to the model, we find that the public’s expectations land close to our substantial change scenario.

the innovation effects here are likely a lower bound.

the model cannot generate the singularities

Labor productivity rises through this channel

Automating research increases the amount of research inputs that go into the production of new knowledge, but research remains bottle- necked by physical tasks

which broadly correspond to the wide range of published predictions

n contrast, AI use implies higher demand for capital, which raises its return and share of total factor income.

We build a simple, integrated economic framework that converts a small set of parameters — chiefly, what fraction of tasks in the economy are affected by AI, how widely AI is used, how large the productivity gains per task are, and how much AI automates versus augments labor — into paths for productivity, growth, wages, the labor share, job reallocation, and unemployment from now until 2030.

The model maps future paths of AI capabilities into implied paths for GDP, the labor share, wages, labor reallocation, and unemployment.

Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory

frictions

"The model maps future paths of AI capabilities into implied paths for GDP, the labor share, wages, labor reallocation, and unemployment" (abstract); "The scenarios are not predictions, and we attach no probabilities to them" (p. 3).

What will be the total cost of growth factors12Common growth factors used in CM cell culture include: FGF-2 (Fibroblast Growth Factor 2) — most widely used; promotes proliferation IGF-1 (Insulin-like Growth Factor 1) — supports growth and survival EGF (Epidermal Growth Factor) — stimulates cell division TGF-β (Transforming Growth Factor beta) — regulates differentiation Insulin — metabolic support; relatively cheap Transferrin — iron transport; serum supplement Approaches to reduce GF costs include: precision fermentation (yeast/bacteria-produced), plant molecular farming, autocrine cell lines (self-producing), thermostable variants, and small-molecule substitutes. per kg of undifferentiated13Undifferentiated cell biomass: Proliferating cell mass prior to any directed differentiation step. This is what most published TEAs (Humbird, Pasitka et al., CE Delft) model. Cost estimates for differentiated products would typically be higher. cultured chicken cell biomass (wet weight, at harvest14Wet weight = the mass of cells as harvested from the bioreactor, including water. Typically ~75–90% water depending on cell line, density, and how much dewatering occurs during separation. We use 80% as our reference assumption. See the 'Cultured-chicken meat' definition above for full details.) produced in 2036?

plest Model → NoteModel version and September–October updates October 26 (target and basis, documentation only): the team set the model’s target: unconditional and hypothetical cost of a large plant built with the best technology available in 2036, whether or not such plants exist, in 2026 US dollars (estimand). It also set a planned standard output basis, washed cell biomass at a fixed solids content plus cost per kg of protein (output basis). The “downstream processing” switch is relabelled “forming and texturization” (outside the estimate). No model outputs changed. October 26 (engine 2026-10-26.2): results are also shown per kg at a target solids content and per kg of protein, using placeholder conversions you can change under Structural stress tests. A placeholder harvest, washing and dewatering cost is available under Model Structure, off by default. Default results are unchanged. October 26 (engine 2026-10-26.3): a third Fresh-media calculation option under Structural stress tests derives medium per kg from nutrient yield (grams of cell solids per litre of medium), so it no longer falls automatically as density rises. Off by default; default results are unchanged. See the structure check. October 29 (engine 2026-10-29.1): hybrid product for comparisons. Products will not be made entirely of cultivated cells, so the number shown beside each dry-cell-mass cost, and the probability thresholds, are now for a kg of hybrid product: 25% cultivated cells (at chicken-meat moisture) plus 75% plant-based ingredients at $3/kg. The default median is about $17/kg of hybrid product ($213/kg dry). Change the inclusion rate under Hybrid product in the sidebar. October 26 (engine 2026-10-26.4): default results change. The headline cost is now per kg of dry cell mass (about five times the old wet figures), thresholds then used a chicken-meat equivalent (all cells; replaced by the hybrid product on October 29), density inputs are g dry/L and $/kg inputs are per kg dry. Cost basis “wet” under Structural stress tests reproduces the earlier results exactly. See the basis note above. This page runs the current model (engine 2026-10-29.1). See the September 2026 review and correction log. For comparisons with beliefs recorded at the May 8 workshop, use the hosted workshop-era model or its tagged source. The archived model is preserved as it was deployed and includes errors corrected here. Two supplied research critiques prompted optional median-preserving priors, explicit growth-factor price ranges, and structural comparisons for media use, GF dosage, and maturity dependence. Baseline numerical assumptions are retained. New controls include provenance tooltips and links to the discussion responses, justification, and remaining questions. These additions are AI-implemented scenario tests, pending expert review. September 20: the model pages now connect the technical limitations to the workshop’s skeptical–optimistic disagreement map and structured response round. This adds no new parameter values or headline forecast. It makes the top-down microbial benchmark, interaction risk, and remaining measurement needs visible, with provenance and caveats in the response note. October 26: an opt-in Process-mode selection setting in Structural stress tests replaces the mode weights with feasibility probabilities and uses the cheapest feasible mode in each simulated world, optionally with decision noise. The default (draw mode by weights) and its results are unchanged. See least-cost mode selection for the method and caveats. NoteNew to this model? Start with the Simplest Model → — a shorter version focusing on some key levers with line-of-sight explanations. You can carry your settings over to this Advanced Model when you’re ready. ►Audio overviews(AI-generated · provenance)How Cultured Meat is Made ~11 minDownload MP3View/annotate scriptThe Cost Model Explained ~10 minDownload MP3View/annotate scriptCritique, revisions, and open questions ~32 minDownload MP3Transcript and provenance WarningModel Status & Limitations This model is largely AI-generated and, as of September 2026, has not been independently validated. It is provided to fix ideas, make assumptions inspectable, and support discussion and comparison. Do not treat the outputs as authoritative cost estimates. Several limitations are decision-relevant, not cosmetic: the single latent-maturity factor is ad hoc; source dollars are not yet normalized to one real-dollar year; source data on different weight bases are converted with a single assumed solids content (see the basis note below); and the parameter ranges have not yet been checked against the target quantity decided in October 2026, the cost of a large plant built with the best technology available in 2036, whether or not such plants exist (estimand). The sensitivity chart is a dollar-swing ranking, not a variance decomposition. We are addressing these incrementally. Read the full critique and our responses → Limits/Critique We especially welcome expert review of: growth factor quantities and prices, bioreactor CAPEX ranges, and the maturity correlation structure — see below and the Sources section. ImportantBasis change: costs are now per kg of DRY cell mass Since October 26, 2026 (engine 2026-10-26.4), the headline cost is per kg of dry cell mass, not per kg of wet biomass. Dry-mass figures are about five times the old wet figures (the default median is about $213/kg dry, against $43 per wet kg before), because a kg of harvested cells is only about 20% solids. Nothing about the underlying scenario changed. Why. Harvested cells are mostly water, and washing and dewatering change what a “wet kg” contains: a paste at 15% solids carries half the protein of one at 30%. Dry mass does not change with washing or dewatering, maps to nutrition, and is the natural basis for nutrient stoichiometry (how much medium it takes to build a gram of cells). What else you’ll see. Beside every dry-basis cost, the cost per kg of a hybrid product: 25% cultivated cells, counted at the water content of chicken meat (27.5% solids, a convention agreed on 28 October 2026), and 75% plant-based ingredients at $3/kg. Products will not be made entirely of cells, and there is no consensus inclusion rate, so 25% is a convention you can change in the sidebar. The probability thresholds ($5, $10, $25, $50) and chicken comparisons use the hybrid product. Cost per kg of protein is reported too. Cell density is entered in g dry cells per litre, and $/kg inputs are per kg dry. The old wet-basis output is one switch away (Structural stress tests → Cost basis) and reproduces the earlier numbers exactly. Shared links made before this change that set density or $/kg priors will now be read as dry-basis values. Why the basis matters, and how sources report it → AI-drafted (Claude, 26 October 2026) from David Reinstein’s design decision; not yet reviewed by a person. NoteFrom workshop evidence to model revision The May 8, 2026 workshop is complete. Its outputs now form a linked evidence-to-model workflow: the summary records the technical arguments, the beliefs analysis shows the forecast spread, the demand bridge connects factory-gate cost to an adoption model, and the proposed modeling hack turns the remaining disagreements into testable model changes. The follow-up now separates disagreements about definitions, factual estimates, evidence, source credibility, interpretation, model structure, and the importance of a claim to the final result. This matters because agreement that one input can become cheaper does not establish that the total system becomes cheap. The workshop synthesis explains the distinction, and the crux response form asks skeptics and optimists to diagnose it directly. Workshop summary → · Where disagreement sits → · Respond to the crux map → · Cost → demand bridge → · Sensitivity methods → The named beliefs-analysis dashboard remains an internal review artifact and is intentionally not linked from this public site; a redacted public synthesis should be produced separately. NoteWe Want Your Feedback Comment on specific text or a parameter with Hypothesis (click the < tab on the right edge). For longer comments, email contact@unjournal.org or open a GitHub issue. Listen: Technical Review (22 min MP3) — Audio walkthrough of model architecture and areas for review function toggleHypothesis() { const btn = document.getElementById('toggle-hypothesis') || document.getElementById('vc-hyp'); // Track state via body class (avoids checking Hypothesis's internal display/transform state) const hidden = document.body.classList.contains('hyp-force-hidden'); if (hidden) { document.body.classList.remove('hyp-force-hidden'); if (btn) btn.textContent = '◉ Annotations'; } else { document.body.classList.add('hyp-force-hidden'); if (btn) btn.textContent = '▷ Annotations'; } } function toggleFullWidth() { const body = document.body; const btn = document.getElementById('toggle-fullwidth') || document.getElementById('vc-wide'); // Quarto full-page layout uses column classes and sidebar const contentSelectors = [ 'main.content', 'main', '.page-columns', '#quarto-content', '.column-page', '.column-body', '.column-body-outset', '.panel-fill', '.panel-sidebar', '#quarto-sidebar', '.page-layout-full .content' ]; if (body.classList.contains('fullwidth-mode')) { body.classList.remove('fullwidth-mode'); // Remove injected style const injected = document.getElementById('fullwidth-style'); if (injected) injected.remove(); btn.textContent = 'Expand Content'; } else { body.classList.add('fullwidth-mode'); // Inject a style tag to override Quarto's layout constraints if (!document.getElementById('fullwidth-style')) { const style = document.createElement('style'); style.id = 'fullwidth-style'; style.textContent = ` body.fullwidth-mode #quarto-content, body.fullwidth-mode main.content, body.fullwidth-mode main, body.fullwidth-mode .page-columns, body.fullwidth-mode .column-body, body.fullwidth-mode .column-page { max-width: 100% !important; width: 100% !important; padding-left: 1rem !important; padding-right: 1rem !important; } body.fullwidth-mode #quarto-sidebar, body.fullwidth-mode #quarto-margin-sidebar, body.fullwidth-mode .sidebar { display: none !important; } body.fullwidth-mode .page-columns { grid-template-columns: 1fr !important; } `; document.head.appendChild(style); } btn.textContent = 'Normal Width'; // Also hide hypothesis when going fullwidth hideHypothesis(); } } function toggleParams() { const btn = document.getElementById('toggle-params'); const vcBtn = document.getElementById('vc-params'); const sidebar = document.querySelector('.panel-sidebar'); const fill = document.querySelector('.panel-fill'); if (!sidebar) return; const isHidden = window.getComputedStyle(sidebar).display === 'none'; if (isHidden) { sidebar.classList.remove('params-hidden'); sidebar.style.removeProperty('display'); if (fill) { fill.style.removeProperty('grid-column'); fill.style.removeProperty('width'); } if (btn) btn.textContent = '◀ Hide Parameters (expand charts)'; if (vcBtn) { vcBtn.textContent = '◀ Parameters'; vcBtn.style.background = '#e8f4e8'; vcBtn.style.borderColor = '#5a7a5a'; } } else { sidebar.classList.add('params-hidden'); if (fill) fill.style.gridColumn = '1 / -1'; if (btn) btn.textContent = '▶ Show Parameters'; if (vcBtn) { vcBtn.textContent = '▶ Parameters'; vcBtn.style.background = '#f8f9fa'; vcBtn.style.borderColor = '#ccc'; } } } // Wide sidebar: turn the parameters panel into a fixed-position drawer overlay. // Using position:fixed avoids all CSS grid conflicts — the sidebar lifts out of // the document flow and floats on top. A dimmed backdrop lets users click outside // to close. The OJS cells inside still work because the DOM nodes are the same. function toggleWideSidebar() { const body = document.body; const sidebar = document.querySelector('.panel-sidebar'); const backdrop = document.getElementById('params-backdrop'); const btnV = document.getElementById('vc-wide-params'); const btnI = document.getElementById('toggle-wide-params'); if (body.classList.contains('sidebar-wide')) { // Close drawer body.classList.remove('sidebar-wide'); const s = document.getElementById('sidebar-wide-style'); if (s) s.remove(); if (backdrop) backdrop.style.display = 'none'; if (btnV) { btnV.textContent = '⟺ Wide params'; btnV.style.background = '#f8f9fa'; btnV.style.borderColor = '#ccc'; } if (btnI) btnI.textContent = '⟺ Wide view'; } else { // Open drawer: make sidebar a fixed overlay panel body.classList.add('sidebar-wide'); if (!document.getElementById('sidebar-wide-style')) { const s = document.createElement('style'); s.id = 'sidebar-wide-style'; s.textContent = ` body.sidebar-wide .panel-sidebar { position: fixed !important; top: var(--quarto-navbar-height, 62px) !important; left: 0 !important; width: min(62vw, 820px) !important; max-width: 92vw !important; height: calc(100vh - var(--quarto-navbar-height, 62px)) !important; max-height: calc(100vh - var(--quarto-navbar-height, 62px)) !important; overflow-y: auto !important; overflow-x: hidden !important; z-index: 6000 !important; background: white !important; box-shadow: 8px 0 32px rgba(0,0,0,0.22) !important; padding: 0.5rem 2rem 5rem 1.5rem !important; border-right: 3px solid #3498db !important; scrollbar-width: thin !important; } `; document.head.appendChild(s); } if (backdrop) backdrop.style.display = 'block'; if (btnV) { btnV.textContent = '↘ Narrow params'; btnV.style.background = '#e8f0fa'; btnV.style.borderColor = '#3498db'; } if (btnI) btnI.textContent = '↘ Narrow view'; // Scroll drawer to top when opening if (sidebar) sidebar.scrollTop = 0; } } // Re-initialise Tippy tooltips for abbr[title] and span[title] elements added by OJS. // OJS renders after Quarto's DOMContentLoaded Tippy pass, so elements inside OJS // html`` templates are missed on first pass. Span elements (used for ⓘ inline notes // in the main content area) are also included — native browser title tooltips clip // long text and are unreliable across browsers. function initOjsTooltips() { if (typeof tippy === 'undefined') return; document.querySelectorAll( '.panel-sidebar abbr[title], .panel-fill abbr[title], .panel-sidebar span[title], .panel-fill span[title]' ).forEach(function(el) { const t = el.getAttribute('title'); if (!el._tippy && t) { el.removeAttribute('title'); // prevent duplicate native browser tooltip tippy(el, { content: t, placement: 'bottom', maxWidth: 500, allowHTML: false }); } }); } setTimeout(initOjsTooltips, 2500); // first pass after OJS settles // Re-run whenever sidebar OR main content changes (OJS re-renders on slider change) document.addEventListener('DOMContentLoaded', function() { var obs = new MutationObserver(function() { initOjsTooltips(); }); var sidebar = document.querySelector('.panel-sidebar'); var fill = document.querySelector('.panel-fill'); if (sidebar) obs.observe(sidebar, { childList: true, subtree: true }); if (fill) obs.observe(fill, { childList: true, subtree: true }); }); // Reset all parameters: clear URL state and reload (most reliable approach, // since OJS viewof inputs can't be reset reliably from plain JS) function resetAllDefaults() { window.location.href = window.location.pathname + window.location.hash; } var _expandAllActive = false; var _expandAllObserver = null; function collapseAllDetails() { _expandAllActive = false; if (_expandAllObserver) { _expandAllObserver.disconnect(); _expandAllObserver = null; } document.querySelectorAll('details[open]').forEach(function(d) { d.removeAttribute('open'); }); } function expandAllDetails() { _expandAllActive = true; document.querySelectorAll('details').forEach(function(d) { d.setAttribute('open', ''); }); // OJS re-renders reactive cells after every slider change, replacing <details> elements // with fresh closed ones. The MutationObserver re-applies 'open' to any new ones. if (!_expandAllObserver) { _expandAllObserver = new MutationObserver(function(mutations) { if (!_expandAllActive) return; mutations.forEach(function(m) { m.addedNodes.forEach(function(node) { if (node.nodeType !== 1) return; if (node.tagName === 'DETAILS') node.setAttribute('open', ''); if (node.querySelectorAll) node.querySelectorAll('details').forEach(function(d) { d.setAttribute('open', ''); }); }); }); }); _expandAllObserver.observe(document.body, { childList: true, subtree: true }); } } function toggleViewControls() { const content = document.getElementById('view-controls-content'); const icon = document.getElementById('vc-minimize'); if (!content) return; const isHidden = content.style.display === 'none'; content.style.display = isHidden ? 'flex' : 'none'; if (icon) icon.textContent = isHidden ? '−' : '≡'; } function toggleToc() { const btn = document.getElementById('toggle-toc') || document.getElementById('vc-toc'); // Quarto TOC can be in several locations const selectors = [ '#quarto-margin-sidebar', '#TOC', '.sidebar.toc-left', 'nav.toc', '#quarto-sidebar' ]; let toc = null; for (const sel of selectors) { toc = document.querySelector(sel); if (toc) break; } if (!toc) { btn.textContent = 'TOC not found'; return; } if (toc.style.display === 'none') { toc.style.display = ''; btn.textContent = 'Hide Table of Contents'; } else { toc.style.display = 'none'; btn.textContent = 'Show Table of Contents'; } } function hideHypothesis() { document.body.classList.add('hyp-force-hidden'); const btn = document.getElementById('toggle-hypothesis') || document.getElementById('vc-hyp'); if (btn) btn.textContent = '▷ Annotations'; } View ≡ ◀ Parameters ⟺ Wide params ▷ Contents ▷ Annotations ↔ Expand ▲ Collapse all ▼ Expand all /* ── Fullwidth mode: hide parameter sidebar, let charts fill the page ── */ .fullwidth-mode .page-columns { grid-template-columns: 1fr !important; max-width: 100% !important; width: 100% !important; } .fullwidth-mode #quarto-content, .fullwidth-mode main.content, .fullwidth-mode main, .fullwidth-mode .column-page, .fullwidth-mode .column-body, .fullwidth-mode .panel-fill { max-width: 100% !important; width: 100% !important; padding-left: 1rem !important; padding-right: 1rem !important; } /* Hide parameter sidebar and TOC in fullwidth mode */ .fullwidth-mode .panel-sidebar, .fullwidth-mode #quarto-sidebar, .fullwidth-mode #quarto-margin-sidebar, .fullwidth-mode .sidebar { display: none !important; } /* Hide Hypothesis in fullwidth mode, and when manually toggled off */ .fullwidth-mode .annotator-frame, .fullwidth-mode .hypothesis-sidebar, .fullwidth-mode hypothesis-sidebar, .hyp-force-hidden .annotator-frame, .hyp-force-hidden .hypothesis-sidebar, .hyp-force-hidden hypothesis-sidebar, .hyp-force-hidden iframe[src*="hypothes.is"] { display: none !important; visibility: hidden !important; pointer-events: none !important; } /* Sidebar overflow fix — offset by Quarto navbar so top buttons aren't hidden */ .panel-sidebar { overflow-y: auto !important; overflow-x: hidden !important; position: sticky !important; top: var(--quarto-navbar-height, 62px) !important; max-height: calc(100vh - var(--quarto-navbar-height, 62px)) !important; } .panel-sidebar details[open] { max-width: 100%; overflow-wrap: break-word; } /* When params sidebar is hidden, let charts take full width */ .params-hidden { display: none !important; } NoteNew to Cultured Meat? Read our deep dive: How Cultured Chicken is Made — a detailed guide covering cell banking, bioreactors, media composition, growth factors, and why each step affects costs. Quick summary: Cultured chicken is produced by growing avian muscle cells in bioreactors. The main cost drivers are media (amino acids, nutrients), growth factors (signaling proteins), bioreactors (capital equipment), and operating costs. For a side-by-side comparison of how published TEAs differ in their assumptions and estimates, see our TEA Comparison page. This model stops at a factory-gate cost; for what a given cost implies for consumer adoption and market share, see our Cost → Demand bridge, which connects to Pablo AMC’s demand-side model. CELL BANK → SEED TRAIN → PRODUCTION → HARVEST → PRODUCT [O] [OOO] [OOOOOOO] [===] [≡≡≡] NoteLearn: Understanding Uncertainty & Distributions This model uses Monte Carlo simulation: we sample thousands of possible futures and see what distribution of costs emerges. Distribution Types Used: Parameter Type Distribution Why Examples in model Costs, intensities Lognormal Always positive, right-skewed (some very high values possible) Media cost per liter, GF price per gram, cell density Probabilities, fractions Beta Bounded between 0 and 1, flexible shape Maturity factor, adoption probabilities, utilization rate (0–1 share). For switching parameters (e.g., hydrolysate adoption), the model first samples an adoption probability from a Beta distribution, then uses a Bernoulli draw to determine whether each run uses the "adopted" or "non-adopted" cost regime. Bounded ranges Uniform Equal probability within bounds; used when we have only a plausible range and no reason to favor values within it Asset life (years), CAPEX scale exponent, fixed OPEX scaling factor Background reading. Each distribution name above links to its Wikipedia article. For a visual, non-technical introduction to distributions, see Seeing Theory: Probability Distributions (Brown University). The switching parameters also use a Bernoulli (yes/no) draw. How the model builds a lognormal from the p5/p95 (or p10/p90) values you enter is explained in Model formulas: how the distribution is constructed. We are open to considering alternative distributional forms (e.g., triangular, PERT, or mixture distributions) and to making the modeling flexible enough for users to select or compare different distributional assumptions. Suggestions are welcome via Hypothesis annotations. Reading the Results: p5 (5th percentile): “Optimistic” - only 5% of simulations are cheaper p50 (median): Middle outcome - half above, half below p95 (95th percentile): “Pessimistic” - 95% of simulations cheaper The “Maturity” Correlation: A key feature is the latent maturity factor that links: Technology adoption rates Reactor costs Financing costs (WACC) In “good worlds” for cultured chicken (high maturity), multiple things improve together. This makes favorable assumptions co-occur more often. Technical success with difficult financing remains possible and should not be ruled out. This approach draws on the concept of copula-based correlation in Monte Carlo simulation, where a shared latent variable induces dependence among otherwise independent marginal distributions. See methodological references. More detail on the maturity factor implementation The maturity factor is sampled once per simulation from a Beta distribution (mean set by the “Industry Maturity” slider, standard deviation 0.20). It then shifts the adoption probabilities (hydrolysates, food-grade, recombinant GFs) and adjusts WACC downward in high-maturity draws. This creates positive correlation among “good news” variables without imposing a rigid structure. The strength of the correlation is moderate by design — maturity explains some but not all of the variation in each parameter. Caveat: This correlation structure is a modelling convenience, not an empirically calibrated design. The choice of a single latent variable, the specific sensitivity coefficients, and the functional form are all ad hoc. Results — especially in tail scenarios — may be sensitive to these choices. We encourage users to stress-test by varying the maturity slider widely and treat correlated-scenario outputs as illustrative rather than calibrated.

Cell Biomass Manufacturing Cost ($/kg dry cell mass) →

Weights not used: cheapest-feasible selection is on (Structural stress tests).

On the wet basis they apply to wet biomass, as before October 26, 2026.

Important: Model Status & Limitations

Basis change: costs are now per kg of DRY cell mass

Current basis. Costs are per kg of wet cell biomass, because the code treats the cell-density inputs as grams of wet biomass per litre at harvest. Harvested cells are mostly water, and a wet kg can carry more or less residual medium depending on washing and dewatering: a paste at 15% solids carries about half the protein of one at 30%. So “cost per kg of wet cell biomass” is not a stable unit.

wet-biomass water content is not standardized

the single latent-maturity factor is ad hoc;

Output basis: what “per kg” means here

For substantive or longer-form discussion — please post on 💬 GitHub Discussions. That’s where the conversation can get involved, where others can reply and build on each other, and where everything stays threaded and organized. See the 📖 Discussion Map for where to post what, or jump to a hub:

P(Fed-batch) (?)

Probability of Each Process Mode (?) Code viewof p_fedbatch = Inputs.range([0, 1], { value: urlNum("p_fedbatch", 0.20), step: 0.05, label: html`P(Fed-batch) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Low density (5–30 g/L), moderate media use (1–2×). Nutrient-concentrated feeds added periodically. Simpler to operate but lower cell density than perfusion.">(?)</abbr>` }) viewof p_perfusion = Inputs.range([0, 1], { value: urlNum("p_perfusion", 0.50), step: 0.05, label: html`P(Perfusion) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Medium-high density (30–150 g/L), higher media throughput (1–5×). Continuous media exchange with cell retention. Currently the industry standard for high-density CM production.">(?)</abbr>` }) viewof p_continuous = Inputs.range([0, 1], { value: urlNum("p_continuous", 0.30), step: 0.05, label: html`P(Continuous) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Highest density (50–200 g/L), efficient media use (0.5–3×). Near-steady-state operation; cells grown and harvested continuously with optimized recycling.">(?)</abbr>` })

Per kg of chicken-meat equivalent13Dry cost × 0.275: a kg of chicken-meat equivalent is taken to hold 0.275 kg of dry cell mass (27.5% dry matter, the midpoint of the 25–30% placeholder the current model uses), and the product is assumed to be made entirely of cultivated cells. Hybrids with a lower cell share are outside this estimate (crux 6).: P10 $70.4 · P50 $101 · P90 $138

Media (basal + GF-independent base)$/kg dry cell mass

n (CM_013CM_01, as asked at the May workshop: the 2036 production cost per kg of wet cultivated-chicken cell biomass, across large-scale plants. The wording was revised on June 25 and again in September; the link shows the history.),

Small teams (and AI tools) each build their own cost model of the same target, the 2036 cost per kg, conditional on one shared set of input parameters. Comparing the models shows how much of the experts' disagreement comes from the input values and how much from how each model is put together.

Overview May ’26 workshop synthesis Skeptical case Workshop record Questions & forecasts Evidence Recordings

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. Do not fill this out if you are human:

simulated $/kg wet biomass (clipped at P99 for display)

he cost model.

How to help: state your assumptions as numbers with ranges, try the browser simulator, or propose a change. You don't need to build models yourself.

ry the browser simulator,

Status: a pilot proposal with no date. We schedule a live event only once the target and a reference model are pinned down and enough people have agreed to lead (go/no-go).

from one shared set of input values

Squiggle cost model → An earlier version (v0.2/v0.3) on Squiggle Hub, no longer equivalent to the dashboard. Handy for forking and quick what-ifs.

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.

Cost → demand bridge: shows what a properly harmonized value-chain cost could feed into, while flagging that the current bridge is only illustrative.

Question update · September 10, 2026. The forms now separate achievable cost, actual cost conditional on production, production probability, and sales. Earlier responses retain their original wording and are not automatically answers to the revised questions. What changed and how to update.

Overview Synthesis

Beyond the paper: none, so this is the report's model with the report's values

f = 0.0001 for 5 years from year 1;

0.995

v_c

The answer is the purple box, E(M): the

Quick start: four questions ? After the dangerous period (the first century), wh

log10

Could the action backfire?

scenario ?

0.0001) last? ?5 years, starting now (report)11 years, from year 4 (RP 2026)50 years, starting now (report)Could the action backfire? ?No (report)Its effect is uncertain and may raise riskThe risk it averts comes back later The answers move the detailed settings below; the "Where the settings come from" line on the right tags each value. Presets: Report default · Report, cubic value · Report figures (f = 0.5) · RP 2026 central values ? Scenario viewof riskType = Inputs.select(["top", "constant", "filters", "decay"], { value: U.risk, label: lab("Risk scenario", "How the yearly chance of extinction changes over time. The report's main case is Time of Perils: high risk for a century, then low risk forever. 'What the risk scenarios are' below has the details."), format: d => ({top: "Time of Perils (report section 3)", constant: "Constant risk (Ord's base case)", filters: "Great Filters (report section 3)", decay: "Exponential decay (report section 4)"})[d] }) viewof valueType = Inputs.select(M.PAGE_VALUE_TYPES, { value: U.value, label: lab("Value growth", "How much each future year is worth compared with a year of today's world. With growth, far-future years count for more, so averting risk now is worth more. Options marked 'beyond' aren't in the report."), format: d => ({constant: "V0 Constant (report)", linear: "V1 Linear (report)", quadratic: "V2 Quadratic (report)", cubic: "V3 Cubic (report)", logistic: "V4 Logistic (report)", polynomial: "Polynomial of order n (beyond)", exponential: "Exponential, uncapped (beyond)", capped: "Exponential with hard cap (beyond)", fluctuating: "Fluctuating around a base path (beyond)"})[d] }) Risk scenario ? Time of Perils (report section 3)Constant risk (Ord's base case)Great Filters (report section 3)Exponential decay (report section 4) riskType = "top" Value growth ? V0 Constant (report)V1 Linear (report)V2 Quadratic (report)V3 Cubic (report)V4 Logistic (report)Polynomial of order n (beyond)Exponential, uncapped (beyond)Exponential with hard cap (beyond)Fluctuating around a base path (beyond) valueType = "constant" What the risk scenarios are Constant risk: rt=rhigh forever. Ord’s base case; with constant value, E(w)=vc(1−rc)/rc. Time of Perils: rhigh for the first L years (report: 100), then rlow forever. Great Filters: F high-risk eras of length ℓd separated by low-risk eras of length ℓs, then low risk forever. Report: F=2, ℓ=(100,500,100,…). Exponential decay: rt=r0e−λ(t−1)+r∞. Notebook defaults: λ=10−4, r∞=10−9. The report’s Figure 3 draws a much faster decay than the one used in its tables (see Departures). Per-year risks come from per-century probabilities: 20% per century gives rhigh=1−0.81/100=0.0022289; 1% per century gives rlow≈0.0001. Risk viewof rHighCentury = rangeIn("rh", lab("High risk (% per century)", "Chance of extinction per century during high-risk periods (or always, under constant risk). The report uses 20%, a common estimate for this century. The page converts it to a yearly rate.")) High risk (% per century) ? rHighCentury = 36 viewof rLowCentury = rangeIn("rl", lab("Low risk (% per century)", "Chance of extinction per century once the dangerous period is over. The report uses about 1% (a yearly risk of 0.0001). This number largely sets how long the future is expected to last, so it drives most results.")) Low risk (% per century

High risk

become? ?

persistence P and the horizon T. Only

Its value is E(M)=E(w′)−E(w), the change in the survival-weighted sum of future value.

Each

What this is and isn’t. A research model for seeing how the answer depends on the assumptions. It isn’t a forecast, an estimate of what any real intervention achieves, or a recommendation about where to give. Like the report, it counts extinction only; catastrophes that permanently curtail humanity’s potential without ending it, and partial catastrophes, are left out (the report’s footnote 1, p. 7). Under the results, each setting in use is tagged by where its value comes from: the 2023 report, Rethink Priorities’ 2026 model, an illustrative choice of ours, or you.

On this page

RiskValueNotebook TToolNotebookPaper (3 s.f.)Tool at 4T (change)P = 1

5116.5327361955315

Help choose research for independent evaluation by The Unjournal

follow that label to the internal prioritization hub (sign-in required) for an independent follow-up assessment. The label marks a workflow step, not an endorsement.

Needs further prioritization assessment

team

ind research worth evaluating: choose a topic, read the evidence, and add your judgment.

judgment

source title

Cause: Catastrophic risks & forecasting

Get involved — Suggest research for this list ↓ | See how this tool will work (planned workflow) ↓ | Would you sponsor an evaluation? (idea, feedback wanted

They are provisional, not ratings of research quality.

Scoring methodology

Conference Report: Threshold 2030 - Modeling AI Economic Futures · 82% match Directly models 2026-2030 AI effects on GDP growth, labor share, wages, unemployment, and labor reallocation.

hat is an impression to check, rather than an established absence of scrutiny.

overlaps it in what way? I thought that was more about forecswters, and this seems more like a theory model?

80Original AI lens before the Opus shift (not used for the synthesis)

The May workshop form asked about costs “across all large-scale plants in the world,” without an explicit condition that those plants exist. A June 25 revision added that condition; the later independent respondents received that wording. The new questions separate these interpretations. Earlier answers have not been reclassified as A1 or A2. Wording history and updated questions. Cross-cohort pooling is an exploratory comparison, not a set of answers to an identically worded question.

Privacy and provenance — human-reviewed 31 July 2026.

equation

The function

which the horizon charts plot.

Engine: survival(r), expectedValue(r, v).

Budget ($M)Voting MethodCredence-WeightedMy Favorite TheoryMECMETNash BargainingMSABorda CountSplit CycleLexicographic Maximin

We encourage you to blend your results with RP's preset views, or retake the quiz and combine multiple runs. Global Health & Development20%

I want to spend on whatever pays off most on average, even if the chance of success is small and the work may do harmEvaluate funds using a standard expected value. This is risk-neutral, so when calculating a fund's score, any negative effects can be offset by positive effects of equal probability and scale. A fund that has only a 1% probability of success still looks good if the effect is large enough.75%

If AI poses serious existential risks, the benefits of other interventions could be cut short so their value could reduce. Use the slider to apply a discount: 0% means no adjustment, 100% means only AI interventions are valued.

We use the public ratings export: per-paper evaluator labels ("Evaluator 1", "Evaluator 2" within each paper), so no one is identifiable and no one can be followed across papers

What this page is. We commissioned an external methodological review of our cost model and belief elicitation form. The reviewer identified several real structural problems. We are sharing the critique openly — and our responses — because we think transparency about limitations is as important as the model itself. (This page reviews the cost model and elicitation design; the subject-matter evaluation of the underlying forecast is Evaluation 2.)

Note on the interaction with cell density: Media cost per liter (/L)andcelldensity(g/L)areconnected—theyarenotparametersyoucanoptimiseindependently.Richermediaformulationsareoftenneededtosustainhighercelldensities,becausecellsindenserculturesconsumeproportionallymoreaminoacids,glucose,andothernutrients.Theeconomicallymeaningfulunitistherefore∗∗mediacostperkgofcellsproduced∗∗(/kg), which reflects both $/L and th

concrete: the public beliefs analysis

Page style:PlainLensTesseraeStudioWe're trying a few looks for this site. Your choice is saved in this browser only. Tell us which you prefer.

Three evaluators assessed this review of the discount rate GiveWell applies in its cost-effectiveness analyses. Two gave quantitative ratings, shown here with their 90% credible intervals.

Recently evaluated

So the concern looks right

back-dating or dropping those predictors barely changes accuracy.

suspected cause was leakage:

Every way of assessing research encodes someone's goals. In this workshop you choose whose. Each table takes on a persona, from a patient community to a government funder, with $1 billion to distribute. Your group sets its goal and agrees what it wants to know about researchers and their work. Then everyone goes into the research graph: find the people and research that match your persona's interests, decide how much each kind of evidence should count, and name the data the graph can't yet provide. Each persona leaves with its own open, inspectable assessment algorithm, ready for the funding workshop that follows.

Modular science breaks research into separately attributable, reusable components, so every question, claim, piece of evidence, and method can be credited, cited, and built on by itse

AT Protocol, the open protocol behind the Bluesky social network, stores your data in a repository you own; apps read and write it there instead of locking it away in their own silos. A growing ecosystem of research apps already builds on the protocol. You'll claim a researcher ID with Aster, then pick a station and start building. Stand up a personal research page with Sifa, turn your Zotero library into a living literature review with Semble, stream live data with Matadisco or Nebra, connect AI workflows to protocol data, or build Bluesky feeds and starter packs around your research topic. Or bring your own use case instead and work through it with the room's experts. We'll close with quick demos of what everyone built.

Together we build a data network owned by the people in the room: IPFS nodes on ten Raspberry Pis and participants' laptops, joined as equal peers with no cloud server anywhere in the sy

prioritization dashboardThis is a topic-focused working collection within the shared research-prioritization project. Each brief links to its dashboard record for scores, source material, and ratings. The dashboard contains a broader AI collection; these briefs are candidate assessments, not completed Unjournal evaluations or commissioning decisions..

What it says

Transcript

hat are they selling?

Top result

Who was closest?

Robot judgeAI judgeor just tap a name

Funniest guess (bonus, optional)

Cloud software that lets companies store, combine and analyse large amounts of data, charged by how much they use.

technology professionals help each other.It later moved to experts-exchange.com, with a hyphen. Read it again without one.See it for yourselfDraft: not yet re-checked online

Davidelsbeth.geldhof@gmail.coFunniest guess (bonus, optional)Davidelsbeth.geldhof@gmail.co

Who was closest?Robot judgeAI judgeor just tap a name

Give a clueTime!

Search it yourselfChecked 2026-10-06

Wikipedia's article on puppy chow

The inventor and his sister call it 'Tarzan' swinging, and a jungle yell is optional. A 2003 re-examination cancelled all four claims.Read the patentChecked 2026-10-06 · US 6,368,227

The inventor and his sister call it 'Tarzan' swinging, and a jungle yell is optional. A 2003 re-examination cancelled all four claims.

Clue 1: It's a cute reference site for people who build websites.

Marmite

Should each group maintain a short, evolving map of its highest-value pivotal questions and most promising impactful research directions? A proposal for discussion—at and beyond the meeting. The examples are starting points, not settled views.

hey sometimes disobey instructions, cheat on tasks,misrepresent their work, and are unable to com-plete some tasks at all, necessitating human inter-vention [23–25].

Preliminary evidence suggests that a software-driven intelligence explosion is possible. If onedoes happen, it could be the most consequen-tial technological development in history: AI sys-tems could rapidly eclipse human experts acrossmost domains and radically accelerate technolog-ical progress.

should urgently obtain more visibility into the automation of AI R&D,develop ways to steer and constrain an intelligence explosion, and prepare society to adapt to anintelligence explosion’s impacts

Preliminary evidence suggests that it coul

could AI progress radically accelerate in an “intelligence explosion,” where years of advances arecompressed into months or less?

The Economics of Recursive Self-Improvement

Because it is already receiving media and policy attention, the timing value for independent evaluation is especially high.

No separate written rationale is stored for it.

The main concern is that this may be more of a synthesis and agenda-setting policy paper than an evaluable quantitative social-science contribution,

Governments should urgently

Heavy early media attention

may

Source record Publication-stage and date evidence should be checked in the linked paper record. TARGETED_CURATED

Unjournal public-attention search

The welfare case is strongest under assumptions that future sentient welfare counts substantially, AI R&D automation could materially accelerate dangerous capabilities, and governance preparation can reduce risk without causing major counterproductive acceleration or panic.

An evaluation would focus on how well the empirical premises (the share of AI R&D already automated, the feedback-loop claims) support the policy asks, and on what is claimed versus assumed about speed.

whether and how states should require reporting on R&D automation.

AI R&D could compress years of progress into months and asking governments for visibility, rate limits and emergency plans.

would

Alan Chan, Christoph Winter, Andrew Barto, Jakub Pachocki, Geoffrey Hinton, Eric Horvitz, Yoshua Bengio, Dawn Song, Jack Clark, Hilary Greaves, Anton Korinek, Sam Manning, et al. · Catastrophic risks, the long-term future, forecasting