> Nowadays people get told that if they think they’ve figured out something about gravity, they’re probably a crackpot. Instead, they should wait for very large government-funded programs full of well-credentialled people to make incremental advances.
Maybe we approach limit of human cognitive abilities and another Newton is not possible for current problems? Maybe to find a cure for cancer you need to spend 10k hours studying biology, 10k studying chemistry, 10k studying machine learning and then spend 100k hours on the problem, which it's not possible for a human being.
Possibly. But then, we've always been good in solving problems with tools. Could we reduce the necessary "10k" learning by building better tools for accessing knowledge?
Right now, scientific papers are a shitshow, and even refined and systematized knowledge in textbooks isn't as accessible as it could be in the digital age. Could we encode knowledge in more powerful tools allowing us to explore through free-form simulations?
The goal here would be to make discovery of promising solutions more time- and effort-efficient.
Related, I feel we need to figure out a way to systematize information in scientific papers to make mining them for cross-cutting insights possible. I suspect there are lots of discoveries hiding in plain sight, if one knew which particular papers from which disciplines touch on the same underlying phenomenon or concept.
Related, I feel we need to figure out a way to systematize information in scientific papers to make mining them for cross-cutting insights possible. I suspect there are lots of discoveries hiding in plain sight, if one knew which particular papers from which disciplines touch on the same underlying phenomenon or concept.
This is interesting problem, unfortunately it's really, really hard.
I'm most familiar with mathematics, so I'll use it as an example, but this is not limited to mathematics.
If you take any new paper on research mathematics, in a hot field like algebraic geometry or partial differential equations, then unless you're an expert in that field, it will almost always be literally impossible for you to understand -- not simply hard to follow the arguments, but simply impossible to understand even what it's about. Look, I just grabbed random example from recent posts on arXiv[1]: try reading an abstract and explaining it back to me. For 99.99% English speakers, this will be indistinguishable from random gibberish in a paper written by recurrent neural network trained on arXiv papers.
However, 0.01% of people will understand something, and for probably 1% of these, the abstract will make perfect sense. However, if you ask these people to explain it, you'll either spend an hour or two on getting some very superficial understanding of what's at stake here, which won't be very useful to you -- you still won't be able to actually read and follow the paper, and use the insights for your own purposes. Alternatively, and if you're intelligent enough, they can spend a year or two teaching you required background. Then you can see the insight for yourself.
The problem here is that you need literally years of background studies to appreciate the insight. There likely is no quick and easy way around it, otherwise some of the extremely smart people involved would already have had figured it out -- assuming otherwise is hubris. This doesn't mean that the system cannot be improved upon: there's tons of ways to make things simpler, clearer, more digestible. However, you'll still be left with hard problems of hard things being hard.
OK, let's check your numbers. There are about 1.5 billion people who speak English, and you are telling us that only .0001 * 1.5 billion = 150,000 of them will get anything out of the abstract? And that only .01 * 150,000 = 1,500 of those will find it to make perfect sense? That doesn't seem right.
Abstract. Using elliptically fibered Kummer surfaces of Picard rank 17, we construct an explicit model for a three-parameter bielliptic plane genus-three curve whose associated Prym variety is two-isogenous to the Jacobian variety of the general three-parameter hyperelliptic genus-two curve in Rosenhain normal form. Our model provides explicit expressions for all coefficients in terms of modular forms.
Oh. Oh my. Checks list of Sokal Squared spoof papers, nope. "Indistinguishable from random gibberish in a paper written by a recurrent neural network" it is then. Rather than being conservative, now I see that you were wildly optimistic in your estimate. There can't possibly be over 1,000 people in the world to whom this would make perfect sense, can there?
IMO it can be both "hard" and "really, really hard", depending on paper's difficulty. Some papers are much easier to read - e.g. those describing experimental results. Often, being the 0.01% who understand something is enough - it's one thing to read a paper, understand the basic reasoning and trust the conclusion / proposed method; it's another and more difficult thing to fully comprehend the paper and verify its correctness. Often you don't need that full comprehension to make progress and put the paper to good use.
> The problem here is that you need literally years of background studies to appreciate the insight. There likely is no quick and easy way around it, otherwise some of the extremely smart people involved would already have had figured it out -- assuming otherwise is hubris.
That's fair; something of a scientific version of efficient market hypothesis, I guess. If following the bleeding edge of a scientific domain didn't require years of background studies, it would be easy, so scientists would quickly zoom through it, until the going got difficult again.
> This doesn't mean that the system cannot be improved upon: there's tons of ways to make things simpler, clearer, more digestible. However, you'll still be left with hard problems of hard things being hard.
Yeah, I was just thinking about ways to tackle the things that can be made "simpler, cleaner, more digestible". I don't deny that there are fundamentally hard problems we have to face directly, but right now, those problems are wrapped in a lot of irrelevant cruft that makes them bigger than they really are.
> try reading an abstract and explaining it back to me
Challenge accepted, though I know the result kind of reinforces your point. But here's what I understood from that abstract:
> Using elliptically fibered Kummer surfaces of Picard rank 17, we construct an explicit model for a three-parameter bielliptic plane genus-three curve whose associated Prym variety is two-isogenous to the Jacobian variety of the general three-parameter hyperelliptic genus-two curve in Rosenhain normal form. Our model provides explicit expressions for all coefficients in terms of modular forms
We took a particular weird abstract shape with interesting properties, and used it to describe a particular different weird abstract shape, whose properties are important to us. Abstract shapes can be written down as maths, and depending on the way you write it, they can have properties exposed directly as "knobs" to tweak - e.g. "circle of radius r" has a radius exposed directly, whereas "circle that fits in that place" hasn't. In this paper, our description of the weird abstract shape has its important knobs exposed.
Depends what you mean by AI... yea if you ask Siri for a cure and she succeeds then you're the instrument, but if I make clever use of the statistics and math behind it all, and find a way to apply it to some drug discovery or procedure, then I'm the hero, and the AI was just a tool.
If you hop into an excavator, you can't really compare yourself with great builders of the past, but you can move ground around much faster than they ever could. This is what I had in mind - not building oracles for scientific discovery (though that would be cool, too), but building excavators for the mind, and airplanes for the mind (and building a Bagger 288 for scientific papers).
I realise that your numbers are only meant to be illustrative, but it's worth realising that this is "only" 65 hours a week for 40 years. A lot? Yes. More then I'd be willing to work at a typical desk job[1]? Probably. But feasible for a passion project? Definitely! (and doesn't even necessarily require starting before you hit 25, although that clearly helps...). The limiting factor is not so much that nobody can do this, but that (short of independent wealth, which I think can bring its own set of constraints and expectations) very few people ever get the opportunity to do this without a lot of interruptions. I don't think it's impossible to imagine a society where that isn't true.
[1] although... "thinking" can stack surprisingly well with some activities we'd consider leisure, like walking or (at least in my case) gardening. So maybe this isn't really going to be 65hrs/week at a desk...
Pretty much -- for example, I had definitely spent 10k hours each studying biology, chemistry and computer science by the time I reached age 21.
But, all this thinking is in my experience pretty useless in terms of real world results (and I don't mean results like Stanford idolizing you). You're not going to "think" your way into curing cancer no matter how many hours of biology or chemistry you take, and I should know, because I've seen people try. It's pretty hard to "think" up a company the size of Amazon, too, especially in a world where a lot of industry have their Amazon. You need capital and a degree of self-confidence bordering on manic delusion and when you have this you still need to not do what Elizabeth Holmes did and know when you've failed and give up what may have been ten or twenty years or a lifetime and start something else.
It's not the 100k hours. It's the very large odds they will have been for nothing, and picking up whatever is left of your life after.
It's not the 100k hours. It's the very large odds they will have been for nothing, and picking up whatever is left of your life after.
That's an interesting take. But my experience is that there is a non-negligible set of people who would see a lifetime spent working on their chosen problem to be well-spent, even if they don't eventually succeed.
No, we build off of each other’s knowledge. We learn some things, simplify them, and then build on that knowledge. We didn’t go right from wires and signals to JavaScript and web browsers, it was incremental, where each generation built off of the last one’s work.
That’s how many hours it would take today. But if we get a few fancy new libraries for ML, it could be trivial to implement. 20 years ago, if I told you a single developer could deploy a website with a db + authentication in about 20 mins, you’d call me a liar. But today, you can use a fullstack generator and deploy to heroku with the press of a few buttons.
Maybe we approach limit of human cognitive abilities and another Newton is not possible for current problems? Maybe to find a cure for cancer you need to spend 10k hours studying biology, 10k studying chemistry, 10k studying machine learning and then spend 100k hours on the problem, which it's not possible for a human being.