Science Is Open Software
jepedersen.dk
jepedersen.dk
At least some journals have this as a stipulation e.g. https://www.nature.com/nature-portfolio/editorial-policies/r...
Particularly the "data availability" and "Availability and peer review of computer code and algorithm".
However, in my limited experience, of trying to reproduce certain scRNA-seq processing pipelines, in practice it's never available as just a Github link. I can understand that some/many researcher's code is not in good shape, so I think it'll be quite a stretch to have this available.
I do think it's laudable though, to try and make it available. It would certainly have been very useful for me in the past.
And with that said, for experimental work, this approach does not make things fully reproducible; it only makes the analysis reproducible. There are always factors that influence experiments: research is by definition at the edge of our understanding, and reality has countless variables, including ones no one has thought of, known about or thought important.
I see a lot of reliance on containers to make code always available, and I have the same misgivings as you do. It'll work a few years into the future, but what happens once packages aren't compatible with each other/the base container is upgraded/etc.
I've already seen this with older bioinformatics code, which is on old repositories that aren't running anymore, or are very unreliable (but weren't at the time that the code was written). And I'm talking about code that's "only" 15 years old; people will be going to these papers for implementation details long after that point.
Zenodo seems like a good step in the right direction. It should be available as long as CERN is going, shouldn't it? And by that stage it should be "too big to fail".
Another problem with reproducibility is the openness of the underlying data. Many academics are terrified of giving away the golden goose and the software is often useless without the data.
However many scientists do work openly, e.g. The Journal of Open Source Software:
I believe they’re always looking for new volunteers to review papers, so please do volunteer if you are able.
I’d suggest data management needs to be a well paid career in academia, with senior level influence to enable long term open access to data.
I was on a grant from the Wellcome Trust putting an epidemiology database online. This remained available until the PI moved on and their successor was decidedly less public spirited, so sadly it’s not available anymore. What can I do about that when I’m seen as a technician? I have over 8000 citations from my database work but there’s really no career in academia for me.
But until much of academia is burned to the ground, or until science can be properly separated from modern academia, this will never be so. The current academic incentives are all wrong: low-quality research is rewarded and results in publications, whereas high-quality research (that takes time, and usually reveals that most exciting publications depend on p-hacking or other highly data-dependent analyses and selective presentations) is not published or actively blocked during peer review.
So instead you get BS arguments about how data can't be released for various privacy concerns (when in reality the vast majority of most datasets are trivial to scrub of identifying factors, and even in more complex datasets where you need to consider k-anonymity, it is still trivial to release data that allows replication of core analyses), and academic science is increasingly irrelevant unless it is tied to tech and industry, where producing junk actually has real negative economic and personal consequences.
I don't know what world this article / post lives in, but it isn't the messy world of actual reality.
All of these tens of thousands of man-hours could be replaced by a few GitHub repos, if only RCTs would just publish their damn data. Then you could just run and rerun the statistics on whatever subgroup you're looking for, instead of combing through decades of review articles answering slightly different questions, looking for the answer between the lines. With LLMs making mining of large scale datasets almost trivial (with the process most likely becoming trustworthy within a few years), the current status quo is looking more and more antiquated.
If you want to be even more radical, hospitals could just publish their data continuously. Of course, it is easy to point to the risks of doing so, but what's often ignored is the benefits. It is hard to overstate just how many medical mysteries a hospital encounters on a daily basis, how much unknown we are navigating in practice. The current norm is that 99.99% of these cases are never published, and are only ever thought about by a small group of people who happened to be at work. Particularly, when someone dies of something no one figured out, it is never published anywhere, because even if you tried it is not interesting reading material for a journal to publish. And no one ever tries because they're scared of being called out for a mistake. A hospital is essentially a continuously running and extremely interesting experiment, where 99.99999% of all results are thrown in the garbage, and the only published data is subject to extreme selection bias.
All of this could be different, and the risks involved are actually quite small in practice. It is easy to automatically anonymize data quite well, but extremely difficult to absolutely guarantee that it is anonymous. And since current ethical norms are extremely averse to any degree of risk, and usually entirely ignore potential benefits, we all suffer for it. It is not entirely unlikely that someone reading this post will one day die because of something that could have been prevented, had things been different.
This has all been profoundly obvious for at least well over a decade or even two now. A consequence has been that too many serious people are driven away from academia and research, to the detriment of science generally.
I've no idea what to do about all this, because people have voiced obvious and easy solutions for decades, but they are all routinely ignored.
Most of these things are easy to do for the majority of papers, especially in the past 20 years with the internet and modern tech and software. Plenty of frameworks exist already that have done most and/or at least some of these things, but, collectively, academia is decades behind overall.
I personally am trying to float using local LLMs to create anonymized case files and auto-suggest publishing cases in my hospital, which knowing how things work will probably never amount to anything.
Or if you want to float truly insane ideas I guess you can shop around with blackhat groups and see if anyone has stolen some juicy records/data during all the ransomware attacks and databreaches over the years, and do some rogue scientific publishing. Obviously that's crazy, but I have to admit that the notion of pirate scientists plundering and publishing data is hilarious to me.
Everything can be prevented given sufficient knowledge. That's not the point. The point is how to prevent as much as possible.
In the 27 years I've been in academia, I've seen a lot of progress in data openness (NCBI GEO was a game-changer) and FOSS analysis software (it's now widely expected that a high impact pub will make all data and code available for review, and then publicly available upon manuscript publication; most major journals will not allow submission without this). It is becoming common for big journals to specifically ask reviewers to review the analysis code. It is starting to become more and more common for papers to release all the code used to generate all the figures (including supplementary figures)
There is still a long way to go, I agree. But it's always better to light candles than curse darkness, etc.
> academic science is increasingly irrelevant unless it is tied to tech and industry
While I have some sympathy with a lot of your bitterness, this statement is insulting silliness that a quick look at the list of Nobel Prizes in physiology and medicine would prove wrong. Almost all major breakthroughs in the applied sphere stem from decades of basic research that happened just because it interested someone.
I am also in academia and regardless, factually this is not true at all for data, not even remotely (less than like 10% of journals even have data availability policies which are recommendations, and in practice only a small percentage of papers actually make anything available), unless by "publicly available" you mean "available to some academics or academic labs after an often tedious and slow approval process requiring an academic email and various signed agreements". Maybe what you are saying is true in some very specific domains (e.g. machine learning research), but in general what you are saying here is IMO wildly out of touch with present realities in the vast majority of fields, but especially those involving human subjects.
> While I have some sympathy with a lot of your bitterness, this statement is insulting silliness that a quick look at the list of Nobel Prizes in physiology and medicine would prove wrong. Almost all major breakthroughs in the applied sphere stem from decades of basic research that happened just because it interested someone.
Nobel Prizes are so rare they don't speak at all to the generalizations I am making here. Also, much medical academic research is arguably successful because it is in fact ultimately industry-funded or tied to industry. It is of course though highly dependent on the academic subfield, for sure, and I was painting with a broad brush.
If I had to narrow things, STEM academic research isn't so bad, so long as we exclude social science from STEM. Much social science research needs to be defunded ASAP. And I'm not claiming industry research doesn't also have warped incentives. But, on balance, I'd wager outside of pure math/physics and certain more algorithmic/pure domains in comp sci, the smartest people today are going to choose (and be found in) industry, not academia.
The important part here is, as you say, to light candles and insist on rigor. Coincidentally, history tells us that that also gets us further. So by pure memetic selection, this strategy should win
Your analysis completely ignores the physical and biological sciences, engineering, and the humanities. It mostly applies to a small subset of academic fields in the social sciences and medicine. You're also ignoring the changes that have happened since the replication crisis. Preregistration, publishing all code and data, reporting null findings, replicating results, etc. are becoming the norm.
Yes, for the most part the fixes have to be in terms of funding and incentives. Funding needs to be more careful, and more careful funding can be a carrot rather than a stick here.
Re: incentives, IMO we clearly need a stick: there need to be harsh negative consequences for engaging in degenerate research programs and methods that have clearly been shown to result in pathological or cargo-cult science. Null-hypothesis significance testing is one clear practice that needs to go, but building entire fields on phony / meaningless uncalibrated metrics (think: a lot of self-report instruments that are never properly calibrated to objective outcomes or real-world behaviours and/or consequences, with results being reported only as standardized effect sizes) are another more pernicious practice permeating far too many fields. Ideological bias also needs to have funding consequences. Replication issues are still only surface problems in many fields, where the research would still all be worthless even if it replicated 100% perfectly.
> Your analysis completely ignores the physical and biological sciences and the humanities
I admitted later to painting with a broad brush, and yes, it is always hard to generalize and cover everything fairly. But IMO humanities has serious ideological and methodological rigor problems as well, and is overdue for disciplining. I would tend to have stronger positive feelings toward the biological sciences generally, yes. Yes, the social sciences are the major source of the problem (in part because they are so bad they tarnish the reputation of all academia).
> These fields have gotten a lot better over the past decade in the wake of the replication crisis. Preregistration, publishing all code and data, reporting null findings, replicating results, etc. are becoming the norm.
IMO "a lot better" is subjective, and I don't see those things as being the norm yet (beyond as lip-service), and the rate is far too slow. I agree we'll get there eventually, but I am worried about the loss of public trust and thus the production of real knowledge if we don't try a bit harder at this. Plus, globally, countries like China do seem to be more willing to actively crack down on research misconduct, at least in the past years, and it might not be unrelated to them increasingly pulling ahead technologically in many areas.
> tech and industry, where producing junk actually has real negative economic and personal consequences
In academia, you can just endlessly produce low-quality garbage, and basically make a career out of this. In industry, things more often eventually at least have to work and survive contact with reality. Academia mostly lacks this basic check.
The scientific method should be the north star, sure. Much of what is happening in academia is cargo-cult / degenerate / pathological science though.
The point at the end "The scientific revolution succeeded because it insisted on transparency, reproducibility, and constant scrutiny." is a good one. But the real kicker that makes the situation so bleak is it's not the scientists who get to decide whether or not these things get applied to science or not, it's government policies and corporate interests.
I don't know if this is true. Lots of this stuff is inherently expensive - clinical trials, research into vast numbers of compounds, and then mass-producing the drug are all things that cannot be feasibly done without at least millions of dollars.
The US and China are the current world leaders in biotech, and it's because both of them have massive infrastructure to funnel billions of dollars into research.
Yeah, the IP law could be reformed (I am a big believer in reducing IP protections in general) but the truth is that SOME FORM of protection is necessary to convince people with money to fund this kind of research. The government is simply not capable of this level of spending for such uncertain rewards; it doesn't have the proper incentives to recognize and promote good research while defunding useless research.
> Please don't comment on whether someone read an article. "Did you even read the article? It mentions that" can be shortened to "The article mentions that".
> TL;DR I claim that modern science is synonymous with open source software
That's a strong statement that's not supported by the arguments and IMO misguided.
I don't have a problem with "open", but rather with "software".
Both science and software deal with models, however the focus is quite different. I suspect you conflate theory with models.
The goal of science is to produce and test theories — that's an inductive/abductive process. A model, regardless of whether it's reified into mathematical formulas or software, is a means of making a theory operational enough that its consequences can be derived and confronted with observations.
Software often starts downstream of this: it's a reification of theories, models, algorithms, or findings that are the result of research. Of course software can also be used as part of the research process itself. The distinction is roughly the familiar one between research and development.
If I'm reading yiur argument right, you're saying that deployed models are downstream versions (reified) of aa theory. Theory, being the actual object of science.
I think this misrepresents science. Science is the ability to build testable knowledge. From that,how would you separate the test from the science? In fact, in an ideal world, why wouldn't you want your theory to be put in a format that's executable? I'm not saying that those things are always the same, my (provocative) title is based on a dream where we can imagine theory and model coexist because software is now a thing.
Software is a static program and basically none of these things.
Software development is kind of like science, in some ways, in that you discover abstractions and patterns, and this requires resolving disagreements and ambiguity between you and your users, but in the end, the user demands are usually fairly concrete and specific (though no one may know how to express those demands precisely, initially), and the process is not really exploratory in the way science is.
It just really isn't a very good comparison IMO.
More dramatically stated as: Science progresses one funeral at a time.
> An important scientific innovation rarely makes its way by gradually winning over and converting its opponents. What does happen is that its opponents gradually die out, and that the growing generation is familiarized with the ideas from the beginning.
Disclosure: Old scientist.
I would also contest the point about science being social. To me, science is the interplay between social constructions and physical reality. Which is why tests and experiments are so fundamental. Your point about software development as kind of like science is spot on: why not be exploratory/empirical in software development? Couldn't we imagine an LLM, say, that would go out and try to solve a question we frame? This seems strikingly close to autoresearch [1]
This is false in far more cases than the cases where it is true, and even in the cases where you are right, you still usually want similar behaviour.
> You need inputs and, as demonstrated by the recent LLM inputs, parameters you can tune to correctly address a specific context
You seem very confused about what "static" means here.
> I would also contest the point about science being social. To me, science is the interplay between social constructions and physical reality.
Well, almost all philosophers of science would disagree with you, and IMO this sentence immediately contradicts itself.
Frankly, you should really work on learning to write more coherently and carefully. You maybe have some good ideas, but they are being communicated extremely poorly and inconsistently.
I am particularly fascinated by labs like DeepMind [https://deepmind.google/science/]. The recent advances in their frontier models that are able to predict diseases before they're diagnosed is incredible. This is what AI should be built for and actually do!
1. Reproducible, meaning executable, as well as modifiable, and 2. Reliable, meaning that the results are consistently trustworthy
Why does modifiable fit reproducible. Modifiable fits the open source theme, but not the science theme. Can someone set me straight on this if I am missing something?
Number 2 is something I've been thinking about more recently, and the definition of science from the post being "testable hypothesis and predictable".
Medicine is testable and predictable, but not reliable. To me reliability is the job of engineering. To take what is known in science and make it reliable.
I used to work at CSIRO (Australia's science and technology org) and I remember my boss telling me that researchers always think their work is ready to go from the lab into the real world, but it is our job as engineers to make that translation so that the work outside of the lab. That's what I think about when I think of reliable.
Thoughts?
A useful way to think about this is to consider existing human languages like English or mathematical notation. When you read these English words, it's like you're executing a program I've written that will change your brain state. A pattern in my brain is encoded, stored, and transmitted in a language that we've both learned in our own separate ways, then decoded to a pattern in your brain. But your brain is quite different from mine, and you're free to implement any sort of sandboxing you want against this information, and even against the language specification itself, in your own head. Erasing this barrier is equivalent to erasing individuality, so I think we ought to be able to do this for digital programs as well.
I hope that in the not too distant future, we will be able to design and manufacture custom hardware on a per-person basis. Because that's what it will take to preserve human autonomy and bodily integrity in the upcoming era of AI and brain augmentations. I don't want my neural interface to have a hardware-level backdoor (https://en.wikipedia.org/wiki/Intel_Management_Engine) like my desktop computer. There is also a bigger evolutionary problem behind this, and I expand more on that in this tangentially related past comment: https://news.ycombinator.com/item?id=49690354
I like how this is dropped as a fact. Dare to explain why though? Especially vs Docker. NixOS is not even standardized. No guarantees it will not be superseded by some descedant or eg. Guix in a near decade.
Best: I still argue that Docker and Conda are more accessible, but my point was to go for reproducible, declarative science. Nix environments are exactly that. They're not perfect and are, as you point out, not standardized. But they cover much more ground that Docker. If you trust the upstream nix repo, you can get bit-level equality at every single build you (or anyone else) does. Docker relies on huge binary blobs that you can't inspect and that can be pretty much arbitrarily swapped around.
I'm totally fine if Guix takes over. Or the next big thing. As long as it's declarative and reproducible.
The Turing Way seems great and I'm all for education. One of the most interesting conversation topics in this thread is, to me, how to create the necessary incentive structures. Know how is only part of the way. We need to secure the credit assignment for "openness" both in academia and industry.
Thus there is indeed a difference in the monetary intensives.
But why does the public have to pay for e. g. Elsevier? We pay for research of scientists already via taxpayers money (at the least in a civilized country), then we have to pay again for a private entity. If science is really open then it also needs to require public publishing. Gangsters such as Elsevier and others should not be able to drain the public here. Taxpayers financing something should also require public access to findings, at all times. Instead, Elsevier, Springer etc... get more public money while keeping things private. That's the antithesis to science.
I refer to what I think the author is asking for, as "push button reproducibility," i.e., the idea that the results will reproduce themselves at the push of a button, anywhere, at any time in the future. I have a couple of misgivings about this.
First, the whole idea of "open" research predates computer technology. Forcing science to keep up with the latest ideas in software distribution is too much of a burden, when science is already too risky and slow. I had the odd privilege of learning the scientific method from my mom, before there was widespread access to computers. Her version was that a study should be reproducible by a reasonably skilled person. This is a greatly relaxed standard, but is realistic for a discipline that spans decades if not centuries.
I supplied all of the data and code for my thesis research (and a sufficient number of mechanical and electrical drawings). But nobody has Turbo Pascal today, and some of the commercial instruments such as specialized lasers were already obsolete by the time I finished. Also, the experiment was dangerous, and might not pass safety review today. It required about $500k of equipment and a dedicated lab.
Today I have the luxury of saying that if my code fails upon loading a new version of a dependency, the person who discovers that failure is probably skilled enough to fix it, and my work rarely hinges on the idiosyncracies of dependency versions. If it goes into a product, they'll totally rewrite it anyway.
Second, science is still at its core an experimental discipline. Even in physics, there are more experimentalists than theoreticians. Reproducibility means roll up your shirt sleeves and head for the lab. To this day, some processes have not been mechanized, and you still need to spend years developing "lab hands" which not all people succeed at.
I think there's a clue in the fact that the social and medical sciences seem to be the most deeply embroiled in the reproducibility crisis. It's because the quality of results depends on the the quality of measurements, and it's just harder when dealing with living subjects or one-of-a-kind specimens (such as the earth's climate). In fact, not much more than a century ago, it was believed that studying those things was beyond the reach of scientific methodology.
Third, we're not going to stop doing science in areas where it's hard, particularly in medicine, but we're also not going to staff up in areas such as software development, to make science work better. People are suffering from disease right now so there's always an urgency to finding cures, plus an obvious profit motive.
Disclosure: Experimental physicist, developing better measurement equipment.
One thing I'd like to surface is the distinction you're making between software and physics. You point out that social and medical sciences are only now becoming approachable with scientific methodology. We can now, to a higher degree than before at least, model and predict what'll happen in social and medical scenarios. Isn't that an excellent example of how subject matters can be virtualized? I can't find good arguments for why this would stop there; why wouldn't our scientific models improve to such a degree that we can predict and "interfacte" with more and more of reality. Put plainly: why can't experiments be done in code one day? Particularly if we imagine having better robotics.
Do you see my point about closing the gap between the theorizing and experimenting? Or are they fundamentally different things? The "reasonably skilled person" seems like an excellent heuristic > 2 years ago that's rapidly being replaced by automated thinking machines.
The distinction between theory and experiment is somewhat arbitrary. Historically, scientists were better at one than the other, with a few rare people being known for both. To make an analogy, there's also no immovable boundary between hardware and software, since most hardware runs on code, but these tend to be different jobs and even departments.
What tends to happen in science is that when a subject is virtualized, e.g., when it can be modeled by a satisfactory theory, scientists move on to new frontiers. Many of the most interesting areas of work are where both theory and experiment can challenge one another.
I've had managers and even engineers ask me in all sincerity why we still need to make measurements when we can just model the data.
Am I the only one that's tired of "my random shower thought turned into full article with AI" articles?
Or like a machine learning lab claiming SOTA on a benchmark, beating a well-known method that they re-implemented, possibly with bugs, on their private dataset, for millions of compute. But you don't get the source to check, and they don't release any intermediate results or ablation experiments. Aka, from the outside you can't distinguish it from corporate marketing.
I won't put any more time into this ... good luck in figuring out what it is you really want to claim and presenting a coherent and cogent argument for it.