kenferry

See https://arxiv.org/abs/2607.13562. AI tends to make people much more confident in their answers. I worry about this for myself, and therefore try to limit my use of AI to either "getting things done" or idle queries. For developing real thought, I think of it as a very political underling who wants you to think it's valuable.

This paper doesn't say ANYTHING about _capacity_ for critical thinking, and by using the words "critical thinking" yes it makes it sound like "more is better", but look at what they're actually doing. What they're measuring I would much more call "oversight".

If I want to know when Obama was born, and check wikipedia and believe what it says, by their measure I am practicing less critical thinking than if I found independent verification. Independent verification is the main way they measured critical thinking. They are NOT measuring that I thought about how much I could trust the result, and decided that yes I could.

And, well, yes, they're right. But for most purposes, wikipedia is pretty credible for something like that, and it's just a waste of time to follow up on it. What's important is that my calibration is correct.

I'm not even saying I disagree with the idea that there is likely a problem here, with people offloading thinking. But I'm thinking critically about this particular study, which seems totally useless.

Hoo, not been my experience.

AI has an obvious issue with incorporating how happy users are in its responses. It LOVES to "take your argument seriously", push back lightly, and then tell you you are really smart and make amazing points. If you start with an argument out of left field, you will shift the Overton window very easily.

I'm either not seeing something or I do not see the point of this study.

Amount of "critical thinking" is basically amount of oversight, no? What's interesting is when the amount of oversight is inappropriate, leading to overconfidence in bad results (or, if too much oversight, lack of efficiency).

This paper does not seem to discuss results and correctness at all, ONLY the amount of critical thinking.

I know it's "too clever", but if I'm writing code no one else will ever see,

  while (i --> 0) {
      // use array[i]
  }

easy to read (if not understand), and correct!

amazing

> It's June 16, 1987. Ted Uhlemann ("iczer") plugs an Apple IIe's 300-baud[1] modem into a phone line his mother had just given him for his birthday, running Magic City Micro-BBS (written in Applesoft BASIC). He names it SDF-1 after his favorite anime

Dude was into anime in 1987?? Truly ahead of his time.

Well, the point of the _computer_ formalization is that it's checkable.

An error in the formalization would be a bug in lean. Which, as we know, certainly seems plausible (I've mused on if these giant proofs are incentivizing LLMs to find lean bugs to exploit), but at least it's a smaller surface area.

> There is one other aspect of the situation that I would like to emphasise. After two weeks of very hard work, I have managed to understand only a small fraction of [TC26a]. When I ask Fable about the paper, it seems to read and understand the whole thing in just a few minutes, even starting from a fresh context. I know that some readers may be uncomfortable with my anthropomorphic terminology, but I believe that “understand” is accurate, in the sense that the LLM seems to be able to correctly and lucidly answer all kinds of questions about the paper, including about some very fiddly technical details. That is exactly how I would tell that a human colleague has “understood” something. This chasm between human and LLM “speed of thought” is unsettling to say the least.

This is very odd, because it just doesn't match my experience with these systems at all. Yes, they do generally SOUND confident, but code (say) that's hard for me to understand is usually hard for the machine to understand as well, and while it is more indefatigable than I am, it's not brighter.

This mathematician is in a very good place to evaluate whether Fable gets it or not, but it's weird and counter to my experience.

This is bonkers take.

If you read this article, the author is basically saying “the way this worked before was very opaque and probably almost no one figured it out, but _I_ knew how it worked, and so the change is bad for me”.

Do they have a good point that they have a valid use case? Yes. Do they have the only use case? No. Did Apple do this to try to make things better? Yes. That doesn’t make them “evil”, that makes them trying to make things better.

I'm not saying Tao is right, but (a) he knows a LOT more about about what's actually important in research math than I do (and probably than you do) (b) he's not just having a knee jerk reaction. Tao has been a major leader and advocate for computer-assisted math, including AI, and I don't think that has really changed. He's advocating for how to do it well. He's far from an AI skeptic protecting his turf. Is there any other person as well-versed in research mathematics AND AI-math as Tao? Maybe a few, but not many.

The dude keeps a continually automatically produced document of his takes on AI. He's not a skeptic overall. https://teorth.github.io/tao-web/ai-views.html . In this interview the interviewer describes him as an evangelist for AI in math and he doesn't disagree. https://teorth.github.io/tao-web/ai-views-interview.html

It seems like your inclination is to discount his experience with research mathematics. I agree with you that that's probably what Taleb would do - Taleb has a giant ego and basically operates under the assumption that everyone else has the same giant ego as he does but is (a) less honest about it (b) dumber than he is.

Beautiful - the site _seems_ like an insane amount of work. Did you have any tooling to make that better?

The chairs are also an insane amount of work but hey, you get chairs!

it's not hard to accept in general, but have you been following Tao for long? The combination of how virtuous he is with how informed he is with how smart he is is pretty special. He's also far from a luddite. He's been a leader of the computer-math charge.

Questions of AI policy are all hard and anyone who confidently says they know the right thing to do is too confident. Tao's fallible like anyone else, but yeah, as voices to pay attention to go he's very good.

How much do you they think they paid? By home improvement standards, might not be out of line at all.

I agree it just feels weird to spend on “software” for home improvement, but sometimes it may make a bigger difference than the hardware!

This isn't me being pedantic, actual question:

Are we talking about whether nature is ever at equilibrium, or about how stable that equilibrium is?

A ball on the crest of a saddle is at equilibrium, but it isn't a very stable equilibrium.

Basically, are you saying the reality is that nature is always in flux, or that it doesn't take much to knock it toward a different stable state.

Chickens will rise again!

Hm, neat! File a bug!

Not with an iPhone pass you couldn't. Apple renders the barcode, it's not just a big image you pass in.

I'm not quite sure what you're getting at, but Apple knows where the barcode is already. The data for the barcode is given as text, apple draws the barcode.

Maybe what you want is for the screen to support turning up brightness for just one part?

Er, I'm not sure what exactly you're getting at, but the stop market drop was not because of his value-add to the company, it was because of the signal it sends by the CEO quitting.

So… I guess you could make an argument that that merits higher pay, but it'd be saying it's due to his blackmail power.

This article is doing comparisons with the same subjects and lighting…

I… guess? I'm not sure why this was posted. None of these changes seem important.

A lot - title could probably use editing. The language has no heap, only stack memory, so the only deallocation is returning from a call stack frame.

> Everything looks like a problem solvable by reduce. (I'm often wrong on that, but I quite enjoy learning why by trying).

I agree, but I think that’s kind of the objection. It can be tempting to write your code as little brainteasers but…

I’m not a fan of this kind of “fancy” optimization anyway.

It’s too fragile. I may make some innocuous change, now the compiler cannot recognize the pattern and performance falls off the cliff.

I’d rather have the reliabile performance than the absolute fastest possible result. Then if there’s an issue I can catch and fix it reliably with profiling, not deal with a heisenbug based on whether the compiler can match the pattern.

That's true enough, but I don't think the conclusion is "don't regulate".

Try it, iterate, figure it out. It'll be painful but so be it.

I think the big difference here is that it's opt-in, off by default, and also not included with shared images by default.

This is a feature because you use it when you WANT to be able to demonstrate an image wasn't altered / was yours, like for a journalist.

This is an actual question, not a dig: what do you mean by non-removable?

You want there to be separate lines for "how long do you want your stems"?

Looking forward to your 18 SKU product matrix of decisions people have to make that they don't know what answer to choose.

The materials cost between those buns is practically nothing. OP is reaching.

I don't know. AI are basically trained to tell you what you want to hear, including when they're disagreeing with you. If you come off of your conversation with AI with a stronger belief than when you went in, is that merited?

"AI Advice Suppresses People’s Willingness To Say “I Don’t Know”, Even When The Advice Is Wrong And Accuracy Is Incentivized" https://osf.io/preprints/psyarxiv/5y6m4_v1