Dunno, I've been learning a lot of Rust in the past few days. Just dove right into a project and asked AI to teach me stuff on a need to know basis. I'm actually getting used to Rust by now.

Personally, I find that its generated prose tends to have an undue weight to it, almost as if every topic I ask about somehow bears a heavy burden, or is otherwise load-bearing, to use its parlance.

Quite puzzling, really.

I view LLMs in education similarly to office hours. Some people abuse it to get homework answers without grappling with the material, but the optimal amount is not zero.

LLM certainly not a replacement for a book, where you get someone’s extended personal approach to a topic, thoughtfully organized, reviewed and edited, often times actual courses taught based on it, with answers checked and errata available online.

human conceptual thinking is very much a multi-dimensional graph, which relies on light "approximate" concepts that are "good enough". LLM AR token generation is extremely one dimensional and doesnt care about the "weight" of the concept behind a token.

LLMs hold billions of parameters in "mind" at once. humans hold like four "concepts".

This is the essential mismatch and the primary reason LLM conversation can be so painful and exhausting.

Explaining this and limiting "concepts" to four at a time tops is one of the very few AGENTS.md / system prompts I always use, and it has proven invaluable time and again.

Thinking traces show how effective this is at forcing the LLM to simplify its thinking.

[edit] Also, myself and nearly all of my peers are struggling to choke down the flaws of LLM tooling along with the benefits. the speed at which LLM adoption is being forced, without truly crafting them into quality tools first, is not ok, and not normal.

LLMs have stirred an inhumane hunger and fear. the tech is fine, but the way tech companies (creators and consumers) are behaving should be deeply questioned.

it's NOT normal. it's not ok.

I will say, opus 5 is an egregiously bad case of this, but other LLMs have this too, just less bad.

Recently switched to OpenAI and I've gotta say Sol is so much better at writing than Claude. Opus has a distinctive sentence structure and Fable somehow manages to be even more obtuse. The personality of these models really does come through...

I wasn’t even really concerned with optimizing low level code before LLMs and that wasn’t why I was hired either.

However following that low level thread: We can look at the reasonable options and immediately know if they’re reasonable or nonsense. Why? We know the code. Now zoom a level out, where I think our expertise really lies.

Building a complex system isn’t easy. There are customers with requirements, there are budgets, SLAs etc. Sometimes one customer needs X and one needs Y. Our expertise is taking all of this in, and producing something that balances all the different variables. It’s knowing that we’ll expect X events a second so we’ll need Y to ensure we can tolerate failure.

Is it possible LLMs will be able to do all of that too? Maybe. But then why would our customers need the enterprises they pay for?

They tell you have "hit the nail on the head" when you really haven't.

They tell you have had a "great insight" when you are really haven't.

They give you the illusion of learning and progress but essentially give you faulty preconceptions will trip you up further down the road.

You can ask the LLM to be more critical and less sycophantic but that only gets you so far:

They want you to continue using, being dependent on and feeding data into the LLM--your independence isn't a priority.

I have stuff to do now, the value of the knowledge in a year or two isn't important if it solves the issues I have today.

The more things you understand, the higher the chance you'll spot a situation to use them in the future.

I think the best innovations come from times when someone is uniquely able to combine two of their previous experiences together. The more experiences you have in your back pocket the more combinations you have access to and the more likely you'll have a unique combination when the right problem comes along.

i can't even get agents to remember core instructions like "use jq instead of writing a python script to parse some json"..

I'm researching causal inference right now, and my main goal was to make sure I understand how to test estimation on synthetic data.

Basically, it's the same way it works with people. If you delegate a task that you don't understand, and you can't have a credibility proof (i.e. doctors, lawyers), then you research a topic well enough to be able to (1) define the task and (2) verify the end result.

So I have the LLM offer a very short explanation of something, and from there's it's just me asking questions. Anything that feels fuzzy or not fully internalized is something I poke at until I'm satisfied.

It really has helped me develop a sensitivity to what I understand vs what I don't, and the ability to drill into any part of it is amazing.

> I get exhausted reading LLM prose

So much this! If I see one more sentence with the words "genuinely" juxtaposed with "load bearing" my head is going to explode!

btw, I am building the tutorial here for anybody interested in this topic: https://github.com/avilay/learn-probml

What is fascinating is how you can witness it at so many levels of organization. One example: Employer executive get enamored with moving from labor to capital. They believe that by using LLMs, they can replace a lot of workers. At my place of employment, we have people that are surprised they can't file a Jira ticket describing a product ask, and have it kick off an implementation. You can build the skill to attempt that, but invariably you'll get back questions like "what do you mean by <x>" and "what do you want to do in this case, a, b, or c?"; questions that a product person or an exec are not well suited to answer.

In the past, programmers did that kind of interpretation and judgment call. So then you're in a quandary; who should do that work? Work that previously, you never imagined was an inherent part of what the replaceable code monkeys do at your beck and call?

And then, how do you hire for that? How do you find the training for the people that are experienced enough with... something... to know what a cohesive error response is, or what kind of telemetry strategy is best for that particular product and organization, what collection of product asks are incredibly complicated for what they're asking and can deliver 95% of the benefits at 5% of the work if we just do this instead, and whether you want to aim more towards thick or thin clients?

Who are those people? Wait, those are programmers? Wait, there's this whole collection of inherently human skills that we devalued, by not appreciating they were always quietly doing that for us in the past?

That's just one example. There's a repeating pattern of discovering where the work truly is, work that was embedded in manual patterns we might not have to involve ourselves with anymore, but is yet still essential. So the nature of our jobs changes massively, but the overall level of employment does not.

At least, not in the medium to long term. There is a lot of painful churn we have to suffer through first.

Trust me when I say that in the hands of someone who doesn't have your experience, the LLMs would not be getting the results you get.

You might think what you're doing is trivial, it may be sessions that flow roughly, "Instrument this, okay this part is slow, profile this part, OK read the profile output and suggest a better approach".

But your experience will be steering it in the right direction, and you're probably unaware of just how much your experience is doing that guiding, as the LLM shoots off at 100mph, you feel like it's taking you with it, but you will be guiding it a lot more than you realise, and that's where learning and experience comes in, even if you're no longer operating at the lowest depth, your knowledge of that layer will be helping.

If nothing else, the experience to know when something is actually slow is a skill in itself. If a function takes 200ms, sometimes that's as quick as it can realistically go, and sometimes that's literally a million times slower than it could be, and there's actual skill and experience wrapped up in knowing what "slow" looks like.

I'm not 'wasting time' but I'm also not really learning.

Opus 5 first built me a detailed plan, but a couple important details were either obviously wrong or felt unnecessary. I went back and forth asking for sources and more information probably like 4 times and every time it did the "in looking at things in more detail it appears my previous advice was incorrect" spiel. It just became exhausting at some point because it feels like it really lays bare how LLMs are just minimizing that loss function but don't actually "understand" anything. It was really useful as a search engine (it correlated some highly relevant source docs), but I just couldn't trust it to believe it was actually done at any step.

The main idea is you can do any style you want or like to learn.

> Time and time again, when talking to people who rely on ChatGPT, Claude, Perplexity, and other general AI tools, I hear them say, “AI is incredible. It handles nearly everything I throw at them.”

> “What does it fumble with?” I’ll ask.

> “Well, it still gets things wrong when it comes to my line of work.”

https://www.dbreunig.com/2025/04/08/on-ai-observational-comi...

In particular it might be valuable to be in the habit of learning things that one is bad at doing.

Or not.

In my experience it's infinitely easier and faster to learn deep, "boring" things when you understand how they relate to your shallow and wide understanding of all of the related components.

The LLM is merely a tool. And you can use it for domain discovery that enables efficient deep learning at an unprecedented rate or you can develop a cursory understanding of a topic and think yourself an expert.

Indeed it's often a waste of time to just focus on talking people fully if you want to learn fast, reading and especially deliberate practice are better for that. But if you don't have the time, energy or focus, then listening to interviews in the background can be useful supplementally

It’s true of most things. Running, dieting, weightlifting being uncomfortable is a sign of progress.

I have done this for all my work this week and it works quite well.

For one it lets you actually query the LLM as to why, their plans give a high level not every single change and it allows you to correct it as you go and the plan will change.

This is the only path to mastery, or understanding if one prefers. There are no shortcuts to a person achieving deep understanding (a.k.a. "Aha!" moments).

Can a tool such as GenAI be beneficial to someone who already has done the work to understand? Absolutely. But it cannot infuse mastery into a person simply by its use.

Only the time and effort a person devotes can do that.

What is your process for creating these resources?

You can ask the LLM how to do this. Start with a topic you know well to get the mechanism working and trust it well.

Hallucinations bring to question what you think you've learned. That's going to cost long-term if you labor under mis-apprehensions until you maybe figure out you learned something wrong.

Perhaps the best example has been a native macOS app that is a completely custom text editor with built-in debugger, lsp support, fuzzy finder, etc stuff you'd expect. Inside the same app is a library of books i can read within the app completely formatted and for every chapter/section of each book that is a quiz to take (LLM generated of course), a "recitation" tab where i am asked a question and say outloud my response to the AI to evaluate me on and then finally practice problems to do within the custom text editor (these are usually programming books). The reader also has ai re-write built in.

As neat as this is, and i worked through K&R like this, i have ultimately fallen back on "just read the damn book and go to the AI when you've got questions."

I just want to understand more.

Also, would you be willing to share the actual text of it that you put in AGENTS.md?

I'd imagine an application that uses LLMs will be created that better manages learning. It's just not clear what that UX is yet- it's obviously not just a chatbot

I have a personal theory: LLMs are *fundamentally* handicapped at perceiving what's going on in the mind of the human (this can't be "innovated away") and that's at the root of what makes them suck at conversation.

Next time you're chatting with someone, notice how much understanding is shared without anything being said. E.g. the other person might share something deeply disappointing, and they can tell without you even saying anything whether you get what they're going through. This unspoken-yet-communicated information guides the conversation. Or as another example: humans can read the room -- you walk into a room and immediately adjust your demeanor based on what you see and sense.

LLMs are totally blind to things like this, and this adds an inescapable awkwardness to interacting with them. I don't believe they'll ever grow out of this. Which thankfully implies more long term demand for humans instead of robots. :)

For example having an LLM summarize a dense topic and to find books so that you can filter faster and spend time reading those books works way better than having the LLM summarize the books or the topic (or even relying on second hand information). Another one is having the LLM quiz you on your topics of interest. With questions tailored to attack specific areas that you struggle with. Its wonderful at this, nothing I've used comes close to what an LLM can do here.

You define for yourself what your goals are, slowly refining them as you learn more, and use LLM as a tool. This ,I find works best for learning.

And you notice when it's a topic you know well or something like software where you can immediately tell the options it's giving you don't exist on the page. Leading to the amusing statement "LLMs are bad at what I do but great at everything else".

I think this is one reason why LLM text is pretty exhausting to read for long stretches.

i run into context window limits, or practical limitations of digitizing the book

I even wonder if this behavior is due to next-token prediction architectures, somehow.

People don’t even have to be lying to be wrong about this stuff. Someone can learn enough about a topic to be halfway up Mt. Stupid in no time flat, and in doing so, think they not only truly understand the topic at hand, but might be particularly adept because they were such quick studies. People that know less are impressed, because why wouldn’t they be? Anybody that knows more than them sounds like an expert. And people that know what they’re talking about cringe at the overconfidence, and probably try not to engage: who wants to have to prove that someone’s boundless confidence is entirely baseless? Most of the time, they think the actual expert is full of shit because they think they’re the expert. It’s incredible how many times I’ve had people in tech confidently, even smugly “explain” design concepts and strategies to me that they did not actually understand, knowing I was an experienced, degree-holding designer… and they didn’t even have a chatbot’s lips on their ass telling them how smart and insightful they were.

But even with Claude, it's it's really the prose getting in the way you can install the caveman plugin or tell it to use that "standard technical English" thing.

And yes, it is not that it is just presenting the facts. By me taking control of the direction the questions and answers go, I can flesh out my mental model. I won't retain every little thing it tells me. But I am much farther ahead than before.

it also researched vision correcting displays for me and i can finally put that idea to bed - i was never really going to pick up an optometry textbook tbh. plus it was able to pull together a bunch of geometric and physical context about light and the eye plugging exactly my personal knowledge gaps.

in general i suspect these materials might not be that interesting to others because they are so custom to my learning style and personal needs and preferences.

these are usually not one shot documents but rather many prompts deep before i get something I’m willing to sit down and read or study. but dramatically quicker than assembling it myself from primary sources. i wouldn’t say it matches master expositors but then they’re not available to write on any topic i happen to need right now.

plus I’ll just have a live voice discussion with the system when i go for a walk and there are still things bothering me on a topic. it takes a little patience but if i’m in the mood it’s amazing.

i generally find that it can help track down specific references if i suspect hallucinations. but especially on factual topics my experience so far has been extremely encouraging.

“asking the right questions” is also a moving target with each model release

People simply underestimate the value of doing the work and think that the end result is all that matters

I know it sounds silly but 1 layer ends uo being way worse than 2.

Otherwise you probably get more confused as you have mentioned.

On the other side, Peter Diamandis describes a situation where a bunch of kids were given a internet-connected computer and they had no teacher. Instead of it there was a “grandma” that checked kids from time to time.

After that there was a knowledge test that revealed “no teacher” approach was more efficient.

But it was a group, not an individual activity…

I assume this will become less of an issue in the future as there is more trust between the AI tools and me.

""" In this project, I require a socratically delivered line of conversation. Here's the typical structure to the conversation. I ask some question. You need to factor and reason about how to conduct and deliver a conversation. Best practices would be to limit terminology, or assess with the user whether they have a firm grasp on terminology before you use it. You must be very strict about this, it's unacceptable to just introduce a new concept, actor, phrase or other complication into the conversation without first labeling who what or why it exists for the conversation.

Conversation structure needs to be front-loaded with a brief interview for the user, "you understand X?", "whats your understanding of Y?".

Conversation structure then needs to proceed with single-sentence questions from the agent. User replies with an answer. Sometimes the agent needs to correct the user, but only ever do so with yet another question. """

^^ these are the instructions I have installed at the root of a "project".

Keep in mind, this is claude opus 5 low effort we're talking about, in the "projects" area of the mobile app. Here's the process I use to set up its knowledge:

1. I take screenshots of the textbook on my iPhone, and upload a chapter at a time.

2. I have it summarize the chapter into markdown by analyzing screenshots. You could probably achieve this simpler, if you just had the textbook in PDF.

3. I walk with my boy Clau-crates.

I've done this for a couple of weeks and haven't seen it revert back into its typical context-dumping behavior.

On second read, there's probably some clean up I could do. Thanks for making me pull it out and look at it. Things that could probably be improved:

- tell it to cross-check resources online to further ground itself

- use simple, short sentences (long sentences make the brain blur a bit)

- not be sycophantic (it seems like project mode has discarded with my root-level anti-sycophancy prompt)

Yes, sometimes its wrong, most times its right, cross checking is fairly easy, not using it because of the possibility its wrong seems a baby with the bath water thing.

I learnt a lot of functional programming from it, stuff I've always wanted to learn, but just didn't have the time and really the sources can be difficult, it really explained things well, and as someone else said in this thread, you can ask questions over and over until you understand, asking a person that (if you can get an expert) would drive them nuts. Maybe my experience isn't typical, its hard to tell, everyone reports something different.

https://adaptive.bounded.cc

Trying to diagrams/animations didn't yield good results even with frontier models. But pure text, any model does a decent job.

It's possible that this quality you describe stems from the extensive training corpora utilized by the major AI labs. These almost certainly include work from the esteemed economist Jacob Silj:

https://www.youtube.com/watch?v=Poc1upTejD8

I know you probably don't consider it dense but wondering if someone can shed insight.

I find them like empty calories, like programming youtube tutorials. They maximize for feeling learnt instead of steady progress

The full PDF is worth a read (Figure 1 may be of interest to many here): https://www.cambridge.org/core/services/aop-cambridge-core/c...

If "attention is all you need" then it's something we do indeed lack, in comparison to LLMs! But it's an interesting question: might machine cognition benefit from similar bottlenecks in an attention algorithm? Advancements like Kimi Linear seem to indicate that we're far from the finish line: https://arxiv.org/abs/2510.26692

The question: what's the net positive gain of turning people who know nothing in a given field into sub-novices, while weighing actual experts down with work slop and marginal returns?

And I wonder what the true cost is of arming so many novices with that level of dangerous knowledge.

The body of literature on learning theory, and beyond that on specific types of learning and specific mediums such as learning from text is so rich there are way more useful models to draw from. Believe it or not, prellm, researchers in the textual learning field had already demonstrated you can achieve performance equal or better than novice tutors using pretty basic computer aids that follow specific hint/pump interaction structures. Guiding an LLM to use these findings has evidence backing it and is way better than telling it "i guess be like socrates". The problem is, to realize there might be richer more effective and highly researched ways of tackling the problem beyond the first fart of a thought you had one afternoon requires the deep respect for expertise and specialization that precisely basically everyone in the AI space right now fundamentally lacks.

It's long been the case that the best way to learn something is to teach something.

And the craft is loose term, it can mean anything you like to get better at.

https://en.wiktionary.org/wiki/eat_one%27s_seed_corn#English

Personally I think this is a bad characterization of using LLMs to fix up LLMs because while you can never guarantee results this way (as the quoted line claims here, which is worthy of criticism), it is, in practice, useful to use LLMs on top of LLMs. And there's no infinite regress. Auto-mode in Claude Code, for example, seems to me like it's been successful at making the system more safe than --dangerously-bypass-permissions without prompting the user for permissions constantly.

Any more detail you can share? Do the others feel more "human"? Are there any that are particularly digestible/human-friendly?

I've been wondering for a while if this is just Claude because I mostly use Claude, so this is very telling.

Totally agree, unfortunately careful simulation games are very rare

So it may be very slow or become unavailable, back end can't handle that, no caching whatsoever.

Does anyone have a read on if this is primarily a Claude issue, or if all LLMs do this?

Which is pretty unfortunate for those that want to learn. I used to enjoy writing documentation at work, it was my favorite part of the job. And it did feel like it benefited me more than it benefited all the people that were (or weren't) reading my documentation. Now I can't really justify spending much time on docmentation when LLM's can do it in a fraction of the time and it's "good enough"

Just hoping folks don’t get hurt due to people not understanding what they’re doing with these things but believing they’re competent.

What triggered my response was the “just review the output with another LLM and it’s perfectly correct”

I tried using a new agent service recently and could tell immediately that it's powered by Claude due to the way it writes.

Full story in the book

Tangentially, but related: I'm old enough to remember when the spirit of your comment was pervasive on HN.

I’m old enough that I worked my first IT summer job the same year slashdot was founded. I’ve seen a lot of tech tribalism form and dissipate, and this one didn’t feel organic. My gut says a lot of the us-vs-them tension originated in a deliberate campaign to cast AI boosters as the tech industry in-crowd, and ‘other’ the people not on-board. Who knows.

And even if they are related - if Opus 4.8 always has a 1:100 chance of a specific hallucination - then running the same model twice does indeed dramatically reduce the odds of an error in the final output.

E.g. the entire framing to combat complaints about shortcomings was,"It's not the tech. It's you. You're just not doing it right. wrong setup, wrong workflow, add this to your .MD, use loops, etc". Every complaint was immediately met with this same treatment by a swarm of vague bro-bots that materialized from the ether. Always the human's fault.

And, don't get me started on the waves of newly minted expert AI creators, making recommendations without showing a single example of what they'd supposedly built.

I'm sure some bandwagon organic creators tried to cash in on the genre, but that too was part of the point.