▲ 12 points

Ask HN: Fellow neurodivergents, do you pair better with LLMs than neurotypicals?

by thoughtpeddler·2y ago·27 comments·view on hn ↗

It’s recently dawned on me that my neurodivergence might explain why I seem to get more out of AI as compared to my neurotypical peers.

I find that I can more intuitively ‘co-reason’ with LLMs, subconsciously emulating how they operate, so that I can more effectively steer their attention into the appropriate semantic subspaces to elicit more accurate responses.

Do others here, whether on the spectrum, with ADHD, or similar traits, also experience this? Do you find it easier to ‘pair’ with LLMs compared to neurotypicals?

Might attributes like our systematizing mindset and literal thinking be the enabling differentiators?

Going a step beyond: are we, in the grand timeline of civilization, uniquely evolved to fuse with this new type of intelligence?

27 comments

I'm a schizotype and a maintenance programmer among other things.

The maintenance programmer in me is highly skeptical of things that almost work.

I've seen transcripts where people talk to ChatGPT and it seduces them and they get giddy and it is like watching a "meet cute" in a movie and you can see somebody having so much fun interacting with it that they don't really see that what it says doesn't really make sense.

That's not me.

I love chatting with chatbots about old sci-fi books by people like Smith, Anderson, Niven and such. I've only met two people in my life who are better at sci-fi chat than a good chatbot.

Lately I have been asking Copilot for help with maintenance work, it does amazingly well at explaining strange but highly repetitive code such as the stuff the Babel transpiler inserts into code (stuff that is probably all over the training set.) Sometimes it beats searching Google and Stack Overflow.

I am working on several "second brain" programs that use LLMs for classification, clustering and other analysis but not doing any generative stuff right now. I use one of these programs to pick out articles from an RSS feed, everything I post to HN was chosen by that system and then chosen by me twice. I am using another to look through a collection of 250,000 images and another copy of that software to look at about 400 notes a friend made in Evernote. I am hoping to merge these together in the future. These projects are very much about building something that works with the unusual way I think.

A few days ago I was listening to a psychologist interviewed about chatbots, and they explained a very critical issue:

Chatbots will always say what you want them to say, in a tone that you like. They will always affirm that you're right.

In contrast, in a relationship with a real person, they won't always agree with you, take a tone that you like, or constantly affirm that your worldview is correct.

So I would assume you getting along well with chatbots is by design.

I think I get what OP means and I would say by "pairing better" means getting the desired output as in solution to a problem, whether coding or otherwise.

Like as in I understand the cases when LLM would be able to solve something and when it's not. E.g. generating code, using copilot, I can tell by which input will it give me the output that I want.

I understand when it might be off or "hallucinate", so it's not a problem at all. I'll use it differently.

And I'd say my chatting style with an LLM is usually intentionally argumentative for the reason that you mentioned, so it wouldn't agree with me when it doesn't make sense to do so.

I'll formulate my prompts intentionally to avoid biasing it in any specific way.

I have had the exact same thoughts as OP about it, because I can find similarity in how I think and reason to how an LLM does it, a flow of tokens. So I can think of ways to get LLMs to find solutions based on how I myself tend to solve problems.

Yes, you fully understood me! Thank you for further explaining what I'm getting at.

I find myself able to 'co-navigate' a question or idea with the LLM, better than most other people around me, as defined by quality of model output. But it's just now dawning on me that this might be explained by more than me having greater general knowledge, or having a certain personality.

Somehow, I have an intuition for what words or phrases the LLM might have a tendency to overfit on, so I avoid those in my prompts. I likewise know what words to specifically use or emphasize so that I can trigger the LLM to tap into a certain knowledge-pool based on those.

When I see my friends engage with LLMs, I witness the lower quality output they get because they are not deploying the same intuitions in their prompts, and it's hard to teach them.

May be interested in https://news.ycombinator.com/item?id=40766684

Whoa, interesting. What did you mean by the 'notebook tab'? Also, if I understood correctly, your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points. That's actually pretty cool! Your premise is correct that the conversational interface is really just putting the dream-state on rails; the model itself doesn't actually "know how to talk".

I’m using a local frontend: https://github.com/oobabooga/text-generation-webui

With local 8x7b: https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-G... (Q4_K_M)

Requires 35gb+ of regular RAM, RTX card is optional but speeds it up a little. Can’t tell if cloud llms ever exposed this “notebook” mode, I’m in a sanctioned country.

Notebook is basically a raw interface to an llm. No system prompt, no bubbles. You give it a starter text and hit generate, it generates until you hit stop or some system limit triggers. You then hit generate again.

your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points

I basically act like voices in a mind of an llm which induce thoughts and direct it without it even noticing, for the most part.

Sometimes, a dialog is still more preferable. E.g. I can think in a person-help-required mode and then it feels more natural. But I still dialog it there in a notebook tab, where I can adjust both my and its replies at any point, as if I was at the same time one level above. Chat is just a text with A:… B:… lines.

I’m not ND btw, found it by experimenting with writing stories.

Chatbots will always say what you want them to say, in a tone that you like. They will always affirm that you're right.

Where did they get that?

It's fair to say that many commercial chatbots are trained to be ingratiating by HFRL. I wouldn't say that they agree with you always but they are certainly "agreeable" in terms of the big 5 personality characteristics.

The positive value that opposes "agreeable" is "critical", but I'd say you can only get away with critical when you're right and chatbots have trouble being right.

You certainly train a Chatbot to be a real ass or have the attitude of Oscar the grouch if you wanted to.

Contrary anecdata: No. I have yet to get any value out of LLMs other than crossword clue solutions and first drafts of poetry. For everything else, I find the bland mediocrity of LLM text as bad as the hallucinations, and the combination of those intolerable. My more typical friends and coworkers are much more forgiving of these flaws in LLMs.

Do you do anything coding related?

I'm a software developer by trade. LLM-based coding buddies (at least as of July 2024) are better than copy-paste for boilerplate, but I've never yet gotten one to complete a novel class method in a way that saved time. And asking questions about Java or JavaScript? 100% wrong answers on the kinds of questions I ask it in practice; it always "rounds off" to the answer to a more common question.

So for example you don't find copilot useful?

The worry is interacting with LLMs reshapes the brain in as-yet-unknown ways. It may depend on one's position on "You are what you read."

One can use AI/ML tools without the conversant component; someday, there may be more visual feedback mechanisms like spreadsheet models or graphs (both nodal and charts), not just text.

Or maybe code--latent compute--is another expression: untangling it takes work, but it's more worthwhile than just "reading the answer."

I guess it's like wanting to understand the "physics" of a model than just using the model.

On the other hand, I hope to usefully apply "How to Say It: Flashcards for Professional Communication." So there's that.

Neurodiverse people tend to think more in the abstract relative to neurotypical individuals (this could be a lengthy analysis and discussion of itself). When working with chatbots at a more abstract level, one may be more likely to leave with a satisfactory result—be it a solution to a question/problem or just the conversation itself. Perhaps that is the dynamic of what you have witnessed and described in this post. I don't think there is anything else useful one could infer beyond that.

> Going a step beyond: are we, in the grand timeline of civilization, uniquely evolved to fuse with this new type of intelligence?

I've seen this notion claimed and sometimes phrased like above in ND circles and forums. It would imply that neurodivergent individuals are evolutionarily relevant, and hence have higher importance (superiority) over neurotypical individuals. That would be elitist. It can parallel arguments made by, say, in the grand timeline of civilization, a racist who could think that a particular skin color is superior or a sexist who could think that men are superior to women and that each has to fulfill specific roles.

To argue that something has evolved to develop certain features, there has to be evidence that the population was facing (possibly survival) difficulties in the environment being studied, that said features are observed increasingly in the population over time, and that those adaptations helped and improved the lives of that population. There is no such evidence to support the claim.

Neurodivergence is not new. It has been recorded throughout history in various ways—one example [0]. ADHD, ASD, and anything else that falls into the ND space has existed for a long time. It's only in recent history, as psychiatry and relevant practices grew and developed that neurodivergent conditions were reclassified repeatedly, to the point now where these conditions and behaviors are being seen as part of/under one umbrella of neurodivergence.

[0]: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3000907/

EDIT: Few typos.

I agree that a key underlying component of ND thinking is greater abstraction (i.e. "systematizing"), which is the thrust of better pairing with LLMs (but concede also that NTs may also exhibit the same).

My speculation about NDs was not meant to imply superiority or elitism -- more an interesting thought experiment that considers a wider aperture of history. You're right that it doesn't hold up against the fact that tracing evolutionary adaptation requires substantial evidence of survival difficulties and selective pressures.

If anything, I'm more fascinated by the parallels between how NDs such as myself navigate the world (e.g. masking, etc), and how LLMs are fine-tuned to interact with humans too (e.g. the "assistant/user" conversational model) It's like scaling down the "Are we all just stochastic parrots?" question to my own experience of often feeling like a "prediction engine" of others' minds.

An Nd friend of mine uses it to rephrase things so he doesn’t come across as too blunt. So llm is helpful for him at least.

I used to be wary of the Scantron style standardized tests administered by schools. We were informed that they would be graded by computer, and of all my classmates I knew best how tolerant of mistakes computers were. Especially since we were given contradictory goals of "erase mistakes completely" and "make dark marks".

I do not feel I can "co-reason" better with a cleverer version of Dissociated Press at all, because Dissociated Press isn't reasoning. It's the neurotypicals who believe that an LLM actually thinks, while I'm on to what it's really doing: It's statistically predicting the next thing to say given what has been said before. I have better things to do with my time than to attempt to goad DP into producing some output I like based on rough guesses of how it operates. I'd much rather train and use the neural network between my ears to work on interesting problems. At least I have access to some of its internal state and it doesn't live in Microsoft's cloud.

I never felt the need to talk to a computer or have things summarised for me. I so far have no need to "increase my productivity". LLMs don't improve my life. This has nothing to do with being neurodivergent. You'd do yourself a huge favour if you touched grass once in a while. Unlike LLMs, regular people will tell you when you're full of shit.

Never thought about this, but always surprised me how many people don't use AI at all; differences in personality might account for that

I've been gradually mixing AI into my google searches.

I find ChatGPT is good when I want something to summarize and give me a very concise response.

Usually, I prefer skimming the original sources.

Unlike my more neurotypical coworkers LLMs tend to provide gibberish or nonsensical answers for me. We haven't been able to figure out why I'm the chosen one who gets a rolls and rolls of non-compiling C function headers while they get ok answers.

I’ve assigned Claude 3.5 Sonnet as my coach/assistant/partner and it’s unbelievably useful.

I call it “rubber duck debugging for life on steroids.”

That sounds very interesting, can you provide any more details?

Nothing fancy. Just register for Claude, and talk with Sonnet 3.5 like you talk with coach or therapist, no prompts needed to begin. I got hooked after a few messages. It’s unbelievably good.

GPT4 is not even close for this type of tasks.

I tell it about my mood, my struggles, what made me stuck today, my latest argument with someone, how my life is going, anything really, and just follow the conversation on where it goes.

If I notice something doesn’t make sense, I call it out. It figures it out. Then I ask it for a few sentences to add to the prompt, and it delivers.

I’ve been doing this for 2 weeks now, best $20 I’ve spent.

It's because you are gaslighting yourself. It's made worse by the "character" llms get finetuned on so it's harder to realize.

Why was this flagged? The intention of my post was to have a discussion on user interaction with emerging cognitive interfaces. Variations in cognitive style could impact how these technologies evolve and are used. It's useful to understand different user experiences for the development of inclusive and effective AI systems.