03/06/2025
> 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.
03/06/2025
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.
03/06/2025
> 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.
03/06/2025
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.
03/06/2025
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.
03/06/2025
People memorize the entire Quran as a religeous obligation...
https://en.wikipedia.org/wiki/Hafiz_(Quran)
03/06/2025
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.
03/06/2025
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.
03/07/2025
> 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.
03/06/2025
And on the flipside, 廣čØäøå¦ę·”å¢Ø[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" [ęæåøé] Magistrates handbooks, Author Zheng Duan [éē«Æ] (compilation), Early Qing Dynasty (1644-1796)
03/06/2025
Yāall 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. Doesnāt necessarily mean most people will, but itās very likely that most people will.
03/06/2025
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.
03/06/2025
In the setting you describe I think it will _reduce_ the errors to 20%
03/06/2025
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
03/06/2025
03/06/2025
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
03/06/2025
That's more a comment on bureaucracy than it is about memory
03/06/2025
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):
āWriting is like a drug that weakens memory.ā
and: āDo 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.ā
Confucius (Analects 2:15):
āLearning without thought is labor lost; thought without learning is perilous.ā
Lao Tzu (Tao Te Ching, Chapter 48):
āIn the pursuit of learning, every day something is acquired. In the pursuit of Tao, every day something is dropped.ā
Jesus (Matthew 16:26):
āFor what shall it profit a man, if he shall gain the whole world, and lose his own soul?ā
Muhammad (Hadith, Sahih Muslim):
āThe 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.ā
(This Hadith symbolically warns against excessive reliance on external consumption diminishing spiritual clarity and internal balance.)
Rumi (Masnavi):
āThese outward forms are but dust and air; Seek the reality beyond appearance and form.ā
Krishna (Bhagavad Gita, 2:42-43):
āThose 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.ā
03/06/2025
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.
03/06/2025
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.
03/06/2025
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"?
03/06/2025
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
03/06/2025
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 Iām not alone in this!
03/06/2025
Sounds like you rediscovered the long held practice of "Rubber Ducking".
03/06/2025
Haha, thatās a good point. I think the main change from using these (reasoning) models is that Iām more cognizant of my thinking process, rather than there being a novel technique.
03/06/2025
eh, writing it out is closer to "proto design document" or just plain ol "whiteboarding"
03/06/2025
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.
03/07/2025
Have you tried? Maybe it can.
03/07/2025
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...
03/06/2025
But now the rubber duck can talk back and also on occasion hallucinate and lie to you to confirm your delusions.
03/07/2025
I guess I'll stop using this really useful tool because someone else is incompetent.
03/06/2025
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.
03/06/2025
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.
03/06/2025
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.
03/06/2025
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.
03/06/2025
It's your reasoning daemon running in the background.
03/06/2025
Maybe demons were those who learned to cogitate better and ended up being the Ted Kacynzkis of their ilk :)
03/06/2025
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.
03/06/2025
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.
03/06/2025
You computer scientists cosplaying as cognitive scientists really never took single psychology class, did you?
03/06/2025
I believe the quote here is "Every time I fire a linguist, the performance of the system goes up."
03/06/2025
I'm slightly terrified I'm mot consciously trying any of these metacognitive things to "self improve" myself.
03/06/2025
do you want to try that again but being more constructive?
03/06/2025
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.
03/07/2025
Consider the recent advancements in reasoning models, Iād say your method is a bit inefficient ;)
Itās 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 ;)
03/06/2025
Super interesting
> As I read over practice questions, I spoke aloud
This is also something thatās 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
03/06/2025
ā¦ some of yall
03/08/2025
This is fascinating. Reminds me of how people improve at chess by studying a chess engine's suggestions.
03/06/2025
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.
03/06/2025
My god, thatās literally just thinking aloud, doing the rubber duck thing.
You are not āemulating R1ā, you are talking to yourself to make sure you understand the concept.
Which is fine but donāt act like AI is making this part of life better in any way with this example. Nonsense
03/06/2025
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.
03/07/2025
Humans train AI, now AI trains Humans
03/06/2025
At this point I can't tell from the title whether it's a self-help psychology fad or an LLM paper.
03/06/2025
Soon enough, they'll have LLMs that reason purely from the first principles of The Subtle Art of Not Giving a Fuck.
03/06/2025
That would actually be really useful
03/06/2025
How much has our knowledge of AI training techniques helped to discover how to train people to think better?
03/06/2025
We've had knowledge of how to eat better to not get extreme scenarios like obesity and look at the effectiveness of that. Until you have a pill that makes you think better only the motivated will do it, and in this case the motivated could already do it.
03/06/2025
You seem to imply the 'motivated' can not improve.
I'd say the motivated often reap the rewards of innovations more so than the average, as they were pushing the boundaries in the first place.
Having a dishwasher or a robot vacuum does not make me lazy. It allows me to do more productive things.
03/06/2025
I guess I was focusing on the most obvious cases. If a pill makes everyone skinny you'll notice much more that there's no more huge people than that skinny people's body fat went from an average of 13% to 11% or some such. I do agree the motivated have more likelihood to improve, but on a societal level I tend to think first of raising the bottom of the group. I agree with your points.
03/06/2025
As someone with an educational background I actually often ask myself the opposite: Why don't AI techniques almost never seem to use the knowledge we have about human learning to train better AI?
03/06/2025
Maybe an area worth exploring, if you think there's something to it!
03/07/2025
Wasnt that exactly what they tried to do for the first 30 or so years, but it wasn't very fruitful?
03/06/2025
[dead]
03/06/2025
So far, I don't think we found anything interesting, yet.
03/06/2025
"models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions"
One of the parts most worth a replication study.
03/06/2025
I sometimes see these reddit threads of people talking about the experience of having an internal monologue. I have no such monologue, at least not one that is accessible to the part of my mind that calls itself 'me', but I have often wondered if that monologue is something like a 'chain of thought'. I feel like maybe without access to that 'idea feed' maybe my planning and executive functioning is less effective than some other people. I do find myself quite more effective with those sort of tasks when I do a little 'chain of thought' notepad.
I also suspect I spend less time ruminating and second-guessing myself and other anxious behaviours that I imagine would come with having someone talking in your ear all day, but that's probably off topic.
03/06/2025
You never form thoughts in your mind in a linguistic way? Can you read a sentence and be aware of it as a sentence in your mind, or are you unable to do that?
I don't doubt you or anything like that, just very curious. As someone with a very strong internal monologue, it's hard for me to imagine not having one.
03/06/2025
No. I don't think in 'first person' words at all. I might consciously compose a phrase if I'm doing something like writing a poem, which is more akin to arranging a puzzle or something or I might recall words of a conversation someone said to me and i do think of song lyrics if I have a song in my head, but there's no voice in my head and it's absolutely baffling to me to imagine otherwise, as I imagine it is for other people to imagine my situation.
When I read a sentence in a book I don't hear any kind of narration or anything, but I do assemble a 'scene' of images, sounds, facial expressions, motions, etc. not like a movie, but more like a series of small related ideas if that makes sense?
I find that I understand dialogue and characters in books much better when I listen to an audiobook than when I read, not sure if that's related or not.
I am a relatively intelligent successful professional, but I wonder sometimes if I am missing some processing hardware other people have access to.
03/06/2025
Not missing hardware, maybe just different. Those of us with an internal monologue will still say there is thinking that happens before that is voiced, so maybe you're still thinking, maybe the same hardware for an internal monologue is there, I guess, but it just isn't being voiced maybe? It doesn't sound like you are missing anything important, given you are still able to reason just fine.
Anyway thank you for answering!
03/06/2025
For me it has linguistics components for sure, but it is many in parallel and a lot less 'linear'.
Where inner language most certainly comes into play is in the 'output' phase, be it spoken or written, as serialization is required there, but to be honest that often feels like a projection or even a reconstruction with an inherrent sense of loss as the inner is so much richer and nuanced.
That is not to say linearization has no merits. Even if it loses so much it forces consistency and rigor in the lower dimensional reasoning.
03/06/2025
If you were in a debate, and had to think through your reasoning and arguments, refining your argument, not the words you want to say but the substance of what you will say, would there be an inner voice in that situation?
03/06/2025
Best I can put it not one but many (not voices but) narratives at once, and shared with stuff that is not linguistic, but more like analogue waves of concept activation.
In a fast debate you have no time for inner single dimension explicit reasoning. It flows more directly.
Compare it to running on uneven terrain. Somehow decisions are being made where to put the next foot, but most of the time not after an introspective deliberation (although these would occur sometimes in e.g. very tricky rock climbing moves). The steering of where to go occurs at a much higher level, and the steps flow from higher level coarse grained directions.
Now in hindsight you can go back and analyze a recording or memory of the debate and see how it all makes sense as if it was rationally reasoned at every step of the way, but during the debate it most certainly is not an "inner voice", homunculus reasoning before consciously uttering the next phrase.
03/07/2025
Thank you for answering. I can't wait to this stuff starts to get examined in more detail and more experiences shared, I think we are only at the tip of the iceberg.
03/06/2025
In my case, no. There is no conscious 'refinement' of speech options like a Disco Elysium scene or something like that. I mostly just say what I think. There is definitely some impulse control type function to keep me from saying, for example, offensive things, but it's not in the form of myself telling myself 'dont say this thing'. It's usually in the form of an intrusive memory of a facial expression or negative emotion that I associate with that branch of conversation, if that makes sense? Same end, I think, just without a verbal component.
I have moderate inattentive type ADHD that manifests as me being hyper focused on specific sometimes minor things and failing to effectively plan larger picture things and often results in poor executive function outcomes. Maybe that's part of it.
03/07/2025
Interesting. Thanks for your answer. For what it's worth I likely have the same type of ADHD but also have a strong inner monologue.
03/06/2025
Non-chain of thought models just "come up with" words to say. It's probably a bit like that.
03/06/2025
So do chain-of-thoughts. The "come up with words to say" aka. selecting the best-fitting token to continue the sequence, is more-less exactly how my inner monologue feels. It sits at the boundary of the conscious and the unconscious, and from the point of view of my own introspection, the words just appear, complete with an emotional undertone of whether they "feel right" and how much.
03/06/2025
Genuine question, how does multi step reasoning work for you then? Like eg if you have some math problem that's trivially to solve individually but needs multiple steps, lets say 16 * 3 + 5? How does 16 * 3 = 48 land in some 'register' of your brain (short term memory), so that you can then add 5 to get to 53? Maybe 16 * 3 + 5 is to easy for you and you'll just 'see' it but the question still stands, just choose a more complex problem.
Isn't the same meta process at play when thinking about more fuzzy topics?
03/06/2025
Not that poster, but for me it's directly manipulating numbers (for example, "16Ć3 + 5" turns into "10Ć3 + 6Ć3 + 5" into "30 + 18 + 5" into "30 + 10 + 8 + 5" into "40 + 13" into "53"). There's no language involved, though in some cases I might use some spatial reasoning by doing something like associating given chunks of an equation with different fingers.
03/06/2025
I don't really follow. Say, during the in-between step "10Ć3 + 6Ć3 + 5", how do you store and cognize the individual numeric and operator elements?
Surely, even if the arithmetics can be simplified and "lookup-table'd", you are still aware of the numbers in Arabic form or whatever equivalent you're using, right? Or do you somehow have 53 individual blobs swirling inside your consciousness?
03/06/2025
Not poster you replied to but it sounds like we might have a similar internal model.
I store numbers as pictures of numbers, or a geometric representation depending on how big or precise the number is.
Are you saying when you think of the concept of 'twelve plus twelve' you have the equivalent of someone in your head saying 'hmm, well twelve is 2 more than ten, so if I add up ten and ten and two and two I get twenty four?'
That's wild if so.
For your reference, I would follow the procedure above approximately, but visually with numbers that just do the thing that feels right. I think under the hood we're probably doing the same thing, just with a different interface layer
03/06/2025
I can just speak for me, obviously, but yes, that is what's happening. But it's not someone, it is more like me explaining / telling it to myself. Depending on the complexity this can be more or less verbal - the more complex, the less verbal I would say.
03/09/2025
In short, I just do. It's not a linguistic representation of the number 10, it's the concept of the number 10, the same way as (to extend on the mention of another comment) I don't need the word "cat" to know how to interact with a pet cat.
03/06/2025
Language (human) and language (theoretical computer science) are different concepts.
03/07/2025
Not sure if it helps you make sense of it, but at least myself I only verbalize more complicated things. 10x3 for example I just "see" as 30 without needing to "think" or "speak" about it.
If you see a cat walking along the road, do you have to think to yourself "oh, that's a cat" or do you just know that it's a cat without verbalizing anything? It has its own abstract concept, right? Same thing with sufficiently simple numeric transformations.
> you are still aware of the numbers in Arabic form or whatever equivalent you're using, right?
Not sure if you're in a Fahrenheit or Celsius sort of place but if someone says that it's 70 degrees out do you really think in terms of numbers? Or do you just "know" what 70 degrees is without thinking about it?
03/06/2025
It's also very probable that the verbalization the majority does internally is just that - a verbalization of the actual underlying thought process. That is what much of current cognitive linguistics points to as far as I have understood.
(Also a reason why I'm very sceptical that the current LLM approach will eventually lead to AGI, BTW)
03/06/2025
I think you're probably right that the verbalization is the 'interface layer' but why does that mean LLMs can't approach AGI? They also only use words as an 'interface' layer. Underlying weights are vectors in an abstract space.
03/07/2025
With humans, it has been shown that reasoning processes for different aspects of human behavior are completely distinct from verbalization, they work fully autonomous, and the (inner) verbalization comes afterwards.
For LLMs, the tokens (i.e. words) are what the weights are based on, as there isn't other input into them.
03/06/2025
I believe I have an internal narrator but Iām not certain exactly what others mean by that so I donāt know for sure.
However the way I think about math is different than the way I plan my day or other things. In my case, it is very much like I have registers that would hold the result of 16 x 3 in it so I can add the 5 to it later. I have a certain number of registers and with effort like repeating what Ive already solved I could temporarily create more.
It also feels somewhat physical, as if the register is an actual box or has a ālocationā or like Iāve put the answer down on the desk like a part of something Iām building. Perhaps not coincidentally I am one of the many people who have a ācalendar shapeā for the months.
03/06/2025
I speak out loud or write on paper, or just do it a tiny bit slower and sometimes have to redo steps when I forget a result.
03/06/2025
I do have an internal monologue. I can also think in pictures and I can also think in terms of neither, just pure raw thought.
I would say most people are like me. They have 3 modes of thinking and they probably have a primary mode which they favor. I favor none and go into all 3 depending on whether Iām reading, writing or doing something else.
The second bigger group has only one primary mode of thinking. The internal monologue. They can only think in terms of an inner voice and this inner voice is so powerful I often encountered people who think this inner voice is the definition of thought. They assumed thinking was COT.
The even rarer versions you get people who assign colors to numbers or people who canāt even perceive to think in pictures. Youāre the first person Iāve encountered who canāt even have an internal monologue.
03/06/2025
I think you'd be surprised. I never knew that internal monologues were a thing until there was a HN thread about it and a lot of people had them and a lot of people didn't.
I always thought it was something that we did in TV shows or books to give you a sense of what a character was feeling, I didn't know this was an actual literal experience people had.
I can certainly have an internal monologue, in the way that you could put on a puppet show. I can conciously think to myself 'self, this is self. Clean your car out' I can form the feeling of those words in my head. But there's nobody 'saying' them if that makes sense. I'm playing back a design of my conscious self.
03/07/2025
There's a fascinating thing called aphantasia where people can't picture things at all in their mind, but such people are able to lead normal lives and may never realize there's something different. This feels like a similar concept but for imaging speech.
https://en.m.wikipedia.org/wiki/Aphantasia
That said, most of my thinking is not done in the form of a linear monologue where I "talk through" steps to myself.
03/06/2025
So you cannot think in language? Sounds kinda scary to be honest.
03/06/2025
True, but a problem is that self-improving AI leads to a somewhat troubling mode of thinking. AIs switch to an internal babbling type language that makes no sense but clearly still conveys meaning to the AIs, then think in that language (if it's a language, though not sure what else it could be) and then produce correct results.
Worse, when you use multiple agents to get AI LLMs talking to one another, all AI agents switch to this internal language and they make progress despite no human understanding what hell is happening. This seems very bad.
Illustration:
> How many r in strawberry?
I'm asked how many r in strawberry. I can just spell the word and a;dklsjaw; a;ewjraqwpeouypaads;lq qepwiouryaqeopw qewrpoiuyoiauysdqw145124rfa.nkjlwh ;45a8345a894ya4a q4p58q45jaq;lkjas;dlfkja;j
<answer>There are 3 (three) r's in strawberry</answer>
03/06/2025
Iāve heard this described as talking in āNeuraleseā. It seems plausible that this will be the most dense language for model-internal dialog (or presumably inter-LLM dialog assuming they share the same weights).
You will penalize this inasmuch as your alignment strategy depends on Deliberative Alignment. But at some point I assume that will come with a real capability cost as Neuralese can be more conceptually dense.
03/06/2025
They are not going to invent a new language by themselves, they by definition can't even "think" in terms of languages they haven't seen. It does not occur to them that the language they use may be suboptimal. And surely, any better ways of thinking can still be described in English. It seems more likely there will be a gradual transition from us teaching LLMs how to reason, to LLMs being able to actually gobble and process enough data to learn more effective ways to reason, which it can then "teach" us. But that's just the LLM reflecting the way it was trained and aligned.
03/06/2025
> four key cognitive behaviors -- verification, backtracking, subgoal setting, and backward chaining -- that both expert human problem solvers and successful language models employ.
Based on what have they claimed that such methods are used by expert human problem solvers?
03/06/2025
Once you suspend disbelief of AI all of this pseudoscience becomes just as plausible.
03/07/2025
I'm actually interested whether there is some larger explanation on these study methods. Maybe there's something worth integrating with myself and get more efficient at learning.
03/07/2025
I think everyone wants that but there really is no silver Bullet in explaining or understanding how to make this technology more effective or more accurate. The companies themselves can radically change the models, behavior intentionally and unintentionally. There is no guarantee that any insights that you gain today will be applicable in the future.
03/06/2025
Does this also mean giving better system prompts that encourage this behaviour also substantially help?
03/06/2025
In my experience models aren't very good at following such prompts. Smart "non-reasoning" models like Claude 3.5 could, but would generate so much text when thinking they ran out of context.
03/06/2025
``think''
In the abstract they use different characters for double quotes here.
03/07/2025
That's LaTeX syntax for opening and closing quotes. Though their rendered paper doesn't render them that way for some reason!
03/06/2025