I haven't, but I'm unclear why it's relevant given that you don't seem to really be disagreeing!
The NSF denies 80% of grant applications because there is a radical oversupply of people who want to be scientists and the NSF has a finite budget. That by itself doesn't create accountability any more than the fact that lots of people want to be movie stars creates accountability for actors. That's not how accountability works.
Accountability means people being held to account for illegitimate acts. If it existed it would look like this: we (the government) gave you money to deliver some genuine research, yet you delivered a paper that simply modelled your own beliefs, cited a retracted paper and cited another paper that actually disagrees with the claimed statement, used a input dataset too tiny to achieve statistical significance, looks suspiciously P-hacked, you then misrepresented your own findings to the press and by extension the government and then to top it all off it doesn't replicate. We will therefore prosecute you for research fraud and failure to meet the terms of your contract.
What actually happens is this: nothing. Journals will happily publish papers with the problems I just listed, universities praise them, the 'scientists' who do this stuff proceed to get lots of citations and the government proceeds to award even more grant money because who are they to argue with citations.
As you admit, fraud is everywhere in science, supposedly due to "high expectations for results". But so what? Lots of people, non scientists included, have high expectations for results placed upon them. The right mechanisms and incentives to stop fraud are not simply having low expectations of scientists, that's absurd and wouldn't be seriously proposed as a solution in any other area of society. It'd be like saying the answer to fraudulent CEOs fiddling the books is to simply stop expecting them to turn a profit, or like the solution to shoplifting is to just stop expecting shoplifters to pay for things.
Expectations on scientists are already rock bottom: large chunks of the research literature doesn't even seem to replicate, other large chunks are not even replicable by design, and nobody seems to care. You can't get much lower expectations than "We don't even care if your claims replicate" and yet this ridiculous suggestion that the solution to fraud is to give fraudsters even more money keeps cropping up on HN and elsewhere. The solution to fraud is tighter contracts to ensure the rules are clear, and systematic prosecutions of people who break them.
> I haven't, but I'm unclear why it's relevant given that you don't seem to really be disagreeing!
It's relevant because you are criticizing the process but you don't seem to understand how it actually works. Your original characterization was that grant money is "this firehose of tax money being sprayed everywhere without even the slightest bit of accountability in how it's used." The reality is, when I get grant money I need to account for how every dollar is spent, and if there are ever any questions about spending, I better have the receipts to back it up. The other reality is, I only get to spend a small fraction of a grant award, as the University takes most of it off the top, and my students take almost all the rest in the form of a tuition remission and a stipend, leaving whatever is left over for equipment and conference costs. Then there are strict conflict of interest regulations which come with their own reporting requirements. I don't even get all of the money at once; I'll get some up front and then I have to show significant midterm progress in order to get more. There's accountability at multiple layers by multiple organizations.
> The NSF denies 80% of grant applications because there is a radical oversupply of people who want to be scientists and the NSF has a finite budget.
It's accountability in the form of: if you didn't do what you promised you'd do, then you don't get any more money and your career is derailed. Isn't that what you want? Anyway, do we have an oversupply of scientists relative to the amount of science that needs to be done? I don't think so. The NSF budget is finite, but it's also embarrassingly miniscule given the upside of research that has come out of NSF funded projects.
> If it existed it would look like this ... We will therefore prosecute you for research fraud and failure to meet the terms of your contract.
Fraud is one thing. I'm not going to say we shouldn't prosecute fraud. But treating a grant proposal like a contract with positive deliverables (no negative results allowed) is the exact problem. Research is not development. Research implies failure. You can't do one without the other. Failure to meet stated objectives shouldn't be met with prosecution. That just further incentivizes fraud.
If there's a replicability crises it just means we need to spend money on replication research. Researchers know no one is going to bother replicating their study because there is no grant money available for redoing someone else's research. Grant agencies don't pay for that kind of thing, and you can't make a career doing that kind of research. No one gets tenure this way. If we want studies to be replicated, we need to allocate money to replicate them, and we need to incentivize people to do so by making it a viable career for a Ph.D.
> Lots of people, non scientists included, have high expectations for results placed upon them. The right mechanisms and incentives to stop fraud are not simply having low expectations of scientists, that's absurd and wouldn't be seriously proposed as a solution in any other area of society.
I didn't say we should have low expectations, I said we should have realistic expectations, and yes, that does involve lowering expectations from where they are right now. Because the current expectation is this: you have to write a proposal that has a <20% chance of getting funding. If you can't get that funding your career is basically over, so you better promise the world, because everyone else is. In this proposal you need to lay out a research plan for the next 3-5 years and you have to convince the funding agency that your research is going to change the world as we know it. If within that time you fail to meet your stated objectives, you will find funding hard to come across, and your tenure will be threatened, meaning you will probably lose your job and have to move your family. On top of that you want to add potential federal prosecution to stakes, thinking that will make things better.
> Expectations on scientists are already rock bottom .. The solution to fraud is tighter contracts to ensure the rules are clear, and systematic prosecutions of people who break them.
Okay, run with this idea: exactly what rules need to be clearer and exactly how do the contracts need to be tightened? Because there are already clear rules and tight contracts, yet the problem persists. Will clearer rules and tighter contracts fix it? How?
I'll tell you what I think will happen with this system: you'll chase out all of the public scientists because the stakes are too high. Already the pay is too good on the corporate side, and now you add potential federal prosecution to the list if I don't meet positive deliverables? No thanks. I'll go work for Microsoft where my research will be privatized. You might be okay with this as you pointed out you believe a profit motive is good for research, but you know who wouldn't be good with this? Microsoft. And Google. And all the other tech companies who were (or will be) built on top of technologies that started as government funded research. All this does is make Microsoft stronger. Is that what we want? What about the next Microsoft or Google? Where will they come from?
I'll give you a concrete example of where your idea fails: the 2004 DARPA Grand Challenge. Millions were spent trying to bootstrap autonomous cars, and what was the result? They all crashed, no one completed the race. What should the response have been, to prosecute everyone involved? No, they tried again and gave everyone more money. Next time around in 2005 more succeeded (mostly because they relaxed the expectations).
Then in 2007 we saw the first real demonstration of autonomous cars in the DARPA Urban Challenge. Today, everything Tesla, Google, GM, Ford, et al. are doing with driverless cars is based on the research that happened in 2004-2007. Without government funded autonomous car research, there would be no Tesla or Waymo today. That's how research works, you try, you fail, you try again, and you have no idea how far your impact will be, and really no one does. If we try to control this process toward producing only successes with contracts and positive deliverables, like it's an engineering project (with prosecution of failure and all), it just means we're going to lose dynamics like the Grand Challenges, and the broader economy will suffer for it.
Take all that money you want to invest in prosecutors, courts, lawyers, and prisons, and invest that in a system where replication studies are well funded and a viable career path for scientists. Increase funding into the NSF and other grant funding agencies to hire more people to consider grants, and increase grant throughput. I guarantee you you'll fix a lot of the problems you're identifying.
I think we are 80% in agreement but still using words differently.
> when I get grant money I need to account for how every dollar is spent
Yes I know, but that's not what I mean by accountability. Again: nobody is upset with academics because of expenses scandals or taking too many expensive flights. Well, except maybe for climatologists who supposedly take more flights than the average academic, but that's due to the perception of hypocrisy rather than concern over cost.
People are getting upset because when they download and read papers, the papers turn out to be bad and there are no visible consequences for that. Even just getting a clearly fraudulent paper retracted is reported to be a nightmare, according to people who search for scientific fraud as a hobby like Elizabeth Bik. And I've read endless reams of terrible papers that were useless or outright deceptive, I tried reporting a few and nobody ever cared.
Now, you're arguing that there is accountability of the following form:
> It's accountability in the form of: if you didn't do what you promised you'd do, then you don't get any more money
This is true given that scientists are promising the NSF to publish papers, not strictly speaking to do research, and therefore by implication promising to come up with interesting claims, not necessarily true claims. But that's not what we want.
This is an inevitable problem with government funding of research. The buyer, the government, cannot really check if the claims they're buying from scientists are true, so they need proxies like did it get published, did it get cited, etc. But those aren't the same things. Corporate research doesn't have this problem because the corporate will try to apply the research at some point and if it was fraudulent they will discover it at that point, and of course they're strongly incentivized to ensure it never gets to that point in the first place.
In theory the government could write grants in such a way that money is awarded independent of what claims end up being made, instead awarding money for the quality of work done. That's what you're arguing for here. And indeed corporate labs write contracts in this exact way. Scientists get a salary in a corporate lab, they don't have to write grants. They do have to convince their management chain that the research is worth funding, but there are many different ways to do that which don't involve continually publishing astonishing claims in scientific journals.
You're asking me to propose how science should work instead but, indeed, you already know my answer: eliminate the NSF completely, and stop subsidizing student loans. All science should be funded by companies. They have already solved the problems you're treating as novel / intractable above. Scientists are awarded salaries and promotions by firms on a more flexible basis than the NSF. Importantly, they are rewarded for doing research not producing claims. Companies can do this because they have management structures sufficiently well staffed to closely monitor what scientists are doing. That means if a firm is truly committed to research then the scientists will get paid even if their programme has some dry years. Plus there's a huge body of law handling fraud and corruption in the workplace.
At the same time, firms are incentivized to eliminate the research that is probably always going to be nearly useless. Outside of firms selling books or self help courses I doubt many would subsidize sociology or gender studies for example, and it's also unclear that would be a loss.
Your argument about who it would or wouldn't be good for seems a bit contradictory and I struggled to follow it. You're arguing it would be both bad for Google and Microsoft yet also make them stronger. I disagree with both possibilities: I think they would hardly notice the difference and it wouldn't affect how powerful they are. Having worked for one of those companies and also worked at a startup where we often read research papers in a certain subfield of CS with views to maybe applying them, my view is that even in the relatively good field of computer science, most academic output is useless and has no impact. These firms do not rely heavily on government funded research:
- The web was very briefly funded for a couple of years as a side project of CERN, but then R&D was taken over by the private sector where it remained ever since. Page & Brin never even finished their PhD before moving their research into the private sector! It's hardly a mystery where the next Google will come from - probably the same place the previous one did, a garage in Silicon Valley.
- What government funded tech was Microsoft built on? The internet? Microsoft is still with us in spite of the internet, not because of it! Or are you going back to military computers in World War 2? Military R&D is different, governments can fund that semi-effectively because they actually use the outputs.
- Neural networks were a backwater until Jeff Dean resurrected the field using the resources of the private sector, academia has been chasing to catch up ever since.
There are a lot of other examples. The DARPA Grand Challenge is not an example of what I'm talking about because:
1. DARPA is military research and therefore structured differently to how the NSF does things. The very structure of it as a Grand Challenge is a clue here: the output of the programme was cars (not) going round a track, not papers and citations.
2. I'm not arguing for prosecution of researchers who end up with null results!
I'll try not to do another wall of text since we're mostly in agreement, but I will make a couple final comments:
> Your argument about who it would or wouldn't be good for seems a bit contradictory and I struggled to follow it. You're arguing it would be both bad for Google and Microsoft yet also make them stronger.
What I meant is, if e.g. Page and Brin in 1998 had no access to government funding and research because it was privatized by e.g. AOL, there wouldn't be a Google today. But if we were to privatize all research, Google of today would certainly like that insofar as it strengthens their market position (jut like the AOL of 1998 would like the situation), but it also means they have to start funding more research because now they can't get any from the public.
> - The web was very briefly funded - What government funded tech was Microsoft built on? - Neural networks were a backwater
But the point is that it all started with government funding, so we need to be very careful about the consequences of privatizing it all. Today, ideas start out funded by the government, they gain legs in academia, move out into corporations, and are productized and disseminated to the public in the form of consumer goods. This is the progress pipeline, and it's proven extremely effective and enduring at driving innovation.
You want to cut out the beginning of the process because you think corporations can handle that part, but I don't think you've really demonstrated that. Can you point to any tech product out there that is exclusively built on in-house, private research? I certainly can't think of one.
For example, you bring up the origin of Page & Brin. Yes, they never finished their Ph.D., but the fact is they did meet in grad school while they were doing NSF funded work. Brin was at Stanford on an NSF fellowship. They built the first prototype of Google on an NSF grant. They were mentored by academics who also were funded by the NSF as professors and graduate students themselves. You take that funding away, and maybe these two people never meet, maybe they never learn what they need to get that spark of insight. So I agree with you that the next Google will come from the same place the previous one did - a government-funded research lab in Silicon Valley. The garage is where they moved their operation only after they had already used a lot of NSF money to get their start.
> 1. DARPA is military research and therefore structured differently to how the NSF does things. The very structure of it as a Grand Challenge is a clue here: the output of the programme was cars (not) going round a track, not papers and citations.
The processes of getting grants from NSF and DARPA are very similar, and in most cases the deliverable is a paper. The Grand Challenges are the exception of DARPA funding, not the rule.
> Military R&D is different, governments can fund that semi-effectively because they actually use the outputs.
Yes and no. DARPA would like to use the fruits of its funded research, but it funds projects on a very long timescale, so what it funds may or may not be used in the long term. Sometimes the research is not to strengthen the military per se, but to strengthen American interests though creating domestic tech sectors. e.g. I'm sure the military would like to use autonomous vehicles, but what's even better is for America to have its own domestic autonomous car sector that can produce those vehicles.
> most academic output is useless and has no impact.
You've tried to make the case that we should optimize toward useful research, and companies are better at identifying useful research because they have a profit motive, but I still think it's difficult to say today what research will be important 30-40 years down the line. DARPA recognizes that it's very hard to tell how useful research will be ahead of time, and that corporations don't like to engage in foundational research when there is no obvious short-term path to profit. This was the entire point of the Grand Challenge series, and it worked out well -- they wanted to bootstrap the autonomous car industry, so they paid researchers to get them rolling and now look where we are. If the government hadn't gotten involved, there probably wouldn't be an autonomous car sector in the US today.
There are plenty of cases in our history where some technology seemed useless initially turned out to be bigger than anyone could have imagined. We need to be careful not squelch those ideas too quickly because they don't return an immediate profit. Things like the Internet and neural networks come to mind. A lot of people, particularly large corporations, thought the Internet was a toy when it first was introduced. Neural networks seemed like a dead end and then found new life. But the fact is they started in academia. The Deepmind arcade paper and essentially the entire deep reinforcement learning field today is based on decades-old research funded by the UK government. What if that research was locked away in a UK corporation? Would Deepmind even exist? That research was a toy for 30 years, until it wasn't.
The whole point of DARPA and other government funding agencies is that they don't know what the winners are ahead of time, and I don't think corporations can know this either. (if they could, why didn't they do more to fund RL research 30 years ago?). Therefore we shouldn't try to optimize for obvious winners because we'll miss out on non-obvious winners, which bring the biggest upsides. This means we have to fund losers and research that ends up not being useful, and we should be okay with that, because things have turned out pretty well over all.
> 2. I'm not arguing for prosecution of researchers who end up with null results!
Sorry I thought you were with this:
We will therefore prosecute you for research fraud and failure to meet the terms of your contract.
I guess you mean failing to meet the terms of your contract and fraudulently representing that. But it still doesn't address the incentive to commit fraud because if you fail to meet your objectives, you're still not going to get published and therefore won't get the next grant, so your career is still derailed. It just means people will try to hide the fraud better.
After I typed all this I realized I failed at my pledge to not give you a wall of text. Oops!
What I mean by prosecution is that if a research body signs a contract with a scientist to do research, then those contracts would need to specify what research actually is, and that is the first step towards penalizing people who aren't really doing it. Indeed the process of flushing more research into the private sector would automatically eliminate a lot of the grey-area fraud that is so prevalent, because it would force a lot more people to write down what precisely they mean by "doing research", as well as continually evaluate that definition via normal management techniques. For example, is a simple modelling exercise "research"? It's often treated as such by e.g. banks, but the big tech labs we're talking about don't engage in a much of that, unless you count AI, but I think that's sufficiently beyond the sort of modelling you find in most science that it's best to treat it separately.
At the moment governments fund science but have no working definition of what science is, which breeds a lot of cynicism of the type I display above w.r.t. sociology. Is gender studies "science"? Most people would say no, but the government says yes. A more subtle example is epidemiology. A close examination of their papers will reveal that it's just plugging public CSV files into a bunch of very over-simplified simulations, and publishing the outputs. Is that science? If it is, can I get paid to play Cities: Skylines all day as long as I write a paper at the end? It sounds like a stupid suggestion but actually yes I can:
In my view this type of thing is not science, but my guess is at this point the science-y ness of epidemiology or urban planning would split 50/50 or most people would just go with the government's definition of "they receive grants and call themselves scientists, therefore they're scientists".
Would Google exist without the NSF? The specific company maybe not, but there were plenty of search engines around before Google, and Page in particular was already keen on creating a tech company when he was very young so would likely have ended up a startup founder sooner or later. An example competitor was Inktomi, which had already started doing pay-per-click ads. It's all forgotten now but Google nearly didn't survive its early years because they got sued over 'stealing' the PPC ad concept. They were able to argue that their own elaborations on the idea were sufficiently different that it wasn't infringement. It's very plausible that one of these other firms would have struck upon the idea of PageRank; they were certainly incentivized to do so especially once Inktomi had realised that PPC ads were a way to monetize search engines.
"The Deepmind arcade paper and essentially the entire deep reinforcement learning field today is based on decades-old research funded by the UK government. What if that research was locked away in a UK corporation? Would Deepmind even exist?"
Well DeepMind is a difficult example to debate here for both of us because of course DeepMind is or was a UK corporation and they do the exact opposite of locking up their research, if anything they're famously publicity and paper hungry. Google/DeepMind are actually a strong counterpoint to the idea we need academia for long range research: DeepMind is nothing but long range research (of unclear utility!) and of course self driving cars have been driven by Google for the last decade, pun totally intended.
If I were arguing in your shoes I'd be trying to argue Google is the exception that proves the rule and/or trying to distract attention from it, because it shows that companies can and will do long range research. Microsoft Research is another example, although it's less "pure" because it's more or less a little recreation of academia inside of Microsoft. I prefer the Google approach where science and technology are fully integrated.
Now the wider issue of governments needing to fund long range research is one I used to fully agree with. It sounds right and it's easy to find examples where you can sort of link them to government funded research. But as you can see, I changed my mind over time and no longer find myself in that camp, because:
1. Government funded basic research isn't free. We have to weigh up costs and benefits. How much of a contribution does government grant money make to the technological successes we take for granted today? For examples like PageRank, self-driving or DeepMind the initial contribution was quite small and mostly in the form of logistics (grand challenges) or theory work (which is cheap). And how much of a cost does it impose?
2. The costs are not just financial. I guess this is what mostly changed my mind. I concluded a big part of the "cost" of government funded research is actually in terms of intellectual pollution of the literature. If you have to wade through 50 useless, deceptive or outright fraudulent papers to find 1 good one because governments aren't paying attention to what they fund, then that poses an externalized cost on everyone who wants to benefit from research. Moreover this work has to be endlessly duplicated because journals are loathe to retract anything, so everyone who wants to push technology forward in a certain area has to do this work within their own small group because there's no coordination mechanism ... or just give up and ignore the literature entirely (this is what eventually happened to me).
I think a stronger argument for government funded research than the "it would never have happened" approach is that government funded science is usually un-patented and freely accessible. But even this argument is kind of weak because universities do patent the results of tax funded science, maybe not in computer science but it happens a lot in other fields, and also because the results of the research are often behind paywalls too! Although that's been getting better with time and is usually not a problem in CS (which IMHO is definitely one of the better fields).
But overall, to me it's just not clear that the benefits of buying papers en-masse outweighs the costs, both in dollar terms, time terms and of course, the inevitable costs when people put bogus research into production and things go wrong.
The NSF denies 80% of grant applications because there is a radical oversupply of people who want to be scientists and the NSF has a finite budget. That by itself doesn't create accountability any more than the fact that lots of people want to be movie stars creates accountability for actors. That's not how accountability works.
Accountability means people being held to account for illegitimate acts. If it existed it would look like this: we (the government) gave you money to deliver some genuine research, yet you delivered a paper that simply modelled your own beliefs, cited a retracted paper and cited another paper that actually disagrees with the claimed statement, used a input dataset too tiny to achieve statistical significance, looks suspiciously P-hacked, you then misrepresented your own findings to the press and by extension the government and then to top it all off it doesn't replicate. We will therefore prosecute you for research fraud and failure to meet the terms of your contract.
What actually happens is this: nothing. Journals will happily publish papers with the problems I just listed, universities praise them, the 'scientists' who do this stuff proceed to get lots of citations and the government proceeds to award even more grant money because who are they to argue with citations.
As you admit, fraud is everywhere in science, supposedly due to "high expectations for results". But so what? Lots of people, non scientists included, have high expectations for results placed upon them. The right mechanisms and incentives to stop fraud are not simply having low expectations of scientists, that's absurd and wouldn't be seriously proposed as a solution in any other area of society. It'd be like saying the answer to fraudulent CEOs fiddling the books is to simply stop expecting them to turn a profit, or like the solution to shoplifting is to just stop expecting shoplifters to pay for things.
Expectations on scientists are already rock bottom: large chunks of the research literature doesn't even seem to replicate, other large chunks are not even replicable by design, and nobody seems to care. You can't get much lower expectations than "We don't even care if your claims replicate" and yet this ridiculous suggestion that the solution to fraud is to give fraudsters even more money keeps cropping up on HN and elsewhere. The solution to fraud is tighter contracts to ensure the rules are clear, and systematic prosecutions of people who break them.