> Bottom line, Tao is pushing for human understanding as the primary goal
100%. This applies to SWEs/math folks/etc. I do infra and I see many SWEs take their hands off the wheel. When they encounter perf issues they ask their agent and agent says GC and they say GC. It's rarely GC.
Now we might be well past the point where we need to remember the kubectl flags for rollouts etc. But basic human understanding of what their bots are doing as a goal has never changed. Humans are still liable for when bad things happen, and that hasn't changed over the roller coaster the last 5-odd years have been. LLMs, as astonishing they are at Navier Stokes, are still eminently capable of nuking your filesystem and saying "I can now see that that was wrong" with zero regrets. If you can't understand you can't sign off.
> math is the most pure expression of human understanding
This I don't know about. I think math acquires meaning when it contacts reality: like an iota is pointless until there's some circuit that it explains. Abstract math can diverge from that and can become an exercise in playing with symbols for their own sake.
Explaining nature causally using mathematical models isn't "the primary goal" of humanity. STEM people tend to have a weird misconception there, probably stemming from their misconceptualization of the humanities.
This in turns leads to this odd idea of "AI will think for us". That's pure (and pretty obvious) insanity. Logically minded people encountering it should ask, where the error in their reasoning is.
I have two issues with the implication of this statement:
1. Meaning is inherently subjective. Reality is just a canvas on which sentient beings create their own meaning.
2. There are many, many examples of where “playing with symbols for their own sake” have yielded deep insights. There’s actually some implicit structure (eg the structure of logic) that is intrinsic to the universe we live in.
After talking to a mathematician friend I can assure you that contact with reality is not the main goal of abstract math. It is mental constructions that have logical consistency and probably this is not the perfect definition either. It is somewhat of an art which is rendered in the logical mind. However physicists (me) and engineers will align with you.
The questions we asked were originally not about math. They were about a thing that we invented maths for to do or explain, a question that existed, because it touched us in some way that was already real to us. There is nothing that would not allow this to happen in the future. All this requires is attention and connection to the world around us. The maths required to answer our questions can be done and developed by something else.
To me, all you need to believe for this to be true is to agree that understanding maths is also not a stated requirement of reality to get a thing, if something else understands the maths (or something that does the same job). This is demonstrated by billions of people who do not understand maths and get things that, currently, require other people to understand the maths.
But the last part is entirely optional as it pertains to reality. That's just the best we can currently do (and in some important sense it is holding us back as a species, and in some other sense doing the opposite).
If that goal is understanding math, you certainly will be able to understand maths, more than ever before.
If that goal is something that required you to understand maths first in the past, you won't have to do that anymore.
I respect Tao and I believe he's trying to think deeply about the issues, but a lot of his thinking seems to revolve around preserving the current roles and prestige of mathematicians, and also makes a lot of assumptions about the capabilities of AI years or decades into the future.
This is not me being snarky, but all meaningful questions can be settled empirically---e.g., "what happens to my body if I jump off the cliff". But empirical trials have a cost (time, money, irreversibility etc.) and we model and predict because it's cheaper than the trial.
The idea that all meaningful questions can be answered empirically is known as "verificationism" and is philosophically quite dubious.
This is a logical positivist view, that not everyone agrees with.
> Simply put, if we don't understand the answers we won't know what the next question should be.
It won't matter, because it won't be us who will be asking the next questions anymore. Whether in math or anything else.
And no, domains that require real-world validation against physical ground truth won't save us, because AI gets to have the same inputs as we do (or better, if using specialized hardware), while beating us at reasoning.
And GP's likening this to previous massive economic shifts due to automation isn't really helping in any way, not anymore, because perspective won't feed us when we're hungry, and just as importantly, this one will affect every single field of human activity, so no one has any answers as to what the future will really hold for us.
That's the frontier labs' preferred narrative while they themselves are still hiring hordes of human "Account Associates", "Android Engineers", and "AI support engineers" instead of automating those jobs as a show of their AI strength. Of course it will be "us" asking the questions, because "AI" are computer programs, and humans build the computers and choose what computational tools to use for any application.
This is what it feels like to be disrupted. It's not the end of the world. You consider the evidence, ponder the path forward, and adapt. It's what humans do and their superpower. It doesn't have to be a negative thing, even if it is dislocating.
We don't need to be saved, there is plenty of agency to go around. Just grasp the opportunity and forge ahead. This sort of pessimism is self-defeating. Humanity has dealt with this before and come out on top, this time is no different. AI is being wildly oversold.
Tao is being entirely rational. He maybe has more to lose as anyone, but he's getting down to brass tacks instead of jumping at shadows and imaginary boogeymen.
What if AI knows better than us?
It might be that P=NP and the algorithms are handed down to us. We can apply them without understanding why P=NP, and we may never be capable of understanding why.