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:
Let's hide the 'github discussions' for now -- I suspect it's too frictiony for most people to engage ATM
P(Fed-batch) (?)
in most versions I tried the fed-batch has a miniscule chance of making $25/kg -- so maybe we rule it out ... no feasible world is it the cost-effective solution
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>` })
really the matter here is whether the probability is that each of these processes is actually the lowest-cost, feasible, legal method.
PILOT / DRAFT — proposal revised July 20, 2026. This is a work-in-progress proposal for a cultivated-meat modeling hack; no event date has been announced. Numbers, scope, and the in-browser simulator are illustrative and will change. Feedback welcome via the propose-a-change form below or by annotating this page with Hypothes.is.
this is taking up a bit too much space. Condense, use tooltips a bit.
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.
this should have a 'date of last update' ticker
Cost → demand bridge: shows what a properly harmonized value-chain cost could feed into, while flagging that the current bridge is only illustrative.
this has not been given as much attention and we'll probably set it aside for now
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
Note that there isn't so much sensitivity when the person has already agreed to put their name on the evaluation, but I appreciate you are being careful here.
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.)
This makes it sound like we asked a person to do this. I believe what we're seeing here is an AI 'deep research' report.
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
latex needs cleanup here
concrete: the public beliefs analysis
this shoudl be more prominent throughout this and other docs
Could the action backfire?
Problem I see here is this binary question – you’re asking me if I could backfire and of course you’re anything could backfire, but when I care about value maximizer is, it’s net affect, or perhaps the distribution of its effect
But more simply, if I turn the sun all of a sudden I think the page arbitrarily switches on some arbitrary backfire affect? I would want when I’m specifying whether I’m concerned about backfire to have a sense of what I’m signing up for in terms of the magnitude of that affect and perhaps be able to immediately make some choices about that magnitude
scenario ?
I’m selecting constant risk, but I was also asked about the high and low risk levels? What’s going on here should I only be asked about those if I’m assuming a nonconstant risk or selecting a model with the nonconstant risk?
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
This is doing to much anchoring? Perhaps?? It seems to cap at 20% there for the longer-term/smaller risk probability prr century
High risk
Maybe there should be a link here at this point in the presentation to allow me to adjust my choices – it’s not so easy to navigate to the slider or toggle at least that was not an iPad
become? ?
iPad. Once I chose one of these options, I couldn’t figure out or could easily find a switch or toggle to specific Pacific specific number choose a specific number.
persistence P and the horizon T. Only
Persistence of the effect of the mitigation? Horison is that the full length that you care about?.
Its value is E(M)=E(w′)−E(w), the change in the survival-weighted sum of future value.
This is underspecified, another explain. Again, try tool tip or link.
Each
In each year
The purple box, E(M), is the answer, measured in years of today’s world
This needs clarification. What do you mean precisely here. I don’t see what this is a measurement. You don’t have to fully fleshed out but put in a tool tip. If not , or link
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.
Say what it is , not what it isn’t. Details in tooltips. The not this it’s that sounds too ai
Page style:PlainLensTesseraeStudioWe're trying a few looks for this site. Your choice is saved in this browser only. Tell us which you prefer.
Style choice<br /> It seems to affect only a single page - shouldn't it affect all the pages on the site?
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.
It makes it look like the evaluation is only about the quantitative one. There's also a detailed report that this sort of obscures at first glance. That should be previewed, linked, or something in a manageable way without overwhelming. This should also have hyperlinks to the actual evaluation here, one presentation of it, whether PubPub or experimental.
Recently evaluated
These are not that recent - It might be worth putting a note or tooltip about the state of the pipeline and our greater focus on pivotal questions recently, and the revitalization of the pipeline with certain priorities as we await grant funding
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.
This question is really puzzling to me. I want to reflect. I would want this to reflect the true risks. I wouldn't want to prioritize a certain portion of risks over others. It's unclear whether you're asking me a value question or an empirical question here.
So the concern looks right
Which concern? This needs a little bit more expansion.
back-dating or dropping those predictors barely changes accuracy.
I'm not convinced that this rejects "leakage" as an important driver of the innacuracy.
Need to think about this a bit more myself: Suppose that model A1 was fit on "training data A" and part of the 'predictors in training data A' actually reflected outcomes, aka, there were 'leaks'.
We can compare it with "model A2" -- a model fit on training data A that did not include the leaks/contaminated predictor.
Suppose that both model A1 and model A2 used appropriate penalization/normalization and cross-validation, so as to not overfit (other than this leaks issue).
Consider "out of time testing data B", presuming the structural relationships did not change between the earlier and later dates
A1 should predict worse on testing data B than it did on data A . However, A2 should predict roughly as well on testing data B as it did on A.
A2 should predict better on B than does A1. That's ~because A2 did not get contaminated by these leaky outcome variables, which would throw off the rest of the model.
When you say "dropping those predictors barely changes accuracy" is this doing something like "running model A2 and then comparing A1 and A2 in predicting B"?
suspected cause was leakage:
Was the issue that I highlighted the most, I'm not sure the evaluators would say this was necessarily the most critical issue for them
Independent computational reproduction; s
confusing -- what 'independent computational reproduction'? That's not what this paper does -- although someone may have done a replication of it
hat is an impression to check, rather than an established absence of scrutiny.
? this seems like AI language got mixed into my actual notes?
overlaps it in what way? I thought that was more about forecswters, and this seems more like a theory model?
This doesn't have enough context -- the word 'overlaps' must have been highlighted as part of a particular discussion
80Original AI lens before the Opus shift (not used for the synthesis)
this 'before the Opus shift' is too inside baseball here -- make it clearer to people seeing this initially, and put explanations of things like 'the opus shift' in tooltips.
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.
This seems very much the sort of thing I'm interested in.
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
This seems somewhat more relevant to what I might want, to the Unjournal, and how I might think about it fitting into the research ecosystem. It also seems sort of interesting to me - is there any way the hangup here would be whether my co-authors are okay with me publishing this stuff "modularly"?
I could probably go ahead with the drafted (and largely AI built) evidence from the JustGiving seed donation experiment, though. Or maybe the https://llm-uj-research-eval.netlify.app/ work, which is, again, largely AI-driven.
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.
It seems sort of interesting, but I can't figure out how this relates to my own work or how to get involved with this without doing a lot of preparation and thinking about what I would find valuable here. I'd like a little bit more background on what all this stuff is about, what it does, and what the big advantages of doing this over maybe more standard approaches would be.
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
This seems interesting, but I'm a bit lost - why are we doing this? What is the goal, and why is this a good approach? What are we learning from this setup?
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..
this should also have a link to the dashboard in addition to the tooltip
What it says
tooltips going off the page on the left -- fix
Transcript
you might as well format the transcript and add headers ... make it a 'summary doc' or 'overview' or some such
hat are they selling?
This category is not particularly fun. Instead of just the name of the company, maybe show the logo with the name, or their ad slogan or jingle, like their catchphrase.
Top result
Tooltips or something should remind us what these categories are about.
Who was closest?
I can't see people's guesses on this screen - you should be able to see what people guessed when we want to see who was closest.
Robot judgeAI judgeor just tap a name
AI judge stopped working pretty quickly. Did we run out of credits?
Funniest guess (bonus, optional)
Some points should be given or whoever had the funniest guess
Cloud software that lets companies store, combine and analyse large amounts of data, charged by how much they use.
This one was super boring. Let's have fewer questions having to do with designing websites and tech support and things.
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
This one is not completely fair because there actually is a dash in that actual website
Davidelsbeth.geldhof@gmail.coFunniest guess (bonus, optional)Davidelsbeth.geldhof@gmail.co
For some reason, I'm not seeing the guesses here. Some data got lost there.
Who was closest?Robot judgeAI judgeor just tap a name
For some reason, it couldn't judge us. Maybe it lost our answers.
Give a clueTime!
The timer is not really working. 1. You shouldn't have a shared timer; each person should have their own timer. 2. There needs to be some penalization if you go over the time.
Search it yourselfChecked 2026-10-06
When I search it myself with this button, I get very different results. Make it clear which search it's using, and make it more likely that when they click this button, they get the same search. Also, does it matter if we put it in quotes?
Wikipedia's article on puppy chow
Does it usually come up with Wikipedia's article? Because I'm getting that a lot, and maybe it's a little bit boring.
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
For bonus points, maybe people have to guess whether the patent still exists or whether it was canceled.
Also, make it clear: are these all patents that were approved, or were some of them denied?
The inventor and his sister call it 'Tarzan' swinging, and a jungle yell is optional. A 2003 re-examination cancelled all four claims.
Give more pictures of things like this, maybe.
Clue 1: It's a cute reference site for people who build websites.
Fewer ones having to do with creating websites and tech stuff. That's not so fun.
Marmite
Try to use the actual logos if possible.
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.
This is obviously AI-generated... You see the "not this but that" and the passive voice, etc. Make it less so. Try to use natural human language and my own voice more, and also clearly acknowledge how it was created and who's monitoring it, etc.
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.
Update/tooltip -- - Note our streamlined policy -- summarize and tooltip this:
Oct 1 2026: For ~the next ten papers/projects (until we have more funding), a piece of research (which anyone can nominate) will be commissioned for evaluation once two management team members "strongly approve". If multiple field specialists strongly approve, we'll also take this very seriously. We'll announce these "conditionally approved" research objects in the relevant Slack channels (let us know if you want email notifications for this), and then everyone on the team, including field specialists, has one week to weigh in. If no one raises a strong doubt, we'll go ahead.
Needs further prioritization assessment
avoid bold. Italics preferred
team
--> "reach Unjournal team review"
ind research worth evaluating: choose a topic, read the evidence, and add your judgment.
Edit to "Find impactful research for your own work, and help The Unjournal prioritize research with impact-potential for evaluation/peer review. Sort and filter by field, sources, and "Cause (topic)" of your choice.
(USe tooltips on some of this, let's not be too repetitive either)
judgment
your judgment --> "your prioritization judgment" (or something like this
source title
I don't need the 'literal title' note each time -- it's distracting
Cause: Catastrophic risks & forecasting
red font is too strong and distracting. adjust
Get involved — Suggest research for this list ↓ | See how this tool will work (planned workflow) ↓ | Would you sponsor an evaluation? (idea, feedback wanted
make this it's own top level thing -- not under 'about this prioritization tool. Also make it less bold and distracting.
They are provisional, not ratings of research quality.
They are not ratings of research quality. These are also provisional, and AI-generated, unless noted below.
Scoring methodology
Let's have scoring methodology hosted/open on another page.
In general anything that goes beyond about 1 paragraph of text, that isn't core to the papers/prioritization, should continue on a linked page instead of an expander
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].
pdf format is is horrible here .much more useful to have a hover link to tell you what the reference actually is, what it says, and how it relates to the claim being made
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.
Being a bit snarky, this is compounding three "coulds" or "suggests that" statements.
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
Surely, it's hard to critique "should" statements, as they're sort of expressing an opinion. However, they could be recast as "This would be welfare-improving in the following sense" or "If you want to achieve this, this would be helpful" - which could then be critiqued more directly. Obviously, this is an abstract and thus needs to be compact.
Preliminary evidence suggests that it coul
"Could" is, at best, a sort of existence of the possibility and maybe a directional statement. It's not very precise.
could AI progress radically accelerate in an “intelligence explosion,” where years of advances arecompressed into months or less?
I guess they mean the equivalent of "years of advances at today's pace" or "years of advances if AI were not used in this process at all." ?
The Economics of Recursive Self-Improvement
You're linking the paper here, not a post on World of Forums. What's the connection?
Because it is already receiving media and policy attention, the timing value for independent evaluation is especially high.
Give some examples of this "media and policy attention ". OK you give it below. WHen you do this, please say "see below" and link the later section if possible
No separate written rationale is stored for it.
this sentence is awkward. Stored by whom? Rationale for what? When will/would this come?
The main concern is that this may be more of a synthesis and agenda-setting policy paper than an evaluable quantitative social-science contribution,
that's a big concern for me (David). Also hard to evaluate if it's mostly a 'here's what we believe and advocate' paper.
Governments should urgently
a "should" claim is not checkable unless it's broken down more precisely into factual and logical claims
Heavy early media attention
evidence of 'heavy early media attention" please
may
"may" is hard to check
Source record Publication-stage and date evidence should be checked in the linked paper record. TARGETED_CURATED
You keep giving the same link to the paper itself as if it's linked to evidence of other things outside the paper.
Unjournal public-attention search
Seems like a hyperlink other than the one you wanted to share.
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.
this VoI ToC seems logical to me. (Although the "future sentient welfare counts substantially" is a bit vague -- either it's obviously true, or there's some assumption buried here about the importance of the long term future)
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.
These focuses seem reasonable and useful to me.
whether and how states should require reporting on R&D automation.
what's the counterargument?
AI R&D could compress years of progress into months and asking governments for visibility, rate limits and emergency plans.
"Could compress" is a bit imprecise. What is the actual main claim or claims the paper makes?
would
'would' is too strong. 'could' is better
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
nice to show affiliations if possible -- maybe as tooltips
Unjournal tools The Unjournal's own tools live on their own page.
You might want to tease some of these tools here, though, in particular https://uj-cm-workshop.netlify.app/crux-map and https://unjournal.github.io/cm_pq_modeling/simple.html as modeling tools are pretty cool. The cultured meat one is not really Unjournal-specific, so maybe it even could be on this page in some way.
Econometrics, statistics and data science notes Working notes with a micro, behavioral and experimental focus. Web book Researching and writing for economics students A practical guide to economics dissertations and research projects. Course notes Microeconomics (MSc) Notes for an MSc microeconomics module at Exeter.
Nice, but also you should make it more clear in some way that these were built a while back entirely without AI. (NB-- I do intend to update these and help fill in the gaps now that we have a range of tools for lookup, coding, etc., leveraging AI. That is somewhat of a priority. ... Once we start that project, I want to clearly maintain the "pre-AI" version that people can see if they truly prefer to see what the source was before all this AI stuff started to happen, but the AI versions are probably going to be the main one.
Increasing effective giving The puzzle of why people don't give more effectively, what the evidence says, and what we still need to learn. Web book Impact of impact information on giving Field experiments and synthesis on whether telling donors about effectiveness changes what they give. Reports EA Market Testing Public reports from a team running marketing and messaging trials with effective-giving organizations. Data analysis
These are research projects, not really interactive tools. I wouldn't give them quite as much space on the page, and they probably don't illustrate much of what the page was meant to illustrate. Yes, keep them, but don't make them quite so prominent. Perhaps an outlink to a page laying them out futher
Games and everyday tools
These navigation things are helpful, but maybe also let people switch to a simple list and give them some sort of filter they can use to filter these by meaningful categories.
AI Foundations for Research Judgment How language models work, for researchers who have to judge their output: tokens, measurement, evidence workflows, compute economics and safety.
Not sure I'd highlight this. Maybe this should be a less forefront one? It's basically just an internal tool that I built for myself to learn that aggregates information from the internet.
and site GTD Trio Guitar, tuba and drums jazz trio in Western Massachusetts. Band site K-House Jazz A flexible live jazz group in Exeter, Devon.
Make these take up less space on the page (the band pages) and make them less prominent. They are not as interesting.
Play and study 18 versions of the 12-bar blues in any key, from the basic three-chord form to Charlie Parker's 'Bird' changes.
If you can do so without too much clutter, maybe with a hover thing or a quick note, put an update date on each of these and explain which models/tools they were built with, such as "built with Claude Opus 5.5." Some of the tools below were built without any AI tools, and that should be noted as well. But please, please avoid clutter.
plus the bands I play tuba and trumpet with
I play tuba, trombone, and flugelhorn, not really trumpet. . Importantly, the band pages are not really interactive tools. Don't make those quite so prominent. Maybe those should just be mini links.
AI-generatedThis is a model-generated prioritization rationale, not source text or a human assessment. Check factual claims against the source abstract and linked paper. (
It's good to say that things are AI-generated, but this standout purple text is a bit distracting - make it gray.
Stored
Why do you say "stored" evaluation rationale - what is stored about it? You don't need to give so much under the hood here - at least not next to every entry - you can explain how we came up with this at the top or in linked pages.
(not AI-generated)
If it's "source provided," you don't need to say "not AI-generated." That's obvious from context.
source title
Don't need this disclaimer on every paper. That's just adding clutter.
scale.
Say "Evaluation priority and global impact potential, not research quality. These are provisional."
Small score differences are not decisive.
this doesn't need to be bold
Sensitivity Analysis: Dollar-Swing Metric
These explainers should have backlinks to the section where the thing is actually used.
Workshop medians span roughly $1 to $100/kg
needs units
You are not asked to become a modeler; you are asked to make your expertise modelable. The
Fix ai tics language like this
visible CM_01 point estimates
Link and tooltip explain what this question is, also give an in-text abbreviated meaningful name for this question
Model and planning decision. Retain the numerical forecasts and commitment ledger. Compare scenarios with faster grantmaker expansion and different cause allocations; do not add this ecosystem total to the AI-linked ledger. For an organization considering expansion, the useful next evidence is an eligible funding program, a decision timetable, and a grant it can actually budget against.
This also requires some more context and explanation, perhaps with detail in additional fields or tooltips
Evaluation capacity deserves attention. The overview counts only nine publicly accessible evaluators with at least three full-time staff in evaluation or grantmaking. This supports investigating capacity constraints, but headcount alone cannot calibrate our dollar cap.
You're really not giving me any context here.
Roughly 30% went to global catastrophic risk reduction, versus 20% excluding those two organizations.
30% of what? This contrast is not clear.
evaluator-recommended destination
What are "evaluators recommended" destinations?
10 July 2026 revision: OpenAI GPT-5.6 Sol reviewed the Fable branch following the authors’ email discussion. Its additions and edits remain AI-generated and unreviewed by the authors. The revision preserves the earlier source text in Git history, corrects several overstatements, and adds a comparison with Epperson, Diederich, and Goeschl’s unit-donation experiment.
Things like this could be tooltips.
This is the “Fable-adapted” branch of the ICRC paper: consolidated, reorganized, and substantially extended by Claude (Fable 5) in July 2026, under David Reinstein’s direction. Every section carries a provenance tag:
Use folding boxes more. There's too much detail. It's too cluttered.
Impact Information
"Impact information" is not completely accurate: 1. Per unit cost, not ultimate impact - food parcels, not lives saved or something like that. 2. We vary the inclusion of this information, but we don't vary the actual impact of the donation because we're only using a single donation to a single charity.... - Maybe this updates people's beliefs about about the impact
In addition to providing evidence on the cost impact or the cost per output, at least, we are also framing it in a way that suggests to them that they are specifically buying an output rather than just donating to a broad pool, which reflects some of Epperson.
What we do is not 'purely clean' perhaps but it is field-relevant
Jan Schmitz University of Münster
Update Jan is not at Münster (why did you think this?). He's at Radboud.
Our closest comparison keeps the monetary ask fixed at CHF 150:
Closest comparison to what?
that a well-designed unit-donation scheme can increase giving, especially with a large unit size. Their full treatment does more than report a unit cost: it reframes the decision as choosing physical units and can restrict choices to a unit grid
This needs expansion or clarification. Perhaps in a footnote or tooltip’s
It does not show that private philanthropy can replace government aid, or identify how to build political support for official development assistance
This not that ai language. Adjust
“DONATE 50CHF TODAY: your donation can supply 3 food parcels (ca. 17CHF/parcel for one month) to a Syrian family”
consider doing a comparison between cost and cost_sug_50 because ??? @jan schmitz
“DONATE TODAY: your donation can supply food parcels (ca. 17CHF/parcel for one month) to a Syrian family”
this treatment relative to the control does 2 things -- tells them about hte cost, and makes it concrete what they are 'buying' w a donation
The mailing went out on April 22, 2021. An unrelated ICRC “door-drop” campaign followed around May 25, and a TV spot ran in German-speaking Switzerland in the same period. This triggered the preregistration’s Case I contingency: the narrow window (gifts through May 31) is the principal sample for hypothesis tests, with the broad three-month window as descriptive robustness. The narrow flag keeps all letters and simply does not count gifts arriving after May 31.
AI -- please make this language less AI sounding, and explain it more fully. Use tooltips for details
I. Impact of providing information about unit cost in a fundraising solicitation — primary questions.
strictly, this is providing information about 'unit cost'
3 Main results: The impact of impact information
this page should link to or prominently show the actual (or translated) text of the varyuing treatments in the letter
information
Link or footnote this “mixed evidence”
anipulation strength and relevance: the cost-per-output treatments must have been meaningful, salient, and representative of what charities actually do and would
This is important and a natural criticism. I think we can make a credible claim of naturalness. We get minor support from “the other dimension of treatment did have an effect here”. But the biggest limiting criticism I see is “did donors even notice this in a meaningful way?”
Our closest arm-level comparison holds the CHF 150 ask fixed. The control text asks for CHF 150 and mention
Recall and reconsider our thinking: cost info means something different when you have a suggested donation vs when it’s open ended?
scheme with a large unit size substantially increased average giving.
Did they find the predictable incidence va amount trade off and nonlinearity? How to compare the size of the ask / size of “large” units across these contexts…? I guess there’s were typical lab experiment type stakes?
not yet reviewed by the authors.
I am going through it right now to see whether it makes sense in a general way, and adding a few comments and suggestions and questions. You will also want to selectively and then fully check things in a manual and “fully human” way.
Aside: I suspect hybrid human ai research becomes more common soon, but for now I think people expect full human oversight
Table 4.2: Preregistered amount contrasts (SUG-HI-LOW-AMT, H10–H11), narrow sample. Outcome: donation amount in CHF including zeros (revenue per letter).
provide base rates in tables like this
Know of high-impact research we should consider for evaluation — or other work this tool should be doing? Leave an email if you're open to follow-up discussion; if we later introduce compensation for useful contributions, earlier contributors will be grandfathered in.
add a note here and elsewhere, along the lines of: We realize that you might have doubts about the value of adding content, ratings and comments on a largely AI-driven site. We'll try to personally respond to each substantial comment or engagement. And if you have doubts or want to flag us specifically, feel free to email contact@unjournal.org.
Since the workshop: a company-reported production run at 22,000 litres and a new hydrolysate review bear directly on the scale and media questions. Dated updates and their limits.
This comes across as strictly positive, but if you look at the evidence, I believe there was also some less positive news about finance.
9, 2026. These developments
The developments reported below make it look like things have only gone in a positive direction, but I believe that on the finance side, things may have been scaled back? This might be adjusted a bit to not present the wrong impression.
kind of food people would buy
This last clause is a bit vague - what do we mean, "the kind of food people would buy"?
generally expected
Quantify and link, potentially in tooltips to avoid too much clutter
wide
"wide" is a vague term. Can't you be more quantitative and reference something specific here, or at least link it?
differed over how far particular substitutions could go,
In a tooltip link direct quotes/sections - everything stated here needs some sort of attribution.
Rate/discuss (Team) Rate/discuss (Public)
Potentially, I could make these the same, just one box rather than two, but then people can fill in the "team options" within that interface.
Quick rate
tooltips should explain the translation into percentiles
Unjournal commission an evaluation? (-- Strong No … ++ Strong Yes)For aggregation and sorting, the five quick ratings translate to percentile-equivalent scores: -- = 10, - = 30, ~ = 50, + = 70, and ++ = 90. These are deliberately spaced away from 0 and 100..
make it clear that 1. these are translated to percentiles and 2. the "quick ratings" are less weighted/noted separately as 'quick ratings'
More filters and source controls
Try to convey this in a less dense way. Consider whether some of these are redundant or overlap others, etc.
Show/hide scoring rubrics & methodology
this is just 'show below' -- make it a link instead
About this prioritization tool
most "expandable" are arrows, but others are boxes
Optional uncertainty range (10th and 90th percentiles)
Just show these boxes -- no need to make that a folder
Can you give an A1 cost estimate?
skip this question
Jump to another question
it's not 'another question' -- just "Jump to any question"
Full definitions and accounting boundaries
, and advice on forming probabilities
at least 2,000 metric tonnes per plant per calendar year (2 kt/year).
this bit should not be in bold
Qualifying
new line before "qualifying scale"
The authors have made the atlas, methods, data, and replication code easy to inspect. We found some research indexing and light public discussion, but no major institution or policy report using the measures yet.
this is what authors did. not quite the same -- also interesting but make that separate
Several plausible routes could lower cultivated meat production costs.
this first sentence seems a bit contentless. "Could" is vague and "lower" is not really quantified ... do we mean 'substantially lower' in some measured way? Bring closer to parity?
Several plausible routes could lower cultivated meat production costs. The harder question is whether they can work together in a reliable process at commercial scale. Our current reading of the evidence favors testing complete, internally consistent production scenarios before treating lower ingredient prices or higher cell densities as a new cost forecast.
This is an AI-generated (Astra, Extra-high effort) synthesis takeaway. Even in its initial statement, the claims should be sourced and linked. The "not this but that" aspect of this reads a bit like AI style. Quote: "Our current reading" -- it's not clear who "us" is. What would it mean to "test complete internally consistent production scenarios"? It's not clear what's being contrasted here - one wouldn't treat lower ingredient prices as a new cost forecast, for example, but they might enter into such a cost forecast.
David Reinstien: To the extent this synthesis is making concrete statements. I'm not sure if I agree with them. At the very least, this should be stated in a much more contingent and tentative format.
Key Takeaways
Perhaps this should be on its own page, linked here and with a prominent tab or subtab at top
Embedded in media discussion
What does this mean?
Framing note: Recent industry consolidation — including several company closures in 2024–2025 — is one set of data points, not a verdict on any particular TEA's methodology. Company outcomes depend on funding conditions, market timing, and management decisions that are largely independent of the underlying cost trajectory. The workshop aimed to evaluate the technical evidence on its merits.
This should be a fold or a tooltip.
PF
I forgot what PF is here?
Public working summary Workshop held May 8, 2026 Last reviewed July 20, 2026 S2 was off-record and is excluded. Its section below gives only the announced session scope.
This isn't really a well-organized "public working summary." It goes into too much detail about the structure of the workshop rather than the overall narrative and a reasonable synthesis of the working results/takeaways.
Informal Pre-Session Wed May 6, 2026 · 11:00am–12:00pm ET · Zoom · recorded Casual orientation for participants with limited May 8 availability — mostly preparation and broad discussion. Walkthrough of the beliefs form and cost model. David Manheim (Technion/ALTER) and Mirjam Capuder (University of Maribor) participated. It covered introductions, a walkthrough of the interactive cost model dashboard, and early framing questions about key modeling uncertainties — including Manheim's perspective on expert elicitation methodology and how to structure belief-updating around contested TEA assumptions. Substantive points from this session are incorporated where relevant in the S1 and S3 notes below.
This is too much detail of the workshop's logistics here. Logistical stuff should be in folding boxes or tooltips.
These map onto the workshop's structure: cell-line engineering → Crux 1; hydrolysate substitution → Crux 2; albumin and recombinant proteins → beliefs-form question E4; "what share of a hybrid product is cultivated?" → Crux 4.
I don't see how this maps on at all - it seems like you're confusing things.
Regulatory milestones are not directly comparable across jurisdictions. The FDA's February 2026 inventory lists five completed US cultured-cell food consultations, including three completed in 2025; market authorization, inspection, and product scope can still differ.
I don't see that discussion in the evaluation 2-source it more directly and look for a quote.
it
"It" is unclear here
Evaluation 2 of this forecast
That was not an evaluation of a specific forecast -- It was an evaluation of a paper involving an expert-driven forecasting exercise.
in response to this and other evidence.
leave out "in response to this and other evidence"
S2-derived claim,
let's discuss this one (not here) -- if it's important to the model I could ask for permission to use if not quote
rigorous
"rigorous" is subjective. Think of a rephrasing here. Humbird was widely cited and influential. Pastika et al was published in Nature Food a peer reviewed journal that buy one metric ranks number 1 in it's subcategory and it has been heavily cited
asitka et al. (2024): ~$6/lb under continuous production; Humbird (2021):
tooltip should give the full title and a link to the work
S2 was off-record and is excluded. Its section below gives only the announced session scope.
make this a tooltip
Tool updated: Aug 28, 2026Normal cadence: papers twice weekly; feedback hourlyPaper discovery, scoring, matching, and the full dashboard rebuild normally run on Monday and Thursday. New rating and suggestion submissions are checked up to hourly. The dated chips show the actual latest completed updates.Last paper scan: Aug 31, 2026Last AI scoring: Aug 31, 2026Crux/PQ match update: Jul 23, 2026Human feedback: 17 current ratings across 13 papersPublic anonymous aggregate refreshed Sep 1, 2026. Each rater's latest rating per paper counts once. Discussion and a display name are included only after explicit public-sharing consent; emails, rating keys, and submission ids are never included.Earlier API cost: at least ~$4.06 reconstructed
less bold font here, this is taking up too much attention. It's good, but it's busy/distracting ... you can put all the updates dates in a more compact form, without so much bold. Actually, maybe fold these or put them in a tooltip ... just say "Last updated" and giv ehte date of the last paper scan, and then a tooltip with the other dates and the 'cadence' etc.
Normal cadence: papers twice weekly; feedback hourly
this is tmi too prominent for an inside thing. Make it a tooltip/less prominent
A public economics column lists the review among its sources. This suggests some circulation outside the World Bank, although the article does not assess the review’s methods or health-policy conclusions in detail.
it's given as a reference but I didn't see much discussion
Mean absolute score movement?
give a column for the overall MAD of scores to give a basis for cpmparison here ... to know if "6.3 is relatively large given how the scale is used' etc
?
tooltip -- explain better. What is the measure of interest? What was 'resampled' and why?
movement?
in tooltip -- when you give horizontal view it's no longer 'columns'
Agreement alpha?
link a description and explanation of these measures in this context. What are normal and comparison values for this? What is the cardinal meaning of this measure ... 0.5 vs 0.25 eetc., and how to interpret it?
ecent AI-impact/governance papers Ordinary rerun
text overlaps w next col
AI score stability metrics, with paper cohorts and variation sources shown side by side AI stability metric
tooltip cuts off on lhs of screen
17.2 points
between what and what? explain here/tooltp
cautious
what is 'cautious'?
rivate review found that one run treated prior public scrutiny as substantial while the other called it unclear, materially changing the estimated value of a new evaluation
that's very interesting!
identified publicly.
it's OK to do so
One same-wording execution pair differed by 30 points.
which one and what was the difference in wording?!
Prior human-rated comparison papers were less stable on both dimensions: ordinary rerun movement averaged 8.5 points and the wording effect averaged 5.8 points. 2/6 (33%) had wording effects larger than ordinary rerun movement; wording-mean alpha was 0.69.
this lower stability -- could it be somehow because of contamination from seeing human wordings, or because of greater variation in the nature of these papers?
Recent AI-impact/governance papers: ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points. 0/6 (0%) had a wording effect larger than their ordinary rerun variation, and wording means produced no broad-tier changes. Prior human-rated comparison papers were less stable on both dimensions: ordinary rerun movement averaged 8.5 points and the wording effect averaged 5.8 points. 2/6 (33%) had wor
worth discussing the pooled analysis too
Recent AI-impact/governance papers: ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points. 0/6 (0%) had a wording effect larger than their ordinary rerun variation, and wording means produced no broad-tier changes.
to me this suggests the wording variation was not particularly important -- is that reasonable? How does this compare to results in similar contexts?
ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points.
4.1 percentage points on the 0-100 scale? How does that compare to the typical rating variation between such papers -- drawing the reference group in a few reasonable ways
Six recent AI impact/governance papers and six papers with prior Unjournal team ratings
which papers? Tooltip and/or link or fold a list
026-08-31T23:54:58.113454+00:00
avoid this weird date format .. .I don't need the milisecond
Held-out check: a third baseline execution is kept outside the balanced headline analysis and used only as a robustness check.
explain (tooltip?) why 'holdout' is important here. I'm not sure I understand what it's doing in this context
What is being tested: every stability number below describes AI model scores. “Previously human-rated papers” names a paper cohort selected because earlier Unjournal ratings are available; those ratings do not enter the stability calculations. Human and AI scores are compared separately in the amber reference panel.
make this a folding box (beyond the first line)
mber reference panel.
I don't see any amber .. OK brown -- add an anchor and hyperlink down the page
paper-level human ratings.
would be fine to do so -- why not? These are already public IIRC. I'm also not worried about sharing AI-generated ratings. Remember, we're rating for priorization/impact, not research quality
calibration
what is meant by calibration here?
Average the two runs within each reviewed, equivalent wording, then compare those wording means. This reduces ordinary rerun noise in the wording comparison.
I got this upon re-read but was confused at first. Explain better. And is this the approach recommended in the apepr, or is there a better decomposition
Run each wording twice with the paper, rubric, model, settings, and output schema fixed. Differences within a wording estimate ordinary run-to-run variation.
Need some context here -- which models (Fable, Sol, etc) and effort levels are these being tested on?
prompt variants
i want to see the nature of these prompt variants early on, to get a sense of things. Show some examples concisely in a tooltip or fold
A shared first pass: Browse research that may inform policy, funding, or further evaluation. Scores rank the value of attention or evaluation; they are not quality grades.
add: metrics of mention on social media, citations in white papers, etc., ?altmetrics type metrics
Forum Post Cruxes and Pivotal Questions Explorer
this page should link research, where available, to each crux
Living reviewFilter to papers surfaced by one or more expert-maintained living literature reviews. A paper keeps this provenance even if it originally entered through another academic source. All Reviews ▾ AI Accountability ReviewBridging BoundariesExistential CrunchLauren Policy: Migration Literature ReviewMonitoring Gene DrivesNew Things Under the SunRegression to the MeatReskilledScaling in Human SocietiesThe Care GapThe Patentist Review / UJ overlapFind review posts that cite papers The Unjournal has already evaluated, has in its pipeline, or previously considered. This is intended as an engagement and follow-up queue. All Any UJ overlap UJ evaluated In pipeline / considered Targeted intakeRequested, topic-focused additions outside the broad recurring scan. Use this to include all records, exclude deliberately oversampled batches, or inspect only one targeted run. Include allExclude targeted intakeOnly targeted intakeOnly: AI impacts on global health and development (29)Only: Animal-welfare focused intake (19)Only: Empirical conflict replication-game intake (14)Only: Large-N GCR and existential-risk quantitative evidence (6)Soil-invertebrate crux sweep (6 cruxes) Crux/PQFilter to papers that map to one or more Unjournal Pivotal Questions or community cruxes (from the cruxes explorer). "PQ match only" restricts to papers mapped to a Pivotal Question. All Has crux/PQ match PQ match only RecencyFilter papers by how recently they were released or last updated. Useful for focusing on the newest research that may benefit most from timely evaluation. Any time Last 30 days Last 3 months Last 6 months Last year Last 2 years Last 5 years ModelThe AI model used to score each paper. New papers receive a standard GPT-5.5 assessment; stronger candidates receive a second, more thorough GPT-5.5 assessment with higher reasoning effort. "UJ historical" papers were prioritized by the Unjournal team (not AI). Papers marked WAIT are temporarily excluded from AI ranking pending a genuine model reassessment. All Models ▾ UJ historicalawaiting genuine model assessmentclaude-sonnet-4-6gpt-5.4gpt-5.4-minigpt-5.5 (codex headless, high)gpt-5.5 (codex headless, medium)gpt-5.5 (codex headless, xhigh)manual_public_followupopus (headless) Also show small-model scored (mini/haiku) Also show UJ historical Also show legal scholarship Targeted intake catalog · 74 records · 5 search families
this is too cluttered -- it should be a filter
Source text (unverified type) source noteText supplied by the discovery source; its status as a formal abstract has not been verified.
the paper has an abstract -- why isn't it being shown here?
AI-assisted prioritization for The Unjournal evaluation
Make it clear here very prominently (with some reminders in key places below) that these are not ratings about research quality -- they consider the potential direct or near-direct impact of the work from a global priorities perspective.
he AI premium is concentrated in frontier and intensive AI use, including closed-source models, paying or seasoned users, and long prompts, rather than casual or open-weight use.
How is this identified? Are they talking about 'firms whose returns correlate more with the aggregate use of these prompts in the AI consumption data"?
interaction-and-communication skill content strongly associated with higher AI exposure, while analytical, scientific, and operations-control skills load more negatively.
this seems particularly interesting -- does this havean implication for the labor market?
predicts
is 'predicts' the claim of interest? This is all backwards looking 'prediction' in a sense. is there an implied claim about future returns?
The paper informs live decisions about AI labor-market adaptation, AI industrial policy, competition and market-concentration monitoring, financial-risk exposure to AI shocks, and prioritization of retraining or adjustment support across occupations. It i
in what ways? Need more examples.
causal welfare effects.
what does 'welfare' have to do with this?