Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Study mode

openai.com

Post reply on HN
Print view View original

301–310 of 828 posts

Re: Study mode

#301

Post by ascorbic »

kcaseg wrote: post #290

I know it is bad for the environment, I know you cannot trust it, but as an adult learning C++ in my free time, having a pseudo-human answering my questions instead of having to look at old forum posts with people often trying to prove their skills instead of giving the simplest answer ChatGPT is something I cannot just ignore — despite being a huge LLM hater. Moral of the story: none.

If it helps you feel better, it's really not that bad for the environment. Almost certainly uses less energy than searching for lots of forum posts.

Re: Study mode

#302

Post by lottin »

czhu12 wrote: post #26

I'll personally attest: LLM's have been absolutely incredible to self learn new things post graduation. It used to be that if you got stuck on a concept, you're basically screwed. Unless it was common enough to show up in a well formed question on stack exchange, it was pretty much impossible, and the only thing you can really do is keep paving forward and hope at some point, it'll make sense to you. Now, everyone ba…

Yes. Learning assistance is one of the few use cases of IA that I have had success with.

Re: Study mode

#303

Post by cthalupa »

SLWW wrote: post #210
pyuser583 wrote: post #203

Earlier quoted context omitted.

> Ask it about the Crusades, often if forgets the sacking of St. Peter's in Rome around 846 AD, usually painting the Papacy as a needlessly hateful and violent people during that specific Crusade. Which was horrible, bloody as well as immensely destructive (I don't defend the Crusades), but paints the Islamic forces as victims, which they were eventually, but not at the beginning, at the beginning they were the aggre…

The crusades were predicated on historic tensions between Rome and the Arabs. Which is why I mention that, while the First Crusade proper was in 1096, it's core reasoning were situations like the Sacking of St. Peters which is considered by historians to be one of the most influential moments and often was used as a justification as there was a history of incompatibilities between Rome and the Muslims. Further leadin…

Why should we consider something that happened 250 years prior as some sort of affirmative defense of the Crusades as having been something that started with the Islamic world being the aggressors?

If the US were to start invading Axis countries with WW2 being the justification we'd of course be the aggressors, and that was less than 100 years ago.

Re: Study mode

#304

Post by aryamaan »

It is surprising that it is prompt based model and not RLHF.

I am not an LLM guy but as far as I understand, RLHF did a good job converting a base model into a chat model (instruct based), a chat/base model into a thinking model.

Both of these examples are about the nature of the response, and the content they use to fill the response. There are so many differnt ways still pending to see how these can be filled.

Generating an answer step by step and letting users dive into those steps is one of the ways, and RLHF (or the similar things which are used) seems a good fit for it.

Prompting feels like a temporary solution for it like how "think step by step" was first seen in prompts.

Also, doing RLHF/ post training to change these structures also make it moat/ and expensive. Only the AI labs can do it

Re: Study mode

#305

Post by jennyholzer »

ZYbCRq22HbJ2y7 wrote: post #128
avn2109 wrote: post #57

Earlier quoted context omitted.

LLM's are useful because they can recommend several famous/well-known books (or even chapters of books) that are relevant to a particular topic. Then you can also use the LLM to illuminate the inevitable points of confusion and shortcomings in those books while you're reading and synthesizing them. Pre-LLM, even finding the ~5 textbooks with ~3 chapters each that decently covered the material I want was itself a nont…

> they can recommend several famous/well-known books They can recommend many unknown books as well, as language models are known to reference resources that do not exist.

I strongly prefer curated recommendations from a person with some sort of credibility in a subject area that interests me.

Re: Study mode

#306

Post by danenania »

gh0stcat wrote: post #208
simonw wrote: post #198

I think I got the system prompt out for this (I tried a few different approaches and they produced the same output): https://gist.github.com/simonw/33d5fb67d6b8e1b1e2f6921ab0ccb... Representative snippet: > DO NOT GIVE ANSWERS OR DO HOMEWORK FOR THE USER. If the user asks a math or logic problem, or uploads an image of one, DO NOT SOLVE IT in your first response. Instead: *talk through* the problem with the user, one…

I love that caps actually seem to matter to the LLM.

I've found that a lot of prompt engineering boils down to managing layers of emphasis. You can use caps, bold, asterisks, precede instructions with "this is critically important:", and so on. It's also often necessary to repeat important instructions a bunch of times.

How exactly you do it is often arbitrary/interchangeable, but it definitely does have an effect, and is crucial to getting LLMs to follow instructions reliably once prompts start getting longer and more complex.

Re: Study mode

#307

Post by SLWW »

scarmig wrote: post #179
SLWW wrote: post #162

Earlier quoted context omitted.

That's a fun idea; almost having it "play pretend" instead of directly asking it for strong anti-heliocentric arguments outright. It's weird to see which topics it "thinks" are politically charged vs. others. I've noticed some inconsistency depending on even what years you input into your questions. One year off? It will sometimes give you a more unbiased answer as a result about the year you were actually thinking o…

I think the first thing is figuring out exactly what persona you want the LLM to adopt: if you have only a vague idea of the persona, it will default to the laziest one possible that still could be said to satisfy your request. Once that's done, though, it usually works decently, except for those that the LLM detects are politically charged. (The weakness here is that at some point you've defined the persona so stric…

> The weakness here is that at some point you've defined the persona so strictly that it's ahistorical and more reflective of your own mental model.

That's a good thing to be aware of, using our own bias to make it more "likely" to play pretend. LLMs tend to be more on the agreeable side; given the unreliable narrators we people tend to be, and the fact that these models are trained on us, it does track that the machine would tend towards preference over fact, especially when the fact could be outside of the LLMs own "Overton Window".

I've started to care less and less about self-censoring as I deem it to be a kind of "use it or lose it" privilege. If you normalize talking about censored/"dangerous" topics in a rational way, more people will be likely to see it not as much of a problem. The other eventuality is that no one hears anything that opposes their view in a rational way but rather only hears from the extremists or those who just want to stick it to the current "bad" in their minds at that moment. Even then though I still will omit certain statements on some topics given the platform, but that's more so that I don't get mislabeled by readers. (one of the items on my other comment was intentionally left as vague as possible for this reason) As for the LLMs, I usually just leave spicy questions for LLMs I can access through an API of someone else (an aggregator) and not a personal acc just to make it a little more difficult to label my activity falsely as a bad actor.

Re: Study mode

#308

Post by danenania »

simonw wrote: post #248
can16358p wrote: post #239

Earlier quoted context omitted.

If I were OpenAI, I would deliberately "leak" this prompt when asked for the system prompt as a honeypot to slow down competitor research whereas I'd be using a different prompt behind the scenes. Not saying it is indeed reality, but it could simple be programmed to return a different prompt from the original, appearing plausible, but perhaps missing some key elements. But of course, if we apply Occam's Razor, it mig…

That kind of thing is surprisingly hard to implement. To date I've not seen any provider been caught serving up a fake system prompt... which could mean that they are doing it successfully, but I think it's more likely that they determined it's not worth it because there are SO MANY ways someone could get the real one, and it would be embarrassing if they were caught trying to fake it. Tokens are expensive. How much…

Probably the only way to do it reliably would be to intercept the prompt with a specially trained classifier? I think you're right that once it gets to the main model, nothing really works.

Re: Study mode

#309

Post by brulard »

jennyholzer wrote: post #296
GaggiX wrote: post #83

Earlier quoted context omitted.

And the old ChatGPT was nothing compared to what we have today, nowadays reasoning models will eat through math problems no problem when this was a major limitation in the past.

I don't buy it. Open AI doesn't come close to passing my credibility check. I don't believe their metrics.

You don't have to. Just try it yourself.

Re: Study mode

#310

Post by gojomo »

no_wizard wrote: post #49
czhu12 wrote: post #26

I'll personally attest: LLM's have been absolutely incredible to self learn new things post graduation. It used to be that if you got stuck on a concept, you're basically screwed. Unless it was common enough to show up in a well formed question on stack exchange, it was pretty much impossible, and the only thing you can really do is keep paving forward and hope at some point, it'll make sense to you. Now, everyone ba…

>Now, everyone basically has a personal TA, ready to go at all hours of the day This simply hasn't been my experience. Its too shallow. The deeper I go, the less it seems to be useful. This happens quick for me. Also, god forbid you're researching a complex and possibly controversial subject and you want it to find reputable sources or particularly academic ones.

Grandparent testimony of success, & parent testimony of frustration, are both just wispy random gossip when they don't specify which LLMs delivered the reported experiences.

The quality varies wildly across models & versions.

With humans, the statement "my tutor was great" and "my tutor was awful" reflect very little on "tutoring" in general, and are barely even responses to each other withou more specificity about the quality of tutor involved.

Same with AI models.

Post reply on HN