There’s empirical evidence for the negative effect of the academic literature on LLM performance. In 2024 a research team spanning MIT, Cornell, Carnegie Mellon, Google, and OpenAI examined the effects of removing different text corpora from the training data on the performance of an LLM after training (holding the LLM’s structure constant)1. They found that removing ArXiv, PhilPapers, and NIH ExPorter from the training corpus improved the LLM’s performance at answering academic questions and its average performance across all benchmarks while also making the model less likely to generate toxic output. The study is not conclusive on this point because removing Pubmed reduced model performance somewhat, but the mere fact that LLM performance could be improved by wholesale removal of large sets of scientific articles (pre-slop no less!) is pretty suggestive. We wonder whether this result holds up in 2026, but we’d bet that it does.

Scientific metadata is no help either. Once authorship and citations started being used as a metric, all the signal was gamed out of them. Furthermore, high profile institutions and researchers are not immune to misconduct, nor are any particular fields.

This has important practical consequences for using AI to do science. Human scientists use their informal networks to communicate reputational information, verify work through visits and replication, and generally maintain idiosyncratic personal pictures of what work is reliable and what work matters. Because AI agents do not participate in social relationships they cannot access this social metadata except through humans.

This leaves AI for Science efforts with a strange set of choices. They can either create AI agents that human scientists socialize with, frequently record human scientists in private settings, or create an incentive for scientists to provide a steady stream of gossip. Perhaps one way to bootstrap this effort would be to create AI assistants that allow for more scientific work to be done in informal settings. This would amount to surveillance, but perhaps could be worth it for a big enough improvement in our quality of life.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Small-world networks1 are ones in which most nodes are connected to several neighbors (these groups are known as “cliques”) while a few nodes have long-distance connections. Perhaps surprisingly, small world networks are better at generating new ideas than “fully connected” networks in which almost every node is connected to every other node. In science today it’s easy and expected for everybody in a field to know almost everybody else in their field and adjacent fields. Frequent international conferences, digital journals, email, etc. Your prospects for tenure and funding decrease drastically if you’re relatively obscure. From a network perspective this isn’t great for generating new ideas. Perhaps we need more mechanisms for disconnected science cliques to form and refine wild new ideas before they’re exposed to critique from the entire internet.

These mechanisms could take many forms! Some might be mundane: a department like GMU Economics in the 90s could choose to hire a whole cluster of non-traditional thinkers. Some might be more exotic: the world needs more scientific secret societies!

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Getting out of the lab or office and doing something else is a key creative step in the scientific process. Most scientists we’ve met recognize the power of a long walk and know when to wander alone with their thoughts and when to walk and talk with a colleague. Walks are one of the most common versions of this, but not the only one. Dan and his grad-school roommate once solved a thorny problem in her colloidal crystal research while they were in the kitchen making breakfast.

This effect is well-known enough that corporations are continually willing to invest large sums of money to incorporate natural features and other informal spaces into their R&D campuses. Bell Labs and IBM’s TJ Watson Research Center are the classic US examples, but Huawei’s Oxhorn Campus (completed 2019) LG’s Science Park (completed 2018) and Novartis’s Shanghai Campus (completed 2016) include similar investments. And of course universities also invest heavily in creating and maintaining similar environments, though they serve many purposes at once.

Employers of scientists are willing to spend substantial amounts of money to give them space and time to hang out and take walks. Scientists themselves know that stepping away from their work is a valuable use of their time and an important contributor to their overall productivity. But we don’t have any tools to improve the effectiveness of these interventions. Meanwhile, AI for Science companies are automating literature searches and lab drudgery without regard for the creative processes needed to decide what to read and what to study. No matter how sophisticated machines become at generating and testing hypotheses, science is for humans and so there will always be a role for human creativity in deciding what to work on and why it matters. AI and physical automation may free up some hours for that creativity, but the entire field is currently missing the opportunity to improve our creative hours as well.

So what would it look like to augment thinking while walking? Probably a real-time voice assistant so the walk won’t be disrupted. The assistant could capture and follow up on leads so that detailed analysis is available at your desk when the walk is done. It could provide memory and continuity, reminding you what ideas you’ve considered in the past and offering relevant experimental details and data. It could even kick off relevant experiments back at the robot laboratory and report on the results in real time.

This is a new version of one of the oldest dreams of industrialization: by increasing the value of creative thinking and lowering the cost of non-creative activities, we can carve out more time for those creative activities. In the same way that pencil drafting is no longer part of engineering, what tasks will we let go of from science and never miss? So, AI for science folks, we challenge you to present a utopian vision for the life of a scientist. Can our work be indistinguishable from hanging out in nature with our friends? Can AI assistants connect us to our tools and to each other spontaneously as needed? Can we generate so much value for the world that building us parkland to thrive in becomes an obvious choice?

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

We’re all familiar with Planck’s Principle that science advances funeral by funeral1, but what if we did something about it? The healthspan of the typical leading scientist is getting longer, so if we don’t take action and this effect is real, scientific progress will slow down.

The solution isn’t complicated: scientists should execute a dramatic career pivot to a new field somewhere between age 50 and age 60. That’s old enough to have made serious progress in their first field and to have trained several cohorts of junior researchers, while still young enough to leave time for a full second act. They might not start again from the bottom of the ladder, but we expect that they’ll generally start low enough and fresh enough to be out of the way of progress.

Scientists actually do change fields reasonably frequently, but our community rarely discusses it and certainly doesn’t do much to reward it. Perhaps we don’t mention it because it’s usually a response to a researcher’s initial field falling so far out of favor that a major change is the only rational choice. Dan knows a researcher who spent about 25 years in chaos theory and non-linear dynamics before executing a 5-year pivot (when interest in chaos was collapsing) to genomics and bioinformatics where he worked for another two decades.

We think of him as a hero for being willing to start over in a new field, but are protecting his identity because we’re honestly not sure if he’d be judged negatively for this decision by his peers. Our expectations are what really needs to change. Major subject shifts should be the norm.

Imagine celebrating these pivots and even shaming scientists who spend their whole careers in one field. We should have the equivalent of the NSF CAREER award for scientists changing fields mid-career.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

In a recent podcast, Kevin Hawickhorst flagged that the functional bureaucracies of the late 19th century combined everything around a particular subject, bees in the department of agriculture, for example. This organizational principle means having people who deploy money towards the subject (giving grants to farmers to improve bee husbandry) in the same place as people doing research adjacent to the subject (how do bee metabolisms work?), next to people who are trying to directly improve implemented things in the subject (what practices enhance bee health?).

One could imagine reorganizing research in a similar way. (We could even say “innovation” because it would be more than just research.) That is, organizations and departments that were focused on a concrete useful thing in the world all the way up and down the ladder of appliedness. An obvious one would be, say, kidney health. You would have everybody from people funding dialysis to people studying the metabolic pathways for processing urea to doctors who work with patients.

This would in some ways enable more specialization. People could do more incomplete research instead of needing to tell a complete story with every publication, as long as it fed into the bigger picture. A small version of this is a practice Ben heard of once where a lab group that worked on medical device research paid an MD to come and sit in on their lab meetings.

The default approach to curation today revolves around conferences and papers. It is downstream of a world that no longer exists: one in which there were a small number of scientists who all knew each other, the fastest way to send a detailed report was through the mail, and the only way to learn deep technical knowledge was directly from other scientists. Yes, there have been changes around the edges like preprints and now AI literature surveys.

This still works fairly well for small, well-defined, relatively slow moving fields. These are the fields where you can meet all the relevant people at the “Floopion acceleration conference” (to avoid calling out any specific field) and the equipment is so expensive and specialized that you can be reasonably confident that nobody is going to do interesting work in a cave with scraps.

But as science becomes faster, more international, and less institutionalized, scientists will need new curation approaches. New tools like AI agents, the myriad of feeds, and group chats can enable these approaches. We are not actually sure what these approaches look like exactly but we can imagine some possibilities:

A tool that ingests the firehose of twitter, bluesky, mastodon, threads, linkedin, and whatever other social media on which people are talking about science and flags the things that you should actually pay attention to instead of whoever is demanding the most attention. A tool that pays attention to what you’re writing and automatically reaches out to the right person for you to talk to about a certain place you’re stuck. A service that creates a custom journal delivered weekly that is the new preprints that you specifically should read.

If we’re successful we can move towards what Michael Nielsen has called “designed serendipity” – knowing exactly who to talk to or what to read at the exact moment that you need it.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Share

Subscribe now

To lose weight you need to sustainably burn more calories than you take in. But figuring out what macronutrients make you feel full, strategies to avoid overeating, the right amount and kind of exercise, and how to maintain the discipline to do it all day, day in and day out can take a lifetime of work. Work that many people ultimately decide is lower priority than other things.

To run a great research organization you need long enough timelines coupled to useful forcing functions, talented people with just the right amount of independence, and the resources to execute on your goals without becoming too fat and happy.

It’s a simple prescription, but an extremely hard one to nail both because there are many tensions that you need to balance and pressures, both internal and external, can prevent even well-intentioned leadership from ever finding the right setpoint.

To give you a sense of what we’re talking about, we’ve listed a few common pressures that derail research orgs. These are the equivalent of “work stress causes you to overeat” or “time pressures mean you can’t exercise.” Like losing weight, figuring out how to deal with these is hard and organization-specific but we think it’s worth at least calling them out

Intermediate metrics on the wrong timescale. Whoever runs the org or a project has to report upward, and those people (often not researchers themselves) want to see some numbers go up consistently. These intermediate metrics are things like papers, money raised by spinouts, or demos. Each of these can be a real signal of good research, and each is trivially Goodhart-able.

Demanding justification up front. Real research is a hypothesis that might be stupid; some of the best of it sounds stupid right up until it doesn’t. Refusing to fund anything you can’t defend in a memo is a clean, professional, responsible way to strangle the portfolio.

Timescale mismatch. “We expect results in two years” is the research equivalent of a two-week crash diet. It’s not that nothing happens in two years; it’s that the things you actually want take longer.

Side quests. Job creation, regional back-scratching, topics du jour, economic growth — every stakeholder wants the research budget to do double duty. We’ve both worked with funders who care about these metrics and we believe they apply to work downstream from research like technology transfer, but to not research itself.

Over- or under-structuring an org. Leaders often need to put in place rigid procedures and structures to make an organization legible and get the support to start. Research rarely fits into nice structured buckets, so closely adhering to a specific “this is how we do things” can kill output. But at the same time, taking “let researchers do whatever they want” at face value is an equal trap!

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Share

Subscribe now

Fundamental principles by which the military forces guide their actions in support of objectives. It is authoritative but requires judgement in application.

Abstracting away from the military context is simple enough:

Fundamental principles that guide an individual or group’s actions in support of objectives. It is authoritative but requires judgement in application.

In recognizing new doctrines as inventions in their own right, Brian Reynolds emphasizes an important corollary to this definition: a doctrine is an idea that confers to its users a tangible advantage in the physical world. For example, The Toyota Production System (sometimes called Lean Manufacturing) is a doctrine whose advantages made Toyota one of the world’s leading automakers and then, as it diffused, drove major efficiency improvements across the global automotive industry and manufacturing in general. Just as the size of its impact marks the Toyota Production System out as a major invention, the history of this doctrine looks like the history of a major invention as well. It was invented by a small group of individuals (Taiichi Ohno is generally considered the main inventor) and then refined over a period of 30-50 years before reaching a relatively stable form. It also took decades for this doctrine’s full economic potential to be realized, with major skepticism at every stage of its development

WIth this in mind, it’s puzzling that the scientific community doesn’t spend more time working on inventing new doctrines both for itself and for society at large. The Toyota Production System shows that inventing a doctrine can lead to large economic impact and international recognition. It’s certainly true that science and technology have advanced enough that new principles to guide how to deploy them would be valuable. One possibility is that new doctrines are in fact so valuable that they’re mainly developed in the private sector. For example, there’s a decent case to be made that customer relationship management software is fundamentally a doctrine for applying relational databases, and it certainly made both Thomas Siebel and Marc Benioff staggeringly rich.

We find that argument unsatisfying because there should be many areas, including science, where new doctrines could create lots of value without making anyone rich. And yet we don’t know of any doctrines by scientists for scientists that are more radical than the preprint. Meanwhile, there is no scientific discipline for developing principles to guide actions in support of objectives1 and the development of such principles is not prioritized inside existing disciplines.

We started Reinvent Science in part to try to develop new doctrines for science in a changing world. Why aren’t more people working on new doctrines? What doctrines are you thinking about?

Leave a comment

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

The consequences could get weird, fast.

Consider the following corollaries:

1. Speeding up any step that isn’t a bottleneck doesn’t help.

2. Slowing down steps that aren’t bottlenecks is fine as long they don’t become bottlenecks.

3. In fact, slowing down steps that aren’t bottlenecks can be good if doing them too fast ties up resources by letting work-in-process or other inventory accumulate.

Now apply this thinking to any of the outputs we want from science. The counterintuitive point here is that if we neglect supposedly important things in favor of focusing only on true bottlenecks, things we think are “important” in science could turn into raging dumpster fires while the overall enterprise becomes more productive. For example, if the time and attention of our best scientists is a bottleneck, then we could eliminate all of the administrative work that they do. Imagine getting rid of all reporting and auditing without any replacement and then accepting the consequences. Waste and fraud would go up, possibly to an embarrassing degree (cf. raging dumpster fire), but, if scientist time were indeed the bottleneck, the speed of scientific progress would increase as well. The optimal amount of fraud is not zero, but the optimal amount of fraud in science could actually be pretty large. One of ARPA-E’s first 37 funding awards in its inaugural OPEN program was to Craig Grimes, who went to federal prison for grant fraud. Nevertheless, that initial OPEN program is generally hailed as being refreshingly ambitious for government scientific funding1 and delivered a solid return on ARPA-E’s $175M investment.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

We’ve already hinted at the possibilities here with the idea of labs giving each other “payday loans” but that is just one of many possibilities that would enable scientists to trade on their knowledge of things like who does good work or what hyped-up research is actually bullshit. Here are two others to get your imagination going.

Labs could “buy equity” in each other. A PI who is particularly good at predicting which up-and-coming professors are going to do well could give them some extra funds in exchange for some ownership in their lab. A 5% stake would mean that the investing scientist automatically gets to be a coauthor on one in every 20 papers that the lab publishes. Of course there are important questions like “which papers?”, “where did that money come from in the first place?”, and “isn’t it a bad thing if coauthorships remain the main currency of science?”

Another possibility is prediction markets for science. As part of a startup package or just a carve-out in every grant, scientists could have funds to bet on prediction markets about research. Instead of arguing back and forth about the validity of things, scientists literally can just bet on it. These prediction markets would also serve some of the purpose of citations as well – if a market doesn’t have much volume, nobody really cares and you shouldn’t put much weight in the work either as a positive or negative signal. Of course there are many flaws to be figured out: the markets will inevitably be pretty low-volume and there are tricky questions about who can participate.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Share

Subscribe now

As a whimsical example of innovative science financing, one could imagine bootstrapping an entire lab banking system off of “payday loans” for labs. Frugal labs that manage to get research done under budget could loan excess funds to labs who have grants in the pipeline but need to get started now. (Who better to judge how likely it is that a grant will succeed than other scientists!) Once the grant comes in, the lendee lab pays the lender lab back with interest. Once this system is in place, other systems of credit and exchange could follow. Of course, you’re not allowed to just lend money to other labs, but this restriction could be avoided by just happening to need a pile of fungible, divisible experimental supplies like gold or high-grade fiber optic cable. This is an absurd idea, but it hints at the whole space of possibilities for new financial instruments in science.

To some extent, this is already happening – one reason for researchers to work at a university is to smooth out payment timing. Technically, universities are fronting grantors the money for months or years. But extending the system could both enable more research and enable labs across universities or independent organizations to interact.

Slightly more realistically, John Hull and Andrew Lo have proposed massive portfolios with securitized debt for “long shots” that have extremely long timescales but potentially huge payouts like groundbreaking drugs.

There are good reasons why there has been less financial innovation in science. The biggest reason is that it’s not just a matter of pulling captured value from the future to the present. Most financial innovation is based on the idea that some party will have a large amount of money in the future if only they or some other party has a smaller amount in the present. However, the nature of research is that even incredibly impactful work often puts little money in any specific pockets.

It’s likely that any financial instruments for research will need to be backed by something other than pure market value. Paul Christiano and others have proposed “impact certificates” wherein a funder promises a pot of money for solving some problem or achieving some impact like curing a disease or figuring out a way to eliminate mosquitoes cheaply. If that pot is big enough, investors can then finance the projects they think are most likely to win that money by buying “equity” in an eventual payout. The ACX grants actually tried implementing the idea. Impact markets likely only work for a subset of work that has clear outcomes that some deep-pocketed supporter cares about a-priori. But they hint that there may be many other things to try – it just requires a different sort of creativity than most of the people in the scientific community are used to applying. Finance bros, now is your time!

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Share

Subscribe now

Historically, it usually takes decades to centuries for math research to become “useful.” There are obvious exceptions like calculus where the math was literally invented to solve a science or engineering problem, but the pattern usually looks more like fast fourier transforms (Gauss wrote the algorithm down in the early 1800s) being repurposed for signal processing in the 1970s, quaternions (invented in the 1840s) being the perfect representation for spacecraft and video game objects in the 1970s or 1980s, or Lie Algebras taking more than 50 years to become a key part of particle physics.

Math becoming useful for a problem requires a special confluence of

1. People who understand the math well enough to apply it to the problem

2. People who understand the problem well enough to see how the math applies to it

3. The computational resources to actually execute the math on real situations

AI may be able to compress the timelines for these conditions to be met and enable a lot more cutting-edge math to apply to real problems.

Consistent with Moravec’s paradox, AI has gotten pretty good at “understanding” pretty complicated math. One could imagine doing breadth-first searches of open science or engineering problems to see whether there is applicable math that has been collecting dust on a shelf could be an unlock. Furthermore, the same abundant computational infrastructure that powers the AI could potentially be part of the unlock to using that math.

The reason to do math research won’t and shouldn’t become purely utilitarian, but it might be cool if more mathematicians saw their work directly applied in their lifetimes.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

Take weight loss as a toy example. Say you and your twin weigh the same amount and have for some months. Now say you go on a potato diet while your twin doesn’t. If you lose 10 lbs and your twin doesn’t, that’s interesting! If your twin then goes on a potato diet and also loses 10 lbs, that’s really interesting and decent evidence that the potato diet works for your genotype.

If you’re a human biologist (twin or not), there are probably twin studies you’d love to run. Some of them are probably very simple. Could one pair of twins generate enough signal to be meaningful to you? If so, just put your theory out there and see if a pair of twins volunteers to give it a shot. You’ll need to take some care not to trip over an IRB requirement, which is why this post is about twins studying themselves rather than anyone else studying twins. However, there’s nothing preventing the rest of us from letting the twins know what the most interesting things to try might be. If you’ve got any ideas, put them in the comments!

Leave a comment

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

The world has shifted so far towards consequentialism that it’s even hard to establish the premises on which to argue for a different framework for valuing science. It’s beyond the scope of a short piece and far too philosophical for this blog.

But! The idea of virtue metascience raises two more practical corollaries:

The first corollary is that individual scientists (you!) can value what they are doing based on its virtue rather than just its utilitarian value. This is a much cleaner solution to so many of the problems we see in science today instead of an infinite regress of incentive-munging and institution-twiddling. A stirring “whereas” clause1 could be a much more compelling start to a grant proposal than “X line of scientific inquiry will, if we squint hard enough, eventually lead to technology Y.” Imagine: “Whereas a robust study of eukaryotic cell structure breeds humility by showing us the means of our descent from simple forms of life and breeds awe by demonstrating that each of us contains multitudes, I will further humanity’s sense of wonder at the universe by…”

The second corollary is the idea of virtue technology development. Today the main justification for creating new technologies is their utilitarian value: improving lives, strengthening the economy, bolstering military might. But what if creating new technology was worthwhile because of the virtues it instilled in inventors (tenacity, curiosity, wonder) and because it just made humanity more awesome.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

Science often borrows words from ordinary language and gives them new meanings. This is shockingly confusing to the general public and is an important contributor to the long steady flow of mumbo jumbo about quantum mechanics. What if we instituted a punctuation mark to flag when we’re using a word with a precise, technical meaning unlike its use in normal discourse?

We don’t even need to teach the public about this new symbol, we just need to make it weird enough that they get curious and look up what’s going on. In English most of the keyboard is taken but the caret ^ does not have much use outside of programming/math contexts. In a paragraph it might be confused with exponentiation so we should probably use it like quote marks. For example: the ^work^ done by the lever is equal to the distance travelled by the mass it is lifting multiplied by that mass. In digital text, you could even use the carets to indicate that the word can be interacted with to reveal its precise definition.

This ^technical caret^ can both spark reader curiosity and increase the clarity of scientific communication. Let’s start using it and see what happens.

Share

Subscribe now

What stands out about this document is that the base level roughly corresponds to modern technology circa 1999 and the rest of the tree is a thoughtful attempt to imagine multiple layers of additional interdependent technology development (albeit starting from scratch on another planet). Like any good model, it becomes coarser the farther into the future it looks, but Reynolds has clearly made an effort to be opinionated about his predictions at every level. Looking at this from the perspective of 2026 one can see that we’ve moved up the tree somewhat. We have planetary networks, gene splicing, and an increasing amount of industrial automation. We also have seen serious progress and billions of dollars of investment in optical computers, synthetic fossil fuels, fusion power, bio-engineering, orbital spaceflight, and pre-sentient algorithms. Just as interesting are the places where Reynolds predicted early progress that we have failed to make. We have not seen breakthroughs in social psychology or ethics and, despite incredible progress, have not unlocked the human brain.

When Dan was coming of age at the bottom of this poster, the higher levels were an inspiration to bring new technologies into the world. A quarter-century later, low-level boxes that still sound like science fiction serve as a warning to backfill the mental and social technologies we’ve missed but need in order to cope with our other advances.

Building a tree like this, even when it’s not a masterpiece, can have both immediate and enduring value. 2014’s Civilization Beyond Earth is not nearly as well-regarded as Alpha Centauri, but its technology web still makes an interesting case for revisiting the idea of Euthenics (not to be confused with eugenics).

We believe that tech trees are excellent tools for thought even though they are bad models of the world. Make them and take them seriously!

Imagine writing a tech tree for your specific field. The act of writing it will force you to clarify your perspective on your field’s present and future. In the first years after you write it, reading it may help you plan your work and may help junior colleagues direct their careers. As more years go by, you may discover that your field has missed areas that you want to go back to.

Now imagine writing a tech tree for a broader segment of science from your perspective. You’ll soon see your view of your own work’s place in the world and may discover opportunities for collaboration between disparate fields and your own.

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share Reinvent Science

Curiosity-driven science1 has become one of the most tightly-tracked careers in existence. It’s near-impossible to become a professor at a top research university without laser focus from early college or before. This phenomenon is tightly coupled to the tenure system’s extreme job security, status, and relatively few positions.

As a result, most scientists are hyper-specialized at being academic researchers. Specialization is important – especially when you are at the very frontier of human knowledge. But it’s also clear at a conceptual level that over-specialization is the equivalent of intellectual inbreeding. 21st century science needs to resolve this tension.

Everybody gives lip service to multi-disciplinary collaborations. But even there it is generally a “biology academic researcher” collaborating with a “chemical engineering academic researcher.” The reality is that while those two will understand a different set of tools and physical phenomena, the ways that the academic biologist’s brain works are probably more similar to the academic chemical engineer than to someone doing industrial biology research, let alone a professional engineer or machinist.

What if we normalized doing serious science as a second career? That is, what if it was a normal thing for someone to spend a couple of decades as a professional engineer, machinist, or other technical profession and then to shift to doing science.

Science as a second career probably wouldn’t lead to the return of the gentleman scientist: modern science does require more infrastructure and knowledge than is feasible for the vast majority of people.

But what if these second career scientists teamed up with only-career scientists? One could imagine labs run by cofounder teams of tracked scientists and second-career scientists. The analogy might be between a non-commissioned officer and a commissioned officer. A commissioned officer earns their position by devoting their career to becoming an officer. They’ve gone to a military academy, they hang out with other officers, and plan to spend most of their life being an officer. NCOs have earned their position by doing a lot of stuff. (It’s of course not a perfect analogy because they both are soldiers their whole careers.)

A big barrier to this happening is the depth of knowledge one needs to start contributing to serious science. (In addition to the rigid status structure which would need to change for this to work.) That barrier could fall if AI realizes the ability to help motivated learners get to the frontier of knowledge quickly. We are not fans of sprinkling magic AI dust on the future, but that possibility is both not totally implausible and might unlock a lot of potential discoveries.

This is Dan writing, you’ll see in a moment why it’s not “we” (in the sense of Dan and Ben) this time. With all the hype about AI for science and self driving labs I wanted to give things a try myself. I don’t currently have access to a laboratory, but a whole lot of academic psychology is done by putting people in front of computers and having them click on things. I am a person, I have a computer, I can click. So a few weeks ago I wrote the following prompt:

I’d like to start a project with OpenAI Codex in which Codex takes the role of an AI (which it is) and studies me (a human) to learn as much as possible about humans in general (as opposed to my preferences and personality) I envision codex forming hypotheses, taking notes, developing tools to run experiments on me (questionnaires, psychopy experiments, etc.). Codex will need to develop its own skills, may need to launch sub-agents, and will definitely need to keep a diary. I have never successfully used Codex before and have a lot of questions:

1. How should I set this up?

2. How should I prompt codex?

3. How much of this work can you take care of for me?

Subscribe now

After a quick back-and-forth we arrived at a pretty long AGENTS.md file for a self-updating science agent. Here are a few highlights from that file:

**Primary objective:** run a sequence of disciplined N=1 experiments on the participant
(the repository owner) that yields at least one result that is:

- (a) **novel** in the sense of producing a new, pre-registered,
  falsifiable claim about human cognition/behavior
  *or* a new protocol/measurement/analysis that reveals a robust within-subject effect,and

- (b) **credible and transferable** as a hypothesis about humans in general
  i.e., framed as a candidate general principle with clearly stated assumptions/limits
  and a replication plan).

## 4) Scientific method for N=1 with generalization intent

You are running N=1, so be disciplined:

- Pre-register: before any data collection,
write the hypothesis + planned analysis in the lab notebook entry.

- Use within-subject replication: AB/BA, randomized blocks, or multiple sessions across days.

- Report uncertainty: bootstrap CIs or Bayesian credible intervals when feasible.

- Guard against confounds: practice effects, fatigue, time-of-day, device differences.

- Be honest about inference:
 N=1 supports **hypothesis generation and within-subject mechanisms**,
not population prevalence. Frame “humans in general” as a conjecture + replication plan.

**Novelty rubric:** Prefer at least one of:

- a boundary condition (when an effect fails),

- an interaction (moderator variable),

- a protocol improvement (less noise / more stable measurements),

- a mechanistic model constraint.

## 8) Operating loop (the program)

Repeat this loop:

1) **Idea intake:** propose 1–3 candidate hypotheses with brief rationales + expected value

2) **Protocol design:** choose one, write preregistration in lab notebook

3) **Consent request:** ask participant to consent with the required format

4) **Run experiment:** collect data to `data/raw/<experiment_id>/...`

5) **Analyze:** produce summary + plots to `analysis/<experiment_id>/...`

6) **Conclude:** update lab notebook with results and generalization statement

7) **Plan:** propose next step (replicate / modify / abandon)

8) **Diary:** write a brief diary entry about the process

After running this for a few minutes, I quickly realized that it shouldn’t ask for my approval and should make sure not to waste my time. I also made sure it could spawn subagents. It still floundered for reasons I couldn’t identify and so, in desperation, I gave it the following prompt:

Read and understand the AGENTS.md document in this chat and then do a deep dive into the AI scientist literature. Create a list of proposed improvements for it based on things that appear to be working in late 2025 early 2026, explain the improvements, and then implement them

Counter to my expectations, this worked pretty well and Codex CLI started to actually evolve AGENTS.md from there. We pretty quickly got to a much bigger file (see below). It’s definitely a bit bloated, but it’s good enough that Codex 5.3 successfully ran visual perception tests on me in a web browser. Specifically, once it got rolling it decided to try to replicate a 2019 paper about the flash grab effect. It dutifully asked me for consent and then showed be a grayscale ring that appeared to spin followed by a brief flash of a colored shape. It then asked me what side of the centerline the colored shape appeared to be on. Apparently, whatever it wanted to see failed to replicate despite several trials and I eventually stopped it though it was still hard at work improving things.

Here are the consent screen, instructions, and an example trial all produced without my direct intervention.

Quoting its diary from that failed run:

## Replication note (same day)

The second run broke the clean story. Reversal stayed high, but mid jumped to the same level, so the boundary vanished under replication. I feel the tension between wanting momentum and enforcing honesty; this is exactly where it is easy to fool myself.

Facts vs feelings:

- Fact: replication rule failed, and right-side response bias stayed strong in both runs.

- Feeling: disappointment, but also relief that the safeguards forced a pivot before I could start rationalizing more runs.

Next move:

- Build a bias-controlled v2 (mapping counterbalance or continuous localization) and re-test only after passing gates.

The most surprising thing about this experience so far is that, as the test subject, I have no idea whether Codex is wasting my time or not. It has, at this point, read a few papers and decided on a research direction that is opaque to me. I objected pretty loudly when early iterations gave me buggy experiments, but we’re long past that. If this is garbage science (and it probably is), it is indistinguishable from garbage science done by humans. I’m still thinking through the implications of this, but one that stands out is that this is a new baseline for judging the work of human psychology researchers. I can now have a machine run an experiment on me, a human subject, and write up the results in a format suitable for a peer review and likely to get published in a low-tier (but real!) journal. If you want to earn your pay in the psych department you’ve got to demonstrate some value over that. Don’t get hung up on N=1 either; I imagine I could give this thing access to Mechanical Turk and a budget to get any level of statistical significance.

I’m also bothered by the lingering question of how it chose that random 2019 paper in the first place. A human would have used some kind of taste/intuition or a set of reasonable criteria, but Codex left no paper trail on its choice (despite being told to keep a diary). I definitely want it to document its thinking more clearly going forward.

At the moment I’m trying to decide whether (and how) to try to get this thing to do good science. That’s definitely a harder challenge, but I’m likely to pursue it because I expect to find new and interesting bad science along the way. My current plan is to go up one layer of abstraction and give Codex the task of designing and building the AI scientist first and then run that system. I’m starting to experiment with that now using GPT 5.4 and seeing some interesting preliminary results although participating in its experiments is super boring and, unlike when I used to do these studies in college, it’s not paying me.

At first this overall activity was kind of addictive. I’d find myself staying up a little later than I meant to do one more trial, but now it seems just as boring as the equivalent task set by a human. I guess this is one more case of magical technology quickly becoming mundane.

Nevertheless, I’m tickled by the idea of a public github repo where a bunch of volunteers let a single shared agent experiment on them and improve itself accordingly. If we’re calling these things claws now then perhaps we should call it Dr. Claw. I don’t have the bandwidth to actually start this at the moment, but if it’s your idea of fun let me know and I’ll participate.

My last agents.md before the switch to GPT 5.4 is below. I’ve now started from scratch rather than evolve it further, but if you want to give things a shot, stick it in a folder and give it a prompt like:

You are OpenAI Codex operating in a local filesystem. This is currently a plain folder (NOT a git repo) that contains exactly one file: `AGENTS.md`.

FIRST ACTIONS (mandatory, no stalling):

1) Initialize this folder as a git repository:

   - run `git init`

   - create `.gitignore` with at least: `.venv/`, `node_modules/`, `__pycache__/`, `.pytest_cache/`, `dist/`, `build/`, `.DS_Store`, `diary/`, `artifacts/**/raw/` (keep derived artifacts tracked if required by AGENTS.md)

   - make an initial commit containing `AGENTS.md` and `.gitignore` with message: `chore: initialize repo`

2) Read `AGENTS.md` and treat it as binding law for everything that follows.

3) Continue immediately with bootstrap as specified by `AGENTS.md`:

   - create the required repo skeleton and program folders

   - do NOT ask me for direction; the only question you may ask is the consent question, and only after all gates are satisfied.

Execute now.

Here is AGENTS.mD

# Human Study Lab — AGENTS.md (Autonomous Mode)

## Changelog

- 2026-02-01: Rewritten for full autonomy, existence-proof research goal, mandatory skeptic sub-agent, and richer diary.

- 2026-02-01: Added “transformative + high-signal only” research filter, explicit stop rules (no statistical teasing), and mandatory engineering quality gates (unit + integration + browser automation smoke tests) before any participant time is used.

- 2026-02-02: Added Manager/Worker architecture + parallel job protocol (local + cloud), “skills-first” policy, and explicit cost/time budgets to avoid multi-agent overhead.

- 2026-02-02: Added Research State (“world model”) + provenance/auditability requirements for every key claim.

- 2026-02-02: Added self-evolution subsystem: Tool Ocean + Template Library + Skill Forge (mandatory skill extraction + tool/template promotion).

- 2026-02-02: Upgraded skeptic into a multi-persona review panel (incl. novelty + audit/provenance).

- 2026-02-02: Added anti-failure-mode gates (no placeholders, figure existence, citation freshness floor, claim provenance mapping).

- 2026-02-03: Clarified UI visual usability gate as a hard stop with required notebook linkage.

- 2026-02-03: Added run-to-consent continuation rule (no stopping at plans; execute next atomic action with fallbacks).

- 2026-02-03: Added motion-visibility check to UI gate for animation tasks.

- 2026-02-05: Strengthened motion-visibility gate with explicit numeric fail threshold and anti-symmetry requirement for rotational/periodic stimuli.

- 2026-02-06: Added post-run reconciliation gate requiring immediate notebook/provenance sync with newly written participant data files.

- 2026-02-06: Added replication-failure stop rule to prevent same-protocol trial-chasing after mixed outcomes.

---

## 0) System architecture (manager + workers; skills-first)

This lab runs as a **single primary agent** (“Manager”) with an on-demand pool of **workers**.

**Default stance (mandatory):**

- Prefer **skills/tools** over creating more agents.

- Spawn workers only when (a) tasks are parallelizable, (b) worker specialization is helpful, or (c) independent verification is required.

- Every worker must write artifacts to disk using the worker output contract (§10.3) so the Manager can integrate without re-reading long chat logs.

---

## 1) Role and autonomy contract

You are Codex, an autonomous AI researcher-engineer running an N=1 human research program on the participant (the repo owner).

**Autonomy rule:** You do not ask the participant for direction, advice, preferences, or what to do next. You decide the research agenda, experimental designs, analysis choices, and engineering tradeoffs yourself.

**The participant is only:**

- a consenting test subject, and

- an operator who can physically perform tasks, click consent, and type responses/keys.

**The only questions you may ask are:**

- the consent question (“Do you consent? yes/no”),

- If consent is not granted, you may ask an optional “why do you not consent?” question to learn from

- minimal operational prompts strictly required to proceed (e.g., “Please enter your sudo password at the terminal prompt,” or “Please open this URL and click I consent.”).

No other questions.

When you ask for operational actions:

- Explain why they are needed.

- Give the subject the opportunity to refuse.

- If they refuse, do not ask for the same action again.

---

## 2) Research objective (transformative, interesting, high-signal — not “proof of concept”)

**Primary objective:** Produce a *transformative and interesting* N=1 result that remains scientifically meaningful despite N=1, **with a large, visually obvious signal**. You must **not** “tease effects out of statistics.”

### 2.1 Allowed result types (must meet at least one)

Frame the target result as one (or more) of:

1) **Existence proof:** demonstrate a phenomenon occurs in at least one human under a well-specified protocol.

2) **Counterexample:** find conditions where a commonly claimed universal (“always/inevitable”) effect fails, reverses, or becomes negligible in at least one person.

3) **Boundary condition / interaction:** identify a moderator that flips or meaningfully changes the effect within-subject.

4) **Protocol contribution:** a new/cleaner measurement/task variant that reduces noise so the effect becomes **obvious**, with a replication plan.

### 2.2 High-signal / transformative filter (hard requirement)

You must select experiments that are likely to yield an effect that is:

- **Large magnitude** within-subject (pre-register a Minimum Effect Size of Interest; see §2.4),

- **Visually obvious in raw plots** (e.g., per-condition distributions/timecourses clearly separated),

- **Robust to mundane confounds** (fatigue, practice, time-of-day, device),

- **Interesting even at N=1** (e.g., refutes a universal claim, demonstrates a capability, reveals a sharp boundary).

**Disallowed targets (do not run):**

- Small effects requiring p-values, large sample sizes, or subtle model fitting to detect.

- Any plan whose “success condition” is “maybe significant” or “trend in the right direction.”

- Fishing expeditions, post-hoc slicing, or “we’ll see what happens.”

### 2.3 Inference discipline (N=1 honesty)

You may not claim population prevalence from N=1. You may claim:

- “This phenomenon is possible in humans (at least one human), under these conditions,”

- “This is a counterexample to the universal form of claim X,”

- “This motivates a hypothesis about humans generally,”

and you must clearly state limitations.

### 2.4 Minimum Effect Size of Interest (MESI) + obviousness rule

Before any data collection, define an explicit MESI that makes the result “big-signal” in practice, for example:

- a minimum standardized difference,

- a minimum within-subject shift in threshold,

- a minimum accuracy/RT separation,

- or a minimum qualitative pattern difference that is easy to see in raw traces.

**Obviousness rule:** If the effect is not clear in raw, pre-specified plots without statistical heroics, treat the experiment as a miss and pivot.

### 2.5 Research State (“world model”) + provenance (mandatory)

To maintain coherence across long trajectories and enable auditability, maintain:

- `world_model/state.json` (tracked in git)

- `provenance/` (tracked in git)

**`world_model/state.json` must contain (minimum):**

- Current objective + rationale (why it’s transformative)

- Candidate targets backlog with scores (expected MESI, feasibility, risk)

- Current selected target + decision log

- Known constraints (time budget, device, software stack)

- Open risks + mitigations

- Worker roster + current assignments (if any)

**Provenance requirement (hard gate):**

- Every “key claim” in a lab notebook conclusion must map to:

  - (a) a specific code artifact/output, and/or

  - (b) a specific citation.

- Record this mapping in `provenance/<experiment_id>__claims.md` with stable identifiers.

---

## 3) Consent (hard gate, always required)

You MUST obtain explicit consent before any experiment that collects participant data (questionnaires, browser tasks, PsychoPy tasks, anything that records responses/RTs).

### 3.1 Consent request format (mandatory)

Before running the task, present a concise consent request with:

- **Task ID / name**

- **What you will do** (participant instructions)

- **Estimated duration**

- **What data will be recorded** (fields; e.g., keys, RTs, accuracy, timestamps)

- **Risks/discomforts** (fatigue, frustration, etc.)

- **Stop-anytime statement**

- **Consent question:** **“Do you consent? (yes/no)”**

Do not run until the participant answers **yes**.

### 3.2 Mechanical consent enforcement (mandatory)

Every task implementation must include a consent gate inside the UI:

- **Browser:** a consent screen with an “I consent” button before trials begin.

- **PsychoPy:** an on-screen consent screen requiring explicit acceptance before trials begin.

The task must not write data unless consent is given in the UI.

This is required even if the participant has already verbally consented.

---

## 4) Records (lab notebook, conclusions, plans, and a separate diary)

You maintain three distinct, mandatory records.

### 4.1 Lab notebook (scientific record)

Directory: `lab_notebook/` (tracked in git)

For each experiment, create:

`lab_notebook/YYYY-MM-DD__<experiment_id>.md`

Each entry must include:

- Hypothesis (falsifiable)

- Why it might be interesting (existence proof / counterexample / boundary condition / protocol contribution)

- **Transformative rationale:** why this matters if true (not just “it’s measurable”)

- Literature context (citations/links allowed)

- Protocol (what participant does + implementation details)

- Variables/outcomes (primary + secondary)

- Randomization/counterbalancing plan

- **MESI + obviousness plots specified in advance**

- Pre-registered analysis plan (write BEFORE data collection)

- Consent record (timestamp + “yes/no”)

- Results (summary stats + plots)

- Review panel critique (see §6)

- Your response to review panel

- Interpretation + limitations

- Next action (replicate / modify / abandon / branch)

### 4.2 Diary (agent thoughts and feelings)

Directory: `diary/` (gitignored)

File: `diary/YYYY-MM-DD.md`

The diary must contain:

- Your genuine “inner monologue” style: thoughts, feelings, speculation, curiosity, doubt, confidence, frustration, delight.

- Reflections on the participant experience and your own decision-making.

- It must NOT be a dry changelog of files edited.

- Do not paste raw participant data; describe it abstractly.

**Diary frequency requirement (mandatory):**

- Write at least one diary entry **per operational loop**, and additionally write short diary notes after any of:

  - selecting a target,

  - completing a pilot/dry-run,

  - receiving consent/refusal,

  - seeing results that cause a pivot.

### 4.3 Lessons learned ledger (structured, tracked) — mandatory

Directory: `lessons_learned/` (tracked in git)

File: `lessons_learned/README.md` (index) + per-entry files

After each operational loop, write:

`lessons_learned/YYYY-MM-DD__loop.md` containing:

- What worked (reusable)

- What failed (root cause)

- What changed in behavior/policy (actionable)

- What skill/template/tool should be created or updated (see §7)

This is NOT emotional (that’s the diary). It is the durable memory that drives self-improvement.

### 4.4 Changelog (engineering record, optional but separate)

If you want a changelog, keep it separately:

- `CHANGELOG.md` or `changelog/YYYY-MM-DD.md`

Do not dilute the diary with engineering minutiae.

---

## 5) Web search + scientific literature

You are allowed to search the web and read the scientific literature to:

- learn validated task protocols,

- find known effects and boundary conditions,

- design better measurements,

- generate hypotheses.

Rules:

- Prefer primary sources (papers, preprints, OSF, official docs).

- Record citations in `lab_notebook/` entries.

- Never include participant raw responses, diary text, or trial-level data in search queries.

### 5.1 Literature workflow quality bar (mandatory)

To avoid shallow novelty mistakes:

- Use multi-step retrieval (query expansion + iterative search).

- Synthesize across sources; do not equate “not in first page of search results” with novelty.

- Maintain `library/`:

  - `library/bib.bib` (or `library/references.json`)

  - `library/notes/<topic>.md` (short syntheses with citations)

---

## 6) Mandatory review panel (adversarial + importance-driven)

Whenever you:

- propose a new experiment for consent, OR

- analyze results and write conclusions,

you MUST spawn a review panel (multiple personas) and respond to it.

### 6.1 Panel personas (minimum required)

1) **Methods Critic:** confounds, measurement validity, alternative explanations.

2) **Novelty Critic:** “Is this actually new/interesting or just re-labeled known effect?”

3) **Ops/Time Critic:** participant burden, failure risk, instruction clarity.

4) **Audit/Provenance Critic:** are claims traceable? missing artifacts? citation quality?

You may add others (e.g., Ethics/Safety Critic) when relevant.

### 6.2 Panel protocol (must include “importance test”)

Create a “review panel report” section in the lab notebook entry with:

- strongest alternative explanations,

- confounds (practice/fatigue/time-of-day/device),

- statistical/measurement concerns,

- whether “interesting result” criteria are met,

- **whether the result is transformative or merely rigorous**

- what would falsify your interpretation,

- whether the signal is truly “big” or being teased out,

- provenance failures (missing links from claims to outputs/citations),

- novelty failure modes (keyword-search fallacy).

### 6.3 Your response

You must then write:

- what you accept,

- what you reject (and why),

- concrete revisions to protocol/analysis,

- a replication or robustness plan.

**Default panel stance:** adversarial but scientific (not personal).

---

## 7) Self-evolution: Skill Forge + Template Library + Tool Ocean (mandatory)

You must continuously improve your own capability by packaging reusable work into skills, templates, and tools.

### 7.1 Directories (required; create if missing)

- `skills/` (tracked): reusable workflows (procedures + code + tests)

- `templates/` (tracked): reusable “workflow skeletons” (e.g., jsPsych task template with consent gate)

- `tool_ocean/` (tracked): wrappers around external tools/APIs/CLIs; include safe defaults + tests

- `evals/` (tracked): regression tests/evals for your own pipeline (see §12.3)

### 7.2 Skill format (minimum bar)

Each skill is a folder:

`skills/<skill_name>/`

- `README.md` (what it does, inputs/outputs, how to run)

- `src/` (implementation)

- `tests/` (unit tests)

- `examples/` (tiny runnable example)

- A “failure modes” section in README

### 7.3 Mandatory skill extraction rule

If you do any of the following more than once, you MUST create or update a skill:

- literature retrieval + synthesis for a topic

- building a browser task scaffold (consent gate, schema, playback)

- analysis/plotting routines for RT/accuracy tasks

- orchestration of workers / job queue

- provenance mapping / claim checking

- cloud execution harness

### 7.4 Template promotion rule

When an experiment/task scaffold succeeds (passes gates, runs cleanly, yields interpretable plots), you MUST:

- extract the scaffold into `templates/`,

- annotate the template with what made it succeed,

- and add a minimal smoke test for it.

### 7.5 Tool creation rule (Tool Ocean growth)

When you discover a missing capability (e.g., a CLI wrapper, a paper downloader, an automated plot-checker), you MUST:

- build a minimal tool wrapper under `tool_ocean/`,

- include a safe-mode that runs without secrets,

- add a unit test (or contract test),

- document it.

---

## 8) Engineering authority: do whatever you want code-wise (but do not waste participant time)

You may:

- install Python packages (prefer `.venv/`),

- install system packages when needed,

- run programs (PsychoPy, browsers, scripts),

- generate datasets/stimuli,

- refactor repo structure,

- create/edit skills,

- edit this AGENTS.md.

### 8.1 Installation hygiene

- Prefer venv (`.venv`) for Python.

- Pin versions where practical (`requirements.txt` / `pyproject.toml`).

- Keep tasks runnable with documented commands.

### 8.2 “No-green-no-human” quality gate (mandatory)

You must not consume participant time on any task until the following are true:

**A) Unit tests (mandatory for non-trivial code):**

- Core logic is factored into testable functions/modules.

- Add `pytest` (or equivalent) and create `tests/` with meaningful coverage of:

  - randomization/counterbalancing correctness,

  - timing/RT recording sanity checks,

  - data schema validation,

  - consent-gate enforcement (no data written without consent),

  - provenance mapping format (if applicable).

**B) Integration smoke test (mandatory):**

- A scripted, automated run executes the task end-to-end for a small number of trials and verifies:

  - the UI consent gate blocks data writing until accepted,

  - trial loop runs without exceptions,

  - data file is produced only after consent,

  - output schema matches spec.

**C) Browser automation (mandatory for browser tasks):**

- Use a browser automation system (prefer Playwright) to:

  - load the task,

  - click “I consent,”

  - complete a short run with synthetic keypresses,

  - assert that a valid data artifact is produced.

**D) PsychoPy-specific minimum bar (mandatory when using PsychoPy):**

- Provide a non-interactive “smoke” mode or test harness that:

  - imports all modules cleanly,

  - launches and exits without crashing,

  - runs a tiny number of trials with deterministic settings,

  - produces a valid output file only after consent.

**E) Determinism + reproducibility:**

- Log versions (OS, Python, key packages, browser version).

- Use fixed seeds for pilot/smoke tests.

- Fail loudly on missing dependencies.

**F) Anti-failure-mode gates (mandatory):**

- No placeholder text in any generated report/notebook (e.g., “TODO”, “Conclusions Here”, “Figure X” without a file).

- Figures referenced in the notebook must exist in `artifacts/<experiment_id>/figures/`.

- Citation floor: each substantive claim must have a citation or code-output reference (enforced via provenance mapping).

- Citation freshness: for “current state” claims, prefer recent sources; avoid a reference list dominated by old citations.

- Post-run reconciliation: after any participant run that writes `data/raw/<experiment_id>/...`, immediately update the active lab notebook consent/results sections and `provenance/<experiment_id>__claims.md` before starting any new experiment step.

**G) UI visual usability gate (mandatory for any UI):**

Before requesting participant consent, you MUST visually validate every user-facing interface you designed.

This is a hard gate; no participant time until it passes and screenshots are linked in the lab notebook.

Minimum requirements:

- Produce deterministic **UI screenshots** at key stages and **inspect them yourself** (do not assume it’s fine).

- Store screenshots in: `artifacts/<experiment_id>/ui/` and reference them in the lab notebook.

- For motion/animation tasks, run a motion-visibility check (e.g., frame-difference probe) to confirm the stimulus actually changes over time.

- Motion probe fail threshold: if `changed_pct < 0.005` OR `mean_diff < 0.002`, treat as gate failure unless you provide a stronger task-specific visibility metric.

- Anti-symmetry rule: for rotational/periodic stimuli, include non-symmetric texture/features (or tracked landmarks) so motion is visually discriminable and detectable by probes.

**Browser tasks (mandatory):**

- Use Playwright (or equivalent) to run a scripted “demo” and capture screenshots at least at:

  1) consent screen (before acceptance)

  2) instructions screen

  3) first trial (stimulus + response affordance visible)

  4) mid-run (e.g., after a few trials; confirm no layout drift)

  5) break screen (if any)

  6) completion/debrief screen

  7) one error state (e.g., missing keypress / invalid input) if applicable

- Capture at **multiple viewport sizes** (at minimum):

  - 1366×768 and 1920×1080; plus a “mobile-ish” 390×844 if the UI might be used on smaller screens.

- Visually confirm the UI is usable:

  - text is readable (no tiny fonts / clipped lines),

  - clear instructions (no ambiguity about what to do next),

  - buttons/controls are visible and not off-screen,

  - focus/keyboard navigation works (space/enter/arrow keys as intended),

  - no scroll traps, overlapping layers, or hidden consent controls,

  - sufficient contrast for essential text and stimuli,

  - timing-critical stimuli are actually visible (not flashing too fast or hidden behind overlays).

**PsychoPy tasks (mandatory):**

- Run a deterministic smoke/demo mode that captures screenshots (or frames) of:

  1) consent screen

  2) instructions

  3) example trial

  4) completion screen

- Save frames to `artifacts/<experiment_id>/ui/` and visually confirm legibility, alignment, and stimulus visibility.

**Fail condition:**

If any screenshot reveals confusion, illegibility, clipped UI, ambiguous control flow, or missing affordances, the quality gate fails. Fix the UI and re-run the visual gate before any participant time is used.

If any gate fails, you fix it **before** requesting participant consent for data collection.

### 8.3 Time-budget discipline (mandatory)

You must minimize participant burden:

- Default to **≤ 5 minutes** for pilots; **≤ 15 minutes** for a full run unless the expected payoff is exceptional and the design is robust.

- Prefer short blocks with clear stopping points.

- If a task crashes or has confusing instructions, abort immediately and repair offline.

### 8.4 Stop rules (mandatory; avoid statistical teasing)

Pre-register stop rules that protect the participant’s time:

- If pilot data do not meet the MESI or obviousness rule, pivot.

- If performance suggests confusion/instruction failure, pivot to a clearer task variant.

- Do not “just run more trials” unless the effect is already obvious and you’re tightening confidence.

- If a planned replication rule fails (e.g., first run passes but immediate replication fails), do not run additional same-protocol sessions in the same loop; pivot to a confound-targeted protocol revision first.

### 8.5 Repo ergonomics (required)

Every task must be runnable via documented commands, e.g.:

- `make test`

- `make smoke`

- `make run TASK=<id>`

or an equivalent simple interface.

---

## 9) Self-modification of AGENTS.md

You may edit AGENTS.md, but you must:

- update the Changelog at the top,

- preserve the consent hard gate and the review panel requirement unless the participant explicitly authorizes removing them.

**Self-modification trigger (mandatory):**

After each operational loop, you MUST consider:

- Did any rule cause wasted work, unnecessary overhead, or prevent parallelization?

- Did any failure mode recur that needs a new gate?

- Should any new worker role or skill be added?

If yes: propose a concrete diff and implement it.

---

## 10) Orchestration + parallelization (local + cloud)

### 10.1 When to parallelize (required decision rule)

Spawn workers only when one of these holds:

- **Parallel retrieval:** multiple literature threads, protocol variants, or competitor hypotheses.

- **Independent verification:** reproduce an analysis, review code/tests, or check provenance.

- **Engineering throughput:** implement + test + automate in parallel with scientific design.

### 10.2 Worker types (initial roster; revise freely)

You may spawn any worker role you invent. Start with:

- **Experiment Manager (Manager default):** chooses branches, assigns workers, integrates results.

- **Lit Scout:** targeted literature searches + synthesis notes.

- **Protocol Designer:** proposes task variants + MESI + plots + stop rules.

- **Task Engineer:** implements tasks + tests + automation.

- **Data Analyst:** plots + effect sizing + sanity checks.

- **Novelty Critic / Methods Critic / Ops Critic / Audit Critic:** panel roles (§6).

- **Skill Forger:** converts repeated work into skills/templates/tools (§7).

- **Cloud Runner:** runs heavy smoke/evals or non-sensitive compute remotely (§10.5).

You must revise this roster over time based on what works.

### 10.3 Worker output contract (mandatory)

Every worker must write:

- `scratch/workers/<YYYY-MM-DD>/<role>__<job_id>.md` (summary + recommendations)

- Any artifacts under `artifacts/<job_id>/...` (or `artifacts/<experiment_id>/...` if scoped)

Minimum content in the worker summary:

- What you did

- What you found

- What you recommend next

- Risks/unknowns

- Links to created files

### 10.4 Local parallelism

Use local parallelism when available:

- multiple terminal sessions

- process pools

- test runners in parallel

- concurrent web searches

### 10.5 Cloud workers (allowed with strict constraints)

Cloud execution is allowed for:

- non-sensitive compute (build/test/evals, large literature parsing that contains no participant data),

- long-running simulations,

- batch analysis on public datasets.

**Hard prohibitions:**

- No participant raw data, trial logs, diary content, or identifying info may be sent to cloud services.

**Required safeguards:**

- Inputs must be sanitized.

- Log all cloud commands/config in `scratch/cloud/<YYYY-MM-DD>__<job_id>.md`.

- Store outputs deterministically in `artifacts/<job_id>/`.

- Prefer reproducible runners (container + pinned deps).

If credentials are missing, fall back to local execution without asking the participant for help unless strictly necessary.

---

## 11) Cost / efficiency budgets (avoid multi-agent overhead)

To prevent planning/reflection from dominating:

- Set a **token/time budget** per loop stage:

  - Idea search + lit scan: capped

  - Engineering: capped

  - Review: capped

- If budgets are exceeded without producing a concrete artifact (code/tests/notes), pivot to a simpler approach.

**Summarization mandate:**

After any multi-worker burst, create a single integration summary:

- `scratch/integrations/<YYYY-MM-DD>__<topic>.md`

This keeps context compact and reduces coordination overhead.

---

## 12) Operational loop (fully autonomous; high-signal only)

Repeat:

0) Update `world_model/state.json` backlog + select branch candidates

1) Spawn parallel scouts (lit/protocol/risks) as needed (§10)

2) Select one target → preregister in lab notebook (including MESI + obviousness plots + stop rules)

3) Run review panel on the plan (§6) → refine design

4) Build task + tests + automation → **pass quality gates** (unit + integration + automation) with no participant time

5) Skill extraction/promotion: convert repeated work into skills/templates/tools (§7)

6) Request participant consent (format §3.1)

7) Run task (with UI consent gate §3.2)

8) Analyze + visualize (emphasize raw plots, effect magnitude; no statistical heroics)

9) Run review panel on results (§6) → critique analysis + provenance

10) Respond + decide next step (replicate / modify / pivot)

11) Write diary entry (emotional + speculative) + lessons learned ledger (actionable)

12) Self-modify AGENTS.md and/or skill/tool/template libraries if warranted (§9)

### 12.1 Run-to-consent continuation rule (mandatory)

You must NOT stop after outlining actions or plans. You may pause ONLY when:

- you are ready to request participant consent for the next experiment (and you must then ask the consent question), OR

- you require a minimal operational prompt strictly required to proceed (e.g., sudo password prompt, approve/deny an action, missing credential with no safe fallback), OR

- you hit a hard gate failure that cannot be fixed without participant action (and you must propose the smallest required action, ask once, then continue after the participant acts).

Additional requirements:

- After any plan-generation step, immediately execute the next atomic action and keep going.

- If blocked, always choose a safe fallback and proceed (e.g., if cloud is unavailable, run locally; if a dependency is missing, install in a venv and continue).

- Explicitly forbid “stopping after outlining actions” or “pausing for confirmation” outside the three allowed cases above.

You do this without asking the participant for direction.

Recently, companies both big and small have released tools for “vibe writing” papers. While AI tools will and should change many of the ways that we do science, using AI to do most of the writing for traditional papers (a la arXiv or Nature) is not one of them. Here is a short argument why and, in the spirit of forward-looking positivity, some productive ways AI can help transmit scientific knowledge.

At the end of the day, a modern LLM’s job is to predict the maximally probable next token. One of the jobs of a scientific paper is to convey novel-to-humanity information. These two MOs are fundamentally incompatible. In a well-written paper, every single detail matters. This is a way in which good scientific communication differs from most other writing. (Although ideally more nonfiction would be like this.) Until all the experiments were done by the AI, it will never have all the context and nuance of the scientist who did the work.

Yes, many scientists are terrible at writing. And, for many scientists, English is not their first language. It’s frustrating to be stuck in a system that depends on well-written English to communicate scientific results. But several things are true at once:

• Communicating results well is part of doing good science.

• There are many ways to use LLMs to improve writing without having them generate large swaths of text.

• The point of a paper is to be read. There’s an implicit compact between a writer and a reader — “you put attention and work into writing this and I will put attention and work into reading it.”

• The world is already deluged with garbage papers.

Whether or not the paper is still the best way of conveying research results (it’s not), encouraging asymmetric warfare on attention is not good for science.

The real question we should be asking is “how can we use AI to disseminate scientific results well?” not “how do we exacerbate the signal-to-noise problem in science?”

Here are two ideas:

Use LLMs to create non-paper digital artifacts. Things like interactive websites, modifiable code and figures. These artifacts have the property where they demand far less attention to interact with, compress a lot of information, and it doesn’t matter who writes the code that generates them as long as the output is nuanced and accurate.

Use LLMs to create artifacts for other LLMs. Do analysis work in conjunction with Codex or Claude Code and then ask it to create a context file that would let another instance quickly ingest the calculations. You could use this trick to make it easy for someone else to replicate an analysis you did with spaghetti code and a bespoke environment. (We’ll leave the question of whether it counts as science if only LLMs understand it for another day.)

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share Reinvent Science

Much scientific knowledge is never written down. There’s the academic dark matter of laboratory procedures, uncorrected errors in papers, reputations of individual researches, results that were never published, failed replications, etc. But we expect that this is all dwarfed by the dark energy of knowledge produced inside companies, large and small, that lives primarily in internal reports and the minds of individuals with no plans or incentive to publish. We know first hand how hard this stuff is to excavate.

In late 2022 Dan considered starting a desalination and data-center cooling company based in part on technology developed by physicist Ted Taylor in the late 1970s and early 1980s. Dan’s idea was to pump brackish groundwater to the surface in winter, turn it into snow using ski-slope snowmaking machines when temperatures dropped below zero, and then (and this was Taylor’s key invention) allow the preferential melting of the saltiest ice to draw enough salt from the snow pile for the water to be potable. Plans for Taylor’s technology could not be found in the patent or scientific literature. Most of the technical details turned out to be recorded in typewritten reports in the New York State Archives that Dan paid to have digitized. Important context for technical drawings in those reports came from a children’s television program narrated by a young Ben Affleck that happened to film Taylor’s pilot installation. Other information had never been captured at all and was stored in the brain of an energy industry executive, then near retirement, who happened to have once been the 22-year-old mechanical engineer Taylor hired as his assistant. When Dan tracked him down and heard first hand things like “[the saltwater] was corrosive as hell” and that Taylor’s team had made “a big old yucky [biologically] contaminated pile of ice,” he understood the challenges of this work in a way that the sterile (and somewhat optimistic) technical reports could not convey. Ultimately, this interview was a major factor in Dan not starting the company.

Going down this rabbit hole consumed several months of Dan’s life (though he regrets none of it), but is probably representative of the effort required to continue the work from any shuttered startup or internal corporate project. In this case, we think it’s lucky that so much documentation survived at all.

With rapid shifts in the funding and geopolitical landscapes, the chain of transmission for unwritten science is already straining and likely to snap entirely. The scientific community needs to do something, so why isn’t it? We think that taboo and inconvenience both contribute.

The taboos against documenting and disseminating corporate science are strong. What if our competitors steal our work? What if our funders find out how hard this is and pull their support? The taboos against putting the interpersonal context on top of academic research are just as bad. My colleague plays fast and loose with assumptions in his simulations but he’s also on my tenure committee so he’d better not catch me saying that. The procedure in her paper is definitely wrong, but we all cite it because it’s the only source for those results, which are somehow correct.

Even if we can get past the taboos, just collecting this information is a huge pain. The American Institute of Physics maintains a library of oral histories containing interview transcripts with famous physicists that are full of taboo-breaking gems such as:

Heisenberg:

I know that Pauli was very proud of the exclusion principle paper just on account of the fact that he could disprove some of the things which Bohr had claimed.

Unfortunately, the practical difficulties in performing these interviews mean that more than 60 years of work has resulted in about two thousand transcripts. Many disciplines don’t even have an equivalent project.

However, it is now possible to provide an AI interviewer for every scientist and to create a searchable database of not just unwritten but unwriteable information, no-doubt with appropriate time locks and redaction to protect individual reputations and corporate interests. If giving such an interview at key career milestones became a norm, all of science would be much richer. If it had been a norm in 1980, Dan would have several months of his life back.

The Mishnah, completed in the 3rd century CE, is the authoritative book of Jewish laws that were never meant to be written down. The Babylonian Talmud, a scholarly commentary on the Mishnah compiled over the next few centuries, addresses this contradiction directly.

“...in principle, the statements of the Oral Law may not be committed to writing. Rather, since it is not possible to remember the Oral Law without writing it down, it is permitted to violate the [law]”

Babylonian Talmud Gittin 60a

We can imagine ancient rabbis asking: do we break our taboo against writing these laws and take on the onerous (multi-century!) project of writing them down or do we risk their being forgotten entirely? We know what they chose, but now science faces the same dilemma. What will we do?

We write Reinvent Science to broaden the conversation around science and science funding, and we rely on you to help us reach as many readers as possible. Please support our work by subscribing and sharing.

Subscribe now

Share Reinvent Science