“if a computer can solve a thirty-year old math problem in a weekend because it’s a math problem, should we…make more problems math problems”
Yes! I’ve had the exciting and uncomfortable realization recently that almost anything in the world can be converted into if/then statements if you’re willing to spend the time encoding your human decisions like that. And then it’s just a question of:
- Can you break each individual step into a simple enough decision than an LLM can make it?
- Can you write a script that strings those decisions together in the right order and with enough checks and balances to not veer off-course midway through?
- Can you afford the token bill?
- Can you live with the knowledge that your own human experience, knowledge, and creativity was ultimately reducible to a string of 1s and 0s?
yeah, I can't help but have the slightly uncomfortable feeling that we're all sort of in a race to figure out out how to turn all our problems into Moneyball (or really, high-frequency trading) problems, and we just don't really realize it yet. And in ten years, when a few people figure that out, we'll all be sitting around wondering how we didn't see it.
I think this is actually one of the few right use of AI: crack the hard problems that humans cant solve but have clear expectation with verifiable math. Maybe "curing cancer" will no longer be a joke but a reality with the right use of AI like this.
At the start of all of this, AI curing cancer felt ridiculous to me, but now, I gotta say, it doesn't seem that crazy, for exactly this reason. Surely, assuming we make it that long, give people 1,000 years, and we would've figured out how to cure cancer. And AI seems like it's just a way to compress that into a much shorter timeline.
And what if AI figures out that most cancers are caused by chronic exposure to certain things (which by and large we already know, e.g. lung cancer and smoking) -- will we then ban these things (e.g. lead in paint), or are we looking for some kind of magical injection so that we can keep smoking? What if the things we crave are made in a way that gives *other people* cancer -- will we ever care?
Righteousness is not a math problem; it is a behavior change -- and it is much simpler than a math problem (i.e. just cease from doing or contributing to evil).
Funny that every example here comes with free wrong answers. A failed counterexample costs nothing, so the loop gets a million swings; the moment a test needs a wet lab or a real customer, the swings aren't free anymore and the loop slows way down.
I generally like your thinking and in this case you point to "restructure data to be more suitable to an LLM" and "change how we write code so that it's optimized for a machine" - I really liked both of those articles/sets of thinking of yours. But in this case, I came across a different set of thinking on the Jacobian conjecture subject (I'm in the "what?" camp unfortunately) - but its conclusion is very different - AI is best where the solution is easily verifiable (the Jacobian conjecture solution was confirmed by Tao very quickly), and that societal problems take years to know whether the solutions really work. Anyways, that was stated much better by Carlo Iacono - here - https://hybridhorizons.substack.com/p/ai-gets-good-where-the-world-can
in combination with genuine, rapid progress from these machines.
Taking our highly complex, often pre-linguistic and tacit understandings of the world and creating meaningful and effective math-shaped versions of those understandings (while exercising appropriate fear and judgement regarding the inevitable risks of the Procrustean bed) so that it can then be worked through by a machine. This is a skillset that can definitively be trained and requires effort, craft and domain-specific expertise.
Interesting article. I recall Nassim Taleb arguing in Antifragile that many important innovations have originated from brute trial and error. Smart theory doesn’t precede the breakthrough, but follows it.
“if a computer can solve a thirty-year old math problem in a weekend because it’s a math problem, should we…make more problems math problems”
Yes! I’ve had the exciting and uncomfortable realization recently that almost anything in the world can be converted into if/then statements if you’re willing to spend the time encoding your human decisions like that. And then it’s just a question of:
- Can you break each individual step into a simple enough decision than an LLM can make it?
- Can you write a script that strings those decisions together in the right order and with enough checks and balances to not veer off-course midway through?
- Can you afford the token bill?
- Can you live with the knowledge that your own human experience, knowledge, and creativity was ultimately reducible to a string of 1s and 0s?
yeah, I can't help but have the slightly uncomfortable feeling that we're all sort of in a race to figure out out how to turn all our problems into Moneyball (or really, high-frequency trading) problems, and we just don't really realize it yet. And in ten years, when a few people figure that out, we'll all be sitting around wondering how we didn't see it.
It's all arbitrage, all the way down.
I think this is actually one of the few right use of AI: crack the hard problems that humans cant solve but have clear expectation with verifiable math. Maybe "curing cancer" will no longer be a joke but a reality with the right use of AI like this.
At the start of all of this, AI curing cancer felt ridiculous to me, but now, I gotta say, it doesn't seem that crazy, for exactly this reason. Surely, assuming we make it that long, give people 1,000 years, and we would've figured out how to cure cancer. And AI seems like it's just a way to compress that into a much shorter timeline.
And what if AI figures out that most cancers are caused by chronic exposure to certain things (which by and large we already know, e.g. lung cancer and smoking) -- will we then ban these things (e.g. lead in paint), or are we looking for some kind of magical injection so that we can keep smoking? What if the things we crave are made in a way that gives *other people* cancer -- will we ever care?
Righteousness is not a math problem; it is a behavior change -- and it is much simpler than a math problem (i.e. just cease from doing or contributing to evil).
Math is hard.
but not *that* hard, apparently.
Funny that every example here comes with free wrong answers. A failed counterexample costs nothing, so the loop gets a million swings; the moment a test needs a wet lab or a real customer, the swings aren't free anymore and the loop slows way down.
I generally like your thinking and in this case you point to "restructure data to be more suitable to an LLM" and "change how we write code so that it's optimized for a machine" - I really liked both of those articles/sets of thinking of yours. But in this case, I came across a different set of thinking on the Jacobian conjecture subject (I'm in the "what?" camp unfortunately) - but its conclusion is very different - AI is best where the solution is easily verifiable (the Jacobian conjecture solution was confirmed by Tao very quickly), and that societal problems take years to know whether the solutions really work. Anyways, that was stated much better by Carlo Iacono - here - https://hybridhorizons.substack.com/p/ai-gets-good-where-the-world-can
Really interesting discussion and framing.
Deeply hopeful for the future of work
in combination with genuine, rapid progress from these machines.
Taking our highly complex, often pre-linguistic and tacit understandings of the world and creating meaningful and effective math-shaped versions of those understandings (while exercising appropriate fear and judgement regarding the inevitable risks of the Procrustean bed) so that it can then be worked through by a machine. This is a skillset that can definitively be trained and requires effort, craft and domain-specific expertise.
Lots to meditate on. Thank you.
Interesting article. I recall Nassim Taleb arguing in Antifragile that many important innovations have originated from brute trial and error. Smart theory doesn’t precede the breakthrough, but follows it.
Dear Chat GPT, write my blog in a crass rip off of the style of Matt Levine