Hacker News Reader

Comment by owenpalmer | Hacker News Reader

> four key cognitive behaviors -- verification, backtracking, subgoal setting, and backward chaining -- that both expert human problem solvers and successful language models employ.

As we make AI better, perhaps we'll inadvertently find ways to make HI (human intelligence) better too.

I had a personal experience with this when I was studying for an exam recently. As I read over practice questions, I spoke aloud, replicating the reasoning methods/personality of Deepseek R1. By spending a lot of time reading long verbose R1 outputs, I've essentially fine-tuned my brain for reasoning tasks. I believe this method contributed to my excellent score on that exam.

View on HN

This is a well-known approach: verbalizing your thought process (either by speaking aloud, or by writing) is something that's long established as a good tactic for making sure that you're actually thinking through something, rather than glossing over it. Ironically, I've seen people bemoaning that use of AI will rob people of that.

I agree that there's potential here, though, and do genuinely hope that we find ways to make human intelligence better as we're going about AI research. Even pessimistically, I think we'll at least surface approaches that people use without thinking about, which is on its own a good thing, because once you know you're doing something, it becomes a lot easier to train yourself to do it better.

This is one of the secrets of the top British universities. They do a lot of debating in small groups. Even their papers are read out loud

> Ironically, I've seen people bemoaning that use of AI will rob people of that.

There's that quote from Socrates, recorded by Plato:

> For this invention will produce forgetfulness in the minds of those who learn to use it, because they will not practice their memory. Their trust in writing, produced by external characters which are no part of themselves, will discourage the use of their own memory within them.

Pretty much every single spiritual philosopher has said some version of that (I'm writing a book on this subject right now, heh):

The Buddha (from the Pali Canon, Vinaya Pitaka, Cullavagga 10:4):

“Writing is like a drug that weakens memory.”

and: “Do not go upon what has been acquired by repeated hearing; nor upon tradition; nor upon rumor... But when you yourselves know: 'These things are good; these things are not blameable; undertaken and observed, these things lead to benefit and happiness,' enter on and abide in them.”

Confucius (Analects 2:15):

“Learning without thought is labor lost; thought without learning is perilous.”

Lao Tzu (Tao Te Ching, Chapter 48):

“In the pursuit of learning, every day something is acquired. In the pursuit of Tao, every day something is dropped.”

Jesus (Matthew 16:26):

“For what shall it profit a man, if he shall gain the whole world, and lose his own soul?”

Muhammad (Hadith, Sahih Muslim):

“The worst vessel to fill is the stomach; sufficient for the child of Adam are a few morsels to keep his back straight. If he must fill it, then one-third food, one-third drink, and one-third air.”

(This Hadith symbolically warns against excessive reliance on external consumption diminishing spiritual clarity and internal balance.)

Rumi (Masnavi):

“These outward forms are but dust and air; Seek the reality beyond appearance and form.”

Krishna (Bhagavad Gita, 2:42-43):

“Those who are attached to pleasure and power, whose minds are drawn away by such things, have no capacity for absorption into higher states of awareness.”

Sometimes it's good to forget something, commit to written record and let it go. People can carry too much old stuff around in their heads, and it can become burdensome.

Even things like confession, or therapy, leverage this - people letting go of bad things that are hanging around in their memory.

A lot of traditions would strongly disagree with that, myself included. I would imagine you would find a lot of Gnostics who disagree, a lot of "witches" who have been literally burned over the years, and also the aboriginal people of Australia. Well, you wouldn't find so many of them, because those lines of thought have literally been beat out of our human societies in the name of progress. If it's not worthy of being in the oral tradition, it's not worthy of the society, it's not worthy thought. (unworthy thought does not need to be rigorously engaged with.)

Also remember, your conclusion itself is "the devil" - the trap of the analytical mind. :) You will likely do everything you can do avoid the fact that you may be disagreeing directly with the word of the creator as given via various prophets, if you go back to the sources, the command is quite clear, however humans will interpret it: because the command is too simple and terrifying to adhere to. It seems impossible to us that we should indeed, be doing nothing but living in nature in a state of oral tradition and anything outside of that is an unintended state, trusting that energy cannot be destroyed and we are nothing but energy. I don't particularly like it either tbh, hence I'm writing a book about it.

And on the flipside, 廣記不如淡墨[0], lit. "a good memory is not as good as pale ink", which is these days more commonly translated as "the faintest ink is more powerful than the strongest memory".

[0] "A Record of Learning about Government" [政學錄] Magistrates handbooks, Author Zheng Duan [鄭端] (compilation), Early Qing Dynasty (1644-1796)

Y’all can throw pithy sayings at each other all you like but memory is not the same as understanding, and AI does offer plenty of opportunity for humans to cognitively disengage. Doesn’t necessarily mean most people will, but it’s very likely that most people will.

Especially if we're more concrete: If your job is, say, in administration and what the machine answers is correct enough that in in 8 out of 10 cases you can basically copy-paste it, I'd say it's extremely likely that it's going to increase the amount of errors made.

In the setting you describe I think it will _reduce_ the errors to 20%

No, since the setting is not specifying the initial rate, it might as well increase or staying stable at 20%.

But there are other factors, like, is the amount of outcomes done also changing, thus affecting the absolute number of errors?

Also, does the side effect of disengage the person in most cases means it has side effects like not paying the same attention to what would stand out as a big issue that needs more attention and consideration than business as usual?

And so on

It's an interesting thing to think about. From the way it's talked about, I would predict that AI will enable people who are more cognitively inclined to think in more complex and refined ways; while other people that over-rely on the results would be the ones that decline.

However, research[1] suggests that relying on AI tools degrades reasoning and cognitive ability regardless of your cognitive ability, and may even cause users to stop making their own choices[2].

1. https://www.404media.co/microsoft-study-finds-ai-makes-human...

2. https://www.nature.com/articles/s41599-023-01787-8

That's more a comment on bureaucracy than it is about memory

Classical era philosophers weren't completely wrong about this. They lived in a more oral literary culture where performers could recite entire works from memory.

I don't think anyone today could recite Beowulf from heart. But 1500 years ago that's exactly how it was enjoyed.

Maybe not Beowulf, though it's not to exclude either, by I have no doubt that some people out there can do that for some work.

Look at these people that will declaim Pi digits, just because.

Also there are difference in education over different cultural era, not only through time but also space. I heard that India for example value more repetition, where western culture is more in love with innovation. These are of course nothing like exclusive tendencies.

Now if you look at antic Greece, it's certainly not like everyone would be able to restitute Homer word for word. Actually it's easy to forget how divided in term of linguistic and social classes this societies where, and focus solely on the most renowned figures as if they where all part of a tight social group full of solidarity and genius. Actually even a guy like Hippias of Elis can be both depicted by plato as exhibiting all the tremendous admirable feats of the time, including mastering the art of mnemonics, and yet turned into a clueless bragger that isn't even able to recognize that he just doesn't know how to define beauty.

People memorize the entire Quran as a religeous obligation...

https://en.wikipedia.org/wiki/Hafiz_(Quran)

I'm pretty sure that "recit[ing] entire works from memory" was an example of an amazing memory feat that was far more common in the past, before writing was as common. I do not think that it was a recommendation that education should consist of memorizing literary works.

I wonder if the number of people who could recite works from memory has substantially changed: while surely the prevalence of such people decreased, world population ballooned! Today we have people that can recite many digits of pi, for example.

> from heart.

Is this the case? I was under the impression they memorized the plot beats and filled in details on the fly. Also using set phrases or epithets like "Gray-eyed Athena" to slow down the narration and let them plan further.

I wonder what Socrates would say if your thinking process depends on GPUs owned by oligarchs and all your attempts at solving a problem are tracked.

What would he say if the collective IQ drops by 30 points in case of a power outage?

What would he say if people need a subscription in order to "think"?

I use this method too for programming problems I would normally procrastinate on and offload to subconscious thinking.

Actually writing out all thinking steps helps with ironing out some wrong steps in my reasoning or going in circles due to having limited working memory.

I started doing this more rigorously after seeing how reasoning based AI does reasoning, because it seemed like a useful thinking technique.

These reasoning AI models help me think on a meta level about my own thinking and shows me tools I can use to improve it.

Great to see that I’m not alone in this!

Sounds like you rediscovered the long held practice of "Rubber Ducking".

eh, writing it out is closer to "proto design document" or just plain ol "whiteboarding"

the only reason I haven't written a design doc or busted out the felt tip for my rubber duck is because it can't read.

Have you tried? Maybe it can.

I mean, I'm sure there's some founder out there pitching an AI-powered rubber duck dev productivity tool.

At the very least, someone at Copilot must have pitched a rubber duck avatar as the new Clippy by now...

But now the rubber duck can talk back and also on occasion hallucinate and lie to you to confirm your delusions.

I guess I'll stop using this really useful tool because someone else is incompetent.

Haha, that’s a good point. I think the main change from using these (reasoning) models is that I’m more cognizant of my thinking process, rather than there being a novel technique.

For me, I think this approach works because I can commit the current thoughts to some type of external (to my brain) storage, freeing up space to think about how to further subdivide those tasks.

In general, this is very helpful for when your executive function feels taxed, as it has the effect of coaching yourself.

Thinking out loud is an age old practice and is the equivalent of "rubber ducking" to yourself.

As someone who comes from a long ancestral line of people who talk to themselves while reasoning through problems - it would occasionally prove to be a minor handicap during proctored exams, as internal monologue isn't really the same thing.

From what I have seen from split-brain experiments, I am of the belief that by vocalizing our thoughts, we are more fully engaging both hemispheres of our brain through the auditory pathway in addition to the Corpus Callosum.

Pictures tell me that the language area is dominant in a single hemisphere, mostly the left side, with motoric stuff and thinking about words in the front (Broca's area) and hearing in the back (Wernicke's area).

So you may have to use the SLI bridge again just to make sense of what the other side is hearing.

Me, working through a problem: "So, that means..."

Girlfriend, coming in from outside: "Who are you talking to?"

Me: "I talk to myself. You know that."

Gf: "Oh right. You also whisper to yourself, which is scary."

Me: "Scary?"

Gf: "It sounds demonic."

Which, to be fair... Evidently, my internal monologuing gets quite a bit vocal even with other people around.

Maybe demons were those who learned to cogitate better and ended up being the Ted Kacynzkis of their ilk :)

It's your reasoning daemon running in the background.

When I was working at a Bike Shop, I was standing on one side of the display, talking about what I was going to do and how I was going to do it. It was a very in depth external monologue (I have both like yourself). As I was coming around the other side of the display, the owner happened to be standing there and said, "Wow, that was some kind of a conversation you just had with yourself. I hoped you were able to solve whatever problem you were discussing." He had a big grin and we both laughed about it.

He told me later he too does a lot of internal monologue for stuff as well and was told by some super successful businessman that this is a good thing and a hallmark of successful people so don't be discouraged by it.

You computer scientists cosplaying as cognitive scientists really never took single psychology class, did you?

I believe the quote here is "Every time I fire a linguist, the performance of the system goes up."

I'm slightly terrified I'm mot consciously trying any of these metacognitive things to "self improve" myself.

do you want to try that again but being more constructive?

Those four parts sound like one unified, cognitive algorithm -- having an ontology of the problem by breaking it into subgoals; checking your work properly; thinking backwards to debug a mistake and retrying; and thinking ahead and reasoning backward from the end result. It's all just one algorithm for solving hard problems. A skill that can be practiced, and then it builds on itself.

Consider the recent advancements in reasoning models, I’d say your method is a bit inefficient ;)

It’s equivalent to the LLMs reasoning in the output and not in the latent space before the final output, which gave the rise to the reasoning models we see today. So speaking out loud might not be the best reasoning method ;)

Super interesting

> As I read over practice questions, I spoke aloud

This is also something that’s expected of the applicant in technical interviews

The interviewers want to hear the applicants thought process and how they develop a strategy to solve the problems presented to them as they work them out

… some of yall

This is fascinating. Reminds me of how people improve at chess by studying a chess engine's suggestions.

Exam is interpolation.

Research is extrapolation.

Neural networks are interpolators, they are notoriously bad at extrapolation. For simple example, look at pendigits data set [1]. The test part of pendigits is taken from different "writers" than train part and neural networks aren't that good at it.

[1] https://archive.ics.uci.edu/dataset/81/pen+based+recognition...

Humans do extrapolation all the time.

My god, that’s literally just thinking aloud, doing the rubber duck thing.

You are not “emulating R1”, you are talking to yourself to make sure you understand the concept.

Which is fine but don’t act like AI is making this part of life better in any way with this example. Nonsense

I think you're missing a key point, which is that by extensively reading the R1 outputs, I'm able to observe how R1 thinks about things, which I can then replicate.

There are good ways and bad ways to think aloud, R1 just gave me a large set of examples of doing it the "good" way.

It's uncommon to read hundreds of paragraphs of a smart person's internal reasoning process. Usually we're only able to read the final results of their thoughts.

Humans train AI, now AI trains Humans

Reply on news.ycombinator.com