Study mode • Hacker News Board

Re: Study mode

#571

Post by ainiriand » Wed, Jul 30, 2025, 5:53 AM UTC

BeetleB wrote: post #264

ainiriand wrote: post #56

Earlier quoted context omitted.

I've learnt Rust in 12 weeks with a study plan that ChatGPT designed for me, catering to my needs and encouraging me to take notes and write articles. This way of learning allowed me to publish https://rustaceo.es for Spanish speakers made from my own notes. I think the potential in this regard is limitless.

Now this is a ringing endorsement. Specific stuff you learned, and actual proof of the outcome. (Only thing missing is the model(s) you used).

Standard ChatGPT 4o.

Re: Study mode

#572

Post by safety1st » Wed, Jul 30, 2025, 5:55 AM UTC

jacobedawson wrote: post #379

An underrated quality of LLMs as study partner is that you can ask "stupid" questions without fear of embarrassment. Adding in a mode that doesn't just dump an answer but works to take you through the material step-by-step is magical. A tireless, capable, well-versed assistant on call 24/7 is an autodidact's dream. I'm puzzled (but not surprised) by the standard HN resistance & skepticism. Learning something online 5…

Firstly, I think skepticism is a healthy trait. It's OK to be a skeptic. I'm glad there are a lot of skeptics because skepticism is the foundation of inquiry, including scientific inquiry. What if it's not actually Zeus throwing those lightning bolts at us? What if the heliocentric model is correct? What if you actually can't get AIDS by hugging someone who's HIV positive? All great questions, all in opposition to the conventional (and in some cases "expert") wisdom of their time.

Now in regards to LLMs, I use them almost every day, so does my team, and I also do a bit of postmortem and reflection on what was accomplished with them. So, skeptical in some regards, but certainly not behaving like a Luddite.

The main issue I have with all the proselytization about them, is that I think people compare getting answers from an LLM to getting answers from Google circa 2022-present. Everyone became so used to just asking Google questions, and then Google started getting worse every year; we have pretty solid evidence that Google's results have deteriorated significantly over time. So I think that when people say the LLM is amazing for getting info, they're comparing it to a low baseline. Yeah maybe the LLM's periodically incorrect answers are better than Google - but are you sure they're not better than just RTFM'ing? (Obviously, it all depends on the inquiry.)

The second, related issue I have is that we are starting to see evidence that the LLM inspires more trust than it deserves due to its humanlike interface. I recently started to track how often Github Copilot gives me a bad or wrong answer, and it's at least 50% of the time. It "feels" great though because I can tell it that it's wrong, give it half the answer, and then it often completes the rest and is very polite and nice in the process. So is this really a productivity win or is it just good feels? There was a study posted on HN recently where they found the LLM actually decreases the productivity of an expert developer.

So I mean I'll continue to use this thing but I'll also continue to be a skeptic, and this also feels like kinda where my head was with Meta's social media products 10 years ago, before I eventually realized the best thing for my mental health was to delete all of them. I don't question the potential of the tech, but I do question the direction that Big Tech may take it, because they're literal repeat offenders at this point.

Re: Study mode

#573

Post by aucisson_masque » Wed, Jul 30, 2025, 5:56 AM UTC

When I studied I found out that being able to speak with a fellow student of the subject we needed to learn was so much more effective than the usual method of reading /trying to remember over and over.

So much that first method would take me an hour as opposed to an entire evening when reading/repeating.

Having such a tool would have been a game changer to me.

I don’t know tho if it’s possible to throw at it entire chapter of learning book.

Re: Study mode

#574

Post by PaulRobinson » Wed, Jul 30, 2025, 5:58 AM UTC

romaniitedomum wrote: post #412

jacobedawson wrote: post #379

An underrated quality of LLMs as study partner is that you can ask "stupid" questions without fear of embarrassment. Adding in a mode that doesn't just dump an answer but works to take you through the material step-by-step is magical. A tireless, capable, well-versed assistant on call 24/7 is an autodidact's dream. I'm puzzled (but not surprised) by the standard HN resistance & skepticism. Learning something online 5…

> Learning something online 5 years ago often involved trawling incorrect, outdated or hostile content and attempting to piece together mental models without the chance to receive immediate feedback on intuition or ask follow up questions. This is leaps and bounds ahead of that experience. But now, you're wondering if the answer the AI gave you is correct or something it hallucinated. Every time I find myself putting…

Despite the name of "Generative" AI, when you ask LLMs to generate things, they're dumb as bricks. You can test this by asking them anything you're an expert at - it would dazzle a novice, but you can see the gaps.

What they are amazing at though is summarisation and rephrasing of content. Give them a long document and ask "where does this document assert X, Y and Z", and it can tell you without hallucinating. Try it.

Not only does it make for an interesting time if you're in the World of intelligent document processing, it makes them perfect as teaching assistants.

Re: Study mode

#575

Post by aucisson_masque » Wed, Jul 30, 2025, 5:59 AM UTC

jacobedawson wrote: post #379

An underrated quality of LLMs as study partner is that you can ask "stupid" questions without fear of embarrassment. Adding in a mode that doesn't just dump an answer but works to take you through the material step-by-step is magical. A tireless, capable, well-versed assistant on call 24/7 is an autodidact's dream. I'm puzzled (but not surprised) by the standard HN resistance & skepticism. Learning something online 5…

Agrée, it would have been a godsend for those of us who were not as fast as the other and were eventually left over in usual schooling system.

Beside there isn’t any of the usual drawback with privacy because no one care if OpenAI learn about some bullshit you were told to learn.

Re: Study mode

#576

Post by Paradigma11 » Wed, Jul 30, 2025, 5:59 AM UTC

apparent wrote: post #562

Paradigma11 wrote: post #539

Earlier quoted context omitted.

Just have a second (cheap) model check if it can find any hallucinations. That should catch nearly all of them in my experience.

What is an efficient process for doing this? For each output from LLM1, you paste it into LLM2 and say "does this sound right?"? If it's that simple, is there a third system that can coordinate these two (and let you choose which two/three/n you want to use?

Markdown files are everything. I use LLMs to create .md files to create and refine other .md files and then somewhere down the road I let another LLM write the code. It can also do fancy mermaid diagrams.

Have it create a .md and then run another one to check that .md for hallucinations.

Re: Study mode

#577

Post by PaulRobinson » Wed, Jul 30, 2025, 6:00 AM UTC

ink_13 wrote: post #526

teleforce wrote: post #479

Earlier quoted context omitted.

Please check this excellent LLM-RAG AI-driven course assistant at UIUC for an example of university course [1]. It provide citations and references mainly for the course notes so the students can verify the answers and further study the course materials. [1] AI-driven chat assistant for ECE 120 course at UIUC (only 1 comment by the website creator): https://news.ycombinator.com/item?id=41431164

Given the propensity of LLMs to hallucinate references, I'm not sure that really solves anything

I've worked on systems where we get clickable links to source documents also added to the RAG store.

It is perfectly possible to use LLMs to provide accurate context. It's just asking a SaaS product to do that purely on data it was trained on, is not how to do that.

Re: Study mode

#578

Post by travisgriggs » Wed, Jul 30, 2025, 6:02 AM UTC

jacobedawson wrote: post #379

An underrated quality of LLMs as study partner is that you can ask "stupid" questions without fear of embarrassment. Adding in a mode that doesn't just dump an answer but works to take you through the material step-by-step is magical. A tireless, capable, well-versed assistant on call 24/7 is an autodidact's dream. I'm puzzled (but not surprised) by the standard HN resistance & skepticism. Learning something online 5…

Yeah. I’ll take this over the “you’re doing it wrong” condescension of comp.lang.lisp, or the Debian mailing list. Don’t even get me started on the systemd channels back in the day.

On the flip, I prefer the human touch of the Kotlin, Python, and Elixir channels.

Re: Study mode

#579

Post by maerch » Wed, Jul 30, 2025, 6:03 AM UTC

QuantumGood wrote: post #417

phailhaus wrote: post #413

Earlier quoted context omitted.

HN is resistant because at the end of the day, these are LLMs. They cannot and do not think. They generate plausible responses. Try this in your favorite LLM: "Suppose you're on a game show trying to win a car. There are three doors, one with a car and two with goats. You pick a door. The host then gives you the option to switch doors. What is the best strategy in this situation?" The LLM will recognize this as SIMIL…

I use the Monty Hall problem to test people in two steps. The second step is, after we discuss it and come up with a framing that they can understand, can they then explain it to a third person. The third person rarely understands, and the process of the explanation reveals how shallow the understanding of the second person is. The shallowest understanding of any similar process that I've usually experienced is an LL…

I am not sure how good your test really is. Or at least how high your bar is.

Paul Erdös was told about this problem with multiple explanations and just rejected the answer. He could not believe it until they ran a simulation.

Re: Study mode

#580

Post by PaulRobinson » Wed, Jul 30, 2025, 6:09 AM UTC

mvieira38 wrote: post #204

andix wrote: post #180

Earlier quoted context omitted.

Absolutely. I used to have a lot of weird IPv6 issues in my home network I didn't understand. ChatGPT helped me to dump some traffic with tcpdump and explained what was happening on the network. In the process it helped me to learn many details about RA and NDP (Router Advertisments/Neighbor Discovery Protocol, which mostly replace DHCP and ARP from IPv4). It made me realize that my WiFi mesh routers do quite a lot o…

Would you say you were using the LLM as a tutor or as tech support, in that instance?

As somebody who has done both tech support, and lectured a couple of semesters at a business school on a technical topic... they're not that far removed from each other, it's just context and audience changes. The work is pretty similar.

So why not have tech support that teaches you, or a tutor that helps with you with a specific example problem you're having?

Providing you don't just rely on training data and can reduce hallucinations, this is the angle of attack that is likely the killer app some people are already seeing.

Vibe coding is nonsense because it's not teaching you to maintain and extend that application when the LLM runs out of steam. Use it to help you fix your problem in a way that you understand and can learn from? Rocket fuel to my mind. We're maybe not far away...