I thought LLMs were a great tool for learning new topics - perhaps even complex ones. But overtime, I have had several frustrations with this. First, I get exhausted reading LLM prose. I really don't want to read anything generated by something like Opus 5 at this point. Second, as I dive deeper, I need a way to organize the information in a useful way as I begin to branch out in many different directions. I have tried to use the LLM to fix this by having it generate a web page with diagrams and organized information flow. It's an improvement, but I still run into the issues I described in my first pint - LLM prose is annoyingly dense, and the useful information gets lost in a bunch of noise. You can direct it do something like "use plain English and avoid LLM prose - provide only as much information as necessary to demonstrate the point", but it is once again only a marginal improvement.

And then I begin to think to myself that I should just read a book on the topic written by a trusted source who put a lot of effort into teaching the topic properly and presenting the information in a thoughtful way. So, I am back to books and mostly try to use LLMs to clarify certain questions or ideas I have.

> What you get is a beautiful animation that is 100% accurate and free of hallucinations.

I'm not sure I follow how this is actually guaranteed? The fact-checking process mentioned just seems to involve asking AI to review its own work.

I am greatly disturbed by the idea that any of the demonstrations involve "complex topics". Whether it was the silicon workflow, LLMs, or EUV, I saw nothing complex. Are these supposed to be freshman undergrad or even high school level explainers? I suppose I didn't see any glaring errors on a quick glance, but definitely lots of details were glossed over.

Perhaps basic special relativity could be done this way, or simple derivatives, but definitely not general relativity or integrals, much less PDEs. I guess I was hoping for a 3B1B type output. Oh well.

What’s everyone’s opinion on learning new tech things in this day and age? My opinion swings between positive and depressing vision of the future.

I still learn new stuff, but I’m afraid it won’t have any value in a year or so.

For example, I’m pretty good at optimizing low level stuff, but right now you can just ask LLMs to do so and they are pretty good at it. They will profile the code and suggest reasonable options like 90% of the time.

I’ve been using LLMs to create readable rewrites of RFCs and specs that interest me. It is not precise enough for implementation use, but it has increased my understanding of the underlying RFC.

Another useful approach has been asking Codex to implement complex things, like a Kademlia DHT or BitTorrent client in a literate style with the explicit purpose to increase understanding by reviewing the source code.

Examples: https://rickcarlino.com/notes/note-dump-and-ai-summaries/ind...

https://github.com/RickCarlino/tiny-bt

The biggest thing I've learned from doing stuff like this is that there are no shortcuts. At some point or another, to truly learn something deeply, you've got to dig in to the boring details and do things the hard way. LLMs can help with this...but I find it's usually tempting to try and just offload the boring stuff to them, which doesn't work.

The main bottleneck as an engineer is no longer writing or testing code. It is how long it takes to understand complex systems. This is a really nice approach that, if you have the tokens and the patience, feels like I great way to learn something and I think we'll see more and more stuff like this.

I had a similar realization a few months back and am working on a tool that generates "mermaid walkthroughs". It is 1000% less pretty but it is fast and is pretty good at explaining how services work or what a code review does or just as a way for your agent to explain some decision to you.

https://github.com/scottrogowski/ariel

The title is not representing what the post is about. “Use LLM to learn complex topics” here actually means that the author asks an agent to describe the problem area, and then implement a simple web-based simulation game, and by playing that game, the author actually learns about the topic and its constraints. They use chip making as an example.

That's actually a fun way to learn processes!

I guess we all learn in different ways. I prefer reading, and can fairly fast find the articles that make sense to me.

Colleagues often suggest podcasts and videos - I very, very rarely listen to them or see them.

The bandwidth is too low. It's not efficient and ultimately I'm bored.

This is a nice project, it looks cute. I watched some of the pages But I want more than that, more information, and faster - still a Wiki fan.

Also, step number 2 in the flow: have the LLM check itself... Naah, I don't believe that.

But you're not the only using gen ai like that. Take care.

My main issue with using AI as a learning platform is that unlike documentation, books, Youtube videos, there is not really a process of having someone "review the learning material". For example, I can always read the review of some book or ciriculum, the comments under a video, or if it's some for of open source documentation you can check the PRs and verify to some extent it's claim. With AI I can't really say what it has halucinated, because I am learning a new thing, I don't have that benefit of previously reviewed material.

I use it by telling it my background, giving it a rough timeline and asking it to create a learning timeline, save progress along the way, and git push / pull periodically so I can use the same thing on both Linux and Mac. I tell it for each phase in the learning timeline, present me information, then challenge me on it. If it's code, it challenges me with a coding challenge, where I use it in an IDE plugin. If I'm learning something that isn't strictly code, then I ask it to give me info, then challenge me with questions and grill til I get it right. I ask it to save what it think I struggled with, so that later we can drill it again and I can also review it in an .md file.

Does anyone else use Claude like this?

It's sped up my learning by 10x. I struggled with 'just reading a book.' Take kubernetes. I hemmed and hawed and spent years periodically reading some dry book or blog or official doc, falling asleep, and forgetting while I got busy. Now I'm aggressively working with it, almost like I'm addicted to a gamification, of getting through our learning timeline, and I'm excited to move forward as quickly as possible and pass its tests.

It's like a fake teacher, because I can also ask it to drill into a topic or re-explain itself if it made no sense.

The only thing that worries me is, sometimes I'll say something like, "Um, are you sure about that?", and it'll apologize and correct itself. I barely challenged it!

I’ve written a skill that I basically feed what I’m looking to do, some ideas I had for accomplishing it and any other details like tech stack, etc.

The skill then riffs with me, judging my ideas and suggesting alternatives. We go back and forth until something useful comes out of it. This process isn’t unlike how I do normal development.

However, once agreed it breaks the work into “steps”. It then creates a tutorial for me, for those steps, explaining each line, why each change happens etc. I can then ask questions, muse about an alternative idea etc. Then I do the steps, and I’ve learned and gotten what I wanted to get done.

This has been how I’ve been learning Godot and making a game for the past month or so. I didn’t go in blind, I started with a course from GDQuest so I could feel confident guiding the tutorials. I will say though, having a tutor to bounce ideas off of has been really useful.

I still try to figure it out myself, consult the docs, discord etc. But if I’m stumped I’ll run my tutor skill and have some fun.

I've been using them by reading some docs/wiki/tutorial, then when I think I understand something trying to do a rough explanation to the LLM and ask if I'm right. I'm usually making some analogy to something I already understand a little. I'm usually partially right but missing some key bits at the first pass. I go back and forward asking for explanations of various bits or asking for resources around the area I'm not understanding. Often times just discovering the relevant name for the area of study opens lots of doors. I basically use it like I would talk to a knowledgeable and patient teacher.

As for how useful it is to understand thins, I believe it's still useful and hope it will continue to be.

I’ve been doing something similar (browser only interactive courses served from GitHub pages) to teach me topics from beginner to advanced

LLMs and systems intersection - https://kernelspace.naigap.com

Distributed systems - https://byzantine.play.naigap.com

If you're using LLMs to learn or for research, and at some point you don't end up engaging with an actual resource (books, papers, lectures, web pages, etc) then you're playing yourself.

I wish there was a LLM tool to explore a topic recursively, as a tree or a mindmap. You would start with some high level concept (say "cryptography") then dig further and further to more specific topics.

I think that would be a way more natural way to explore than being stuck on the classic linear output of a LLM.

I use LLMs to learn deep technical concepts. I really like them because I can spend countless hours a day understanding things and building an investigation file with all my findings. I code examples and test the findings. It has helped me understand basically anything.

I'm using LLMs right now to build a terminal browser, a GUI browser, and a PyTorch/LibTorch replacement. It's really fun to be able to learn and make progress this way. It's like reading multiple interactive books, where every concept can be explained again and again until I understand it.

Gamifying the presentation could make topics more accessible to others. For me the overhead wouldn't help with my own learning. Also I've been burned by just learning things mechanistically (e.g., coding, applying algebraic rules), so I'm leery of learning just by making flashcards or models of the topic.

I find LLM's do great for learning when I ask what are the principles, how the main applications work, what are the key drawbacks, where are the growth plates in the field, etc. - the kind of thing a good advisor points to. Sometimes I have to ask it explicitly to use topological order of topics and show relations, which often highlights the gradient changes in the learning curve. For pruning, it's surprisingly good applying philosophical heuristics - Occam's razor, or Derrida's differance (the difference that makes a difference), etc.

And finally, no learning is effective without problem sets, and for those LLM's at times get me over blocking issues.

The degenerate case is memorizing the glib phrases regurgitated back to me; they're helpful and functional enough to get me into real trouble!

I'm all for this and I'm keen to try it. I certainly don't want to take away from sharing another neat use case for learning.

But

> What you get is a beautiful animation that is 100% accurate and free of hallucinations

100% free of hallucinations when you're not an expert that can check it is impossible. LLM hallucinations are an unsolved problem.

My high hopes were quickly dashed at the step that involved turning a complex topic into rollercoaster tycoon.

> In plan mode (using CC, or OpenCode) I ask a model to build the foundational knowledge for X topic. I ask it to review the accuracy of the knowledge base it built in the previous step.

> What you get is a beautiful animation that is 100% accurate and free of hallucinations.

How do you make that leap?

i've asked llms to write a presentation for me on a given topic. for whatever reason, that finds the hidden layer responsible for explaining things well.

Something that I have realized recently is that it has become so easy to get an answer to almost any question with the help of chatbots that its almost unnecessary to spend any effort thinking about the problem or the solution. I feel like before when I had to spend time researching a problem to find an answer I learned so many things around the topic itself which helped me understand the problem itself better and gained a deeper understanding. Today it feels like you can have an answer to the most complex questions you might have, yet you gain a superficial understanding of the topic and might forget about it quickly.

> I ask it to review the accuracy of the knowledge base it built in the previous step.

> [I ask it to build an interactive thing]

> I then push it to a new repo and enable GitHub Pages for it.

Congratulations. You are an echo chamber for LLMs. Use it to create, and verify, and post to then be scraped and trained on again.

The little tool it outputted is nice, but click around the stages and the text is not high quality at all. The snippy titles, abbrievated explanations, I wish a few more iterations and thought was put into the actual main textual content. Especially for 'complex' stuff

All these LLM-as-review hype pieces don’t acknowledge that it’s turtles all the way down

Agreed. Given how many significant errors LLMs make in my topic of expertise, despite my taking multiple error checking steps, the idea of catching 100% of hallucinations because you told the LLM to check itself is hilarious. It’s just a wild lack of insight: “I’m using the LLM to teach me something I don’t know about, I definitely have the knowledge base to spot any errors that might remain!”

even if you say use RAG or something to a source you can trust, there's no guarantee the agent will still use exactly what the source has.

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

Completely agree. While I didn't set things up to have AI review its output in a loop, my experience trying to build a specific acoustic testing rig with Opus 5 also aligns with the other "it's turtles all the way down" comment.

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.

I could certainly envision a scenario whereby review would increase reliability but not how it would every guarantee 100%, there is a pretty big logical gap there.

You can add "make no mistakes" to the end of the prompt and achieve the same result while burning less tokens.

Yep, that's impossible. The hard truth is that most people this lost to LLM psychosis cannot understand that fact. It's better to treat it like someone in a cult, arguing the facts isn't going to help if they refuse to accept them.

I thought after reading the title that the text was about learning something, yet the actual text seems to be about having a system do something for me.

They're amazing at it, provided you keep asking the right questions.

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.

Generally speaking, the pattern is that people are overestimating how much "work replacement" will happen, and underestimating how much "work shifting" will happen.

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.

> I still learn new stuff, but I’m afraid it won’t have any value in a year or so.

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.

I can only speak for myself, so I hope this resonates with you.

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?

I sometimes get vague ideas for solving maths/science problems. They never pan out but i can talk in detail on group theory and advanced maths and science topics due to investigating such vague ideas over the years. These days the LLM shoots the ideas down instantly and honestly correctly, i know enough to know "yeah that's right, oh well" and move on. Which actually takes away a huge avenue of learning. I'm pretty torn on the outcome of this honestly.

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

There's a big gap between being able to ask questions about something and understanding something.

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.

There’s still an immense value in training the brain to learn and be able to approach new problems with the sort of procedural thinking that LLMs enable. We can explore topics that we are curious about and develop that sort of “muscle” to continue asking questions when we have them. I have no fear that when those bigger (and existential) problems arise we’ll be well equipped to keep asking questions and figuring out ways to solve them.

One possible value in learning new things is developing a habit of learning.

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

Or not.

It depends strongly on the prompt. Like recently I was also doing some performance optimizations, and if I did not mention profiling none of the LLMs even profiled the code; they merely read the code and assumed, based on their own analysis of big-O time complexity. It is after I explicitly asked for profiling that the LLM started actually profiling.

So you think building a warp drive is pointless because 99% of work, per your judgement, will be done by AI? Is your contribution meaningless and artifact useless? I wouldn't think so.

> I still learn new stuff, but I’m afraid it won’t have any value in a year or so.

This is silly. This would be like arguing that encyclopedias made knowing things pointless. I learn new stuff for me.

Professionally, it's important to know enough to know if you're going in the correct direction. Practically, tokens are going to continue to cost money and knowledge can save you tokens.

I don’t agree with the implication that there has to be a practical reason to learn new things. I enjoy learning new technologies because it’s fun.

The way learning new stuff rewires your brain cannot really be predicted, but the effects are positive.

I mean, part of the reason the LLM can do that is because you know enough to direct the LLM to do so and verify the results to some degree, right? It's good to learn new things because:

1. It satisfies you curiosity (and curiosity is always valuable)

2. You can better utilize the LLM to expedite something you now have knowledge about

3. You still improve as an engineer/programmer/prompter/whatever

I still think it's very important not to outsource everything to AI because there is a lot of value in learning and doing things yourself which is an important part of life.

I decided to learn watchmaking instead

I have staff ranging from 10 years of IR experience to right out of college.

I can tell you that there is an enormous gap in ability between them despite them both using LLMs for daily IR work.

The reasons aren’t complicated. The senior responders have tacit knowledge of how breaches evolve and what to look for which gives them a much better framework for where to employ the LLM.

The juniors will normally start from “here are some logs, look for weird” which is fine but leads to tunnel vision and a lack of confidence in their reporting.

I don’t mandate that anyone do work with or without an LLM. I hire seniors based on experience and juniors based on interest. But my experience has so far been that our best up and comers focusing more on learning the technologies instead of leaving those details to the LLM are developing their intuition and understanding faster and in a more robust manner.