Until very recently, I pored over every single line of code Claude generated with razor sharp scrutiny. I would usually catch issues with every response. I'm catching fewer problems these days. Maybe the model is just getting better, and maybe I'm being less careful while under pressure to ship more and more often. But model capability is obviously growing. Even back in March, you could tell it "give me a function that adds two numbers" and you could be 100% confident that it would write the correct function. There was almost no point in looking at the code. Since then, the complexity floor of problems in the category "this is so simple that the model couldn't possibly get it wrong" is rising, and with it, my cognitive surrender to the model is increasing too. Why check it? It's obviously going to be correct.
If AI designs a terawatt fusion plant, then of course we're going to meticulously pore over every detail to ensure safety, reliability, efficiency, whatever. If we find no flaws in the design whatsoever, will we be less careful about the second one? The third one? What about the ten thousandth one? Will "a nuclear fusion plant" become something that models couldn't possibly get wrong?
Terence Tao is arguing that the human involvement in research is crucial, but doesn't convincingly justify why, in my opinion. He says that "human agency is a value of fundamental importance" and that we will need to build "thriving human communities that can understand [AI ideas] together" - not for the sake of correctness, which AI may surpass us on, but for, I guess, the possibility of reclaiming human meaning and purpose. I don't disagree with this at all, but it's not an argument, it's a statement of values. Unfortunately, the stark reality is that if AI does surpass humans, it will become the economically dominant strategy to not verify them and not double check them, but to just do whatever they say. This seems like a great way to raise p(doom). But as the models get better and better, and as I'm scrutinizing Claude's output less and less... I just hope that there are more Terence Taos out there than people like me.
I think you have a fundamental misunderstanding here, and it's not really explained because I think it seems self-evident from within the field. In short: writing code is a means to an end; doing mathematics research is not, but is the end in itself.
The human involvement is crucial because the entire purpose of mathematics research is to increase human understanding of mathematics. It is pursued because it is interesting, not because it is economically useful. In this sense it's a lot closer to the humanities.
A black box oracle that just tells you whether statements are true or false is not the goal of mathematics and would not be particularly interesting to the field (except insofar as it could be harnessed to improve human understanding).
Coding is totally different from this, where it is essentially always done as a means to an end. Likewise with many other fields, like pharmaceutical research or materials science or what have you, that are oriented around solving problems for some practical purpose. Pure math isn't really like that for the most part.
Then in the 80s, you presses 2 buttons and there you had it in your classroom without thinking twice if the electricity arrived correctly at the transistors.
How crazy will the world be once our [current gen] ANN are like that!
What an amazing thought.
So I think the problem is to determine which problems under what instructions we can safely give to a model application to solve and how we test the output for safety and functionality. This would create more usable and safe, albeit a bit more boring, AI-based applications alin to a calculator or general computer. Whether this is posswith current model architecture is another thing.
How do you know?
Memory bits flip randomly. It's not a super rare thing either. You and me have experienced that many times without knowing. The only reason that computers feel deterministic is that we have error-correcting code to fix that. But in the most extreme cases, when multiple bits flip together, once "deterministic" program can generate unexpected output.
So why do you trust computers? Because statistically the case is just very unlikely. Therefore if AI is statistically unlikely to make mistakes there is no reason to not trust them.
With statistical models - such as LLM’s - there is no logic as such, but statistical assumptions based on given data. The output can ge very good or very bad, but you are fool to trust it blindly. Therefore you need a deterministic way to verify, whether meat- or software-based.
- Eric Hoffer
For what it's worth, ten thousand terawatt fusion plants probably approaches the level at which the sheer intensity of energy production would cause significant disruption to the climate (it's roughly 5% of the Earth's entire solar input). Every energy source becomes dirty past a certain point. It would be wiser to learn how to build a utopia within a limited energy budget than find a way to produce enough of it to cook the damn planet, but who am I kidding, we're going to build a million of these things.
You aren't really trying in good faith to think this through are you? This idea is over half a century old. Not getting it by now is willful.
They will never make a logical error yet make terrible assumptions and poor long scale decisions.
Wake me up when an agent swarm can write gcc in a box sealed from the internet.
I think we are a long long looong way from AI designing 'terawatt fusion plants'.
In the big scheme of things is it really that expensive to verify it if a lean proof is generated? The agent itself will likely have already verified such Lean code before calling it "done".
There’s a Twitch-streamer Tsoding who programs on C for fun calling it “recreational coding”. Maybe human programming will be a form of art in the future, virtually useless for big corporations to make money. I don’t care, I love it anyway.
Normatively, this ought to be true. Descriptively, this is of course false. Our entire society is organized around producing commodities, typically by consuming people as inputs.
I don’t understand mRNA vaccines, but they are useful to me. The output of an LLM could at least in theory be the design to a major technological advancement, and its implementation with automatic tooling. We don’t have to understand that for it to be useful.
This is like saying software is useless if the user doesn't read the source code of it. This what happens >99.99999% a person uses software. People want to be entertained or have their problems solved.
That's nonsense. Of course you study them to meet demand.
What the article really says is that we're going to need much smarter mathematicians. That is not possible for puny meat-brain humans. Humans are close to their ceiling. AIs are just getting started.
In practice, we're probably going to hit that limit first in IC design. I once went to a talk by the Intel engineering manager who headed the Pentium Pro effort. That was the first superscalar x86 CPU, and it took about 5,000 engineers at peak to design it. Getting that many people coordinated on one thing was a real achievement. Then Intel stayed with minor tweaks on that design for years.
We're soon going to be seeing designs of even greater complexity cranked out by AIs. No human will understand them at the gate level. Reading AI-written programming language code is bad enough. Reading AI-written Verilog may be beyond human comprehension, except in small sections.
Will all rich people in all countries collectively agree to be moved into tiny apts same size and give up their luxuries?
If not, how do you define who gets what? Who gets to live in the fancy house, etc?
This is beyond naive.
[1]https://online.ucpress.edu/elementa/article/doi/10.1525/elem...
The companies do not do this, because equal distribution of resources is poison to capitalism. The world where everyone are fed, and their basic needs - shelter, medication etc. - are taken care of, does not have trillionaires. Possibly not even billionaires.
It is not profitable to give people humane standards of living. You work a lot harder when you see the homeless on the street on your way to work.
That doesn't mean the transition will be smooth. It doesn't mean that certain classes won't be worse off. I can certainly imagine programmers being among the losers. And there certainly are futures where we slide in dystopia. So let's try building the good futures. Alea iacta est.
I agree with this statement, though I think this Brave New World is incredibly exciting to some and dystopian to others. The former group might include those that value the intellectual process above financial reward and status.
The flip side is there are many people, especially in tech, where their area of expertise has evaporated along with their lucrative and previously high status careers. It used to be possible to have a technical job by essentially following recipes and it turns out AI is far better at that than a human.
The linked article lays out why human understanding of mathematical models remains essential and I think the same applies to software. We're gonna need more software engineers who are able to think critically.
You can't simply prompt a model to be "better" when "better" isnt even properly defined
In a world with ASI, having that capacity is vital.
This is an unfortunate example to choose, being as it is entirely confounded by the canonical illustration of the https://en.wikipedia.org/wiki/Law_of_triviality.
If you believe that math is discovering, it's natural to think that all of that AI math already exists and is just waiting for us to find ways to discover and understand it.
Don't write us out quite yet. :)
Stuff! Inscrutable stuff, maybe, but that's not "doing math for math's sake."
Yes, why not? And, of course, AI doing math for AI.
We may not be needed forever...
It becomes another abstraction, really. As long as we can use it for something useful, it's still valuable.
If the LLM is operating at such a high level that it never actually constructs a useful product for humans to use, then how will that be good for humanity?
If you replace "LLM" with "mathematician" than this is the state of the world today. Stuff like Galois theory is beautiful mathematically, but what has it constructed or enabled for you and me?
A cpu (the physical thing that sits in your mother) is not an abstraction, what are you talking about
We want to understand. Quantum physics, mathematics, how stuff works. Ants don't.
That want is not a given, not all of us have that drive. In fact, very few of us have it. So far though, it seems multiple disconnected civilizations learned to keep that trait going instead of suppressing it and focusing only on practical ant-like activities.
This makes no sense whatsoever.
We need more, because there will always be far more difficult problems yet to be discovered and solved, and that means, we certainly need expert humans to define and verify them.
If you cannot even explain the problem you are facing, not only you don't understand it, but you certainly would not be able to know if the AI solved your problem correctly.
And this will be true for how long? 3-4 months?
The risk of liability is a social problem that is far more difficult to be solved with technical solutions even with AI.