AI tools are fantastic for self-teaching

[tom66]

I just discovered that both Claude and Gemini can run automatic quizzes on subject areas; you just ask them to quiz you on something, and they'll give you 10-ish questions to answer, explaining why an answer is correct or incorrect, and generating more detailed explanations if you ask.  This is going to be really useful for understanding new concepts.  How many people are using this to improve their knowledge?

[Siwastaja]

Not in form of quiz, but yes, when I'm wondering about something, nowadays I'll just chat about it with AI. The contrast to trying to learn the same thing with googling is massive; partially this is because AIs truly are so great, but partially because the baseline (Google) sucks so colossally.

In fact, it's interesting how Googling is not much more effective, or at all, than actually going to a library (travel included), finding the right physical books, and reading through them until you find what you need. Expect to spend tens of hours to build some basic-level understanding either way. Classic Google (web pages) human-slop/noise is such a huge problem, and Google always sucked filtering high-quality content from the large mass of human slop.

With AI, this is different. You can build the knowledge and ask the right questions within an hour of chat. And unlike on forums like this (and this is one of the best - most are much worse), where people misunderstand you if you don't write a perfect question, and even if you do, many just answer even if they didn't understand the question, creating human slop, AI seems to get the gist of your question even from a misformed prompt.

[Tation]

Here using Claude & Gemini as an aid when learning a foreign language.

[johnboxall]

NotebookLLM is also great for studying and making quizzes.

[nctnico]

I just discovered that both Claude and Gemini can run automatic quizzes on subject areas; you just ask them to quiz you on something, and they'll give you 10-ish questions to answer, explaining why an answer is correct or incorrect, and generating more detailed explanations if you ask.  This is going to be really useful for understanding new concepts.  How many people are using this to improve their knowledge?

This has not occured to me. Great tip!

[paulca]

This has been recommended to me: https://obsidian.md/ it's basically a "Mind map" type notebook.  However it integrates with claude code, such that you can set down "key notes" and lay them out as you like.

[mORAV]

To be honest, it is useful if you want to enter an argument.
If you know that a lot of people (historically) have discussed a topic it is an excellent source of information.

I often like to refresh my chats and start with a wrong assumption (that I know is wrong), and wait for the LLM to correct me and then proceed from that point.

It is often the case that if you search for something you don't know much about, it is easy to assume the first result you find is correct.

[Siwastaja]

I often like to refresh my chats and start with a wrong assumption (that I know is wrong), and wait for the LLM to correct me and then proceed from that point.

Just two years ago everyone was sure LLMs can't never work like this because they are so sycophantic and just restate what user said, adding supporting, hallucinated details.

Clearly this has changed. If you know the LLM is wrong, you need to provide some serious reasoning/proof before it fully accepts it.

[mORAV]

Clearly this has changed. If you know the LLM is wrong, you need to provide some serious reasoning/proof before it fully accepts it.

Yes and no, the "refresh" part is quite important. If the wrong assumption gets into the context windows too many times it's impossible to move it from that position, this gets pretty evident is some type of research chats (in my job need to find some obscure electronic components, processes and machinary). If the LLM get the wrong idea itself i can't move from it and even if i do it will come back to the same point a few messages later...

As tools...they are really weird.

The fact that i know how they works but BARELY make them work makes me feel despair for the "average person" who knows nothing about the tech.

[Siwastaja]

Yes and no, the "refresh" part is quite important. If the wrong assumption gets into the context windows too many times it's impossible to move it from that position, this gets pretty evident is some type of research chats (in my job need to find some obscure electronic components, processes and machinary). If the LLM get the wrong idea itself i can't move from it and even if i do it will come back to the same point a few messages later...

This sometimes happens, sometimes not. I suspect it happens when the training data (actual model weighs) support the wrong idea. Then it needs a lot of pushing in the context to turn it around, and small amount of pushing towards the wrong idea is amplified by the training.

[paulca]

Its also likely "equal weight inversion".  "Do not do the thing" has equal weight to do the thing.  They are not as distinct as they should be.  Both include "the thing".  So "the thing" is now in context.  "Don't think of a pink elephant.", you did didn't you?

[Siwastaja]

Its also likely "equal weight inversion".  "Do not do the thing" has equal weight to do the thing.  They are not as distinct as they should be.  Both include "the thing".  So "the thing" is now in context.  "Don't think of a pink elephant.", you did didn't you?

This is exactly why long lists of don't do this, don't do that, don't don't don't - don't work well with AI, and don't work well with humans (kids or adults). People just ignore such instructions and do whatever they want, this includes professions like pilots or doctors.

If doing something is in the training, really the best thing is to allow doing it, and direct towards certain methods with positive examples ("try doing X first"), then the other way will be a natural fallback. "Don't"s is mostly micromanaging, and micromanaging is usually harmful. Should be a special case, not the primary control strategy.

[paulca]

Its also likely "equal weight inversion".  "Do not do the thing" has equal weight to do the thing.  They are not as distinct as they should be.  Both include "the thing".  So "the thing" is now in context.  "Don't think of a pink elephant.", you did didn't you?

This is exactly why long lists of don't do this, don't do that, don't don't don't - don't work well with AI, and don't work well with humans (kids or adults). People just ignore such instructions and do whatever they want, this includes professions like pilots or doctors.

If doing something is in the training, really the best thing is to allow doing it, and direct towards certain methods with positive examples ("try doing X first"), then the other way will be a natural fallback. "Don't"s is mostly micromanaging, and micromanaging is usually harmful. Should be a special case, not the primary control strategy.

I have a tendency to notice it in a awkwardly meta way by the feature inverting itself into my favour.

If I am typing fast and not paying full attention I will actually invert qualifiers by accident usually by omission of a suffix.  Saying "Would" instead of "Wouldn't" or just forgetting to type the "not" word as I'm thinking faster than Im typing and compression occurs.

When this happens with an AI chat, it just assumes straight through it to the contextual aligned "state" inverted or non-inverted.

"Paragraph describing a undesirable state.  I would want that."

It will auto invert the later into "I wouldn't want that." and reply from that direction 9 times out of 10.  In a "chat" this is fine and you can just ignore it.  However if you are trying to set rules it should be noted that whether you put the "n't" or not is almost meaningless to it.