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

Storm: LLM system that researches a topic and generates full-length wiki article

github.com

Post reply on HN
Print view View original

91–100 of 104 posts

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#91

Post by mistrial9 »

lukev wrote: post #27

I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.) Even if they do it well and truthfully (which they don't) current LLMs can only summarize, digest, and restate. There is no non-transient value add. LLMs may have a place to help que…

excellent virtue signalling here -- however, commercial publishers, competitive attorneys, advertising sales and others are literally falling over themselves in an avalanche of doing exactly this, that you advise against (politely).

This moment reminds me very much of the late 1990s when it was common knowledge that "claim jumping a domain name" is very rude and not advisable, or the common knowledge among intellectuals that "ads will ruin the Internet" .. yes, polite people did not make companies to claim jump domain name registration, or push annoying and repetitive ads on the Internet..

but..

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#92

Post by skissane »

msp26 wrote: post #67
CuriouslyC wrote: post #63

Earlier quoted context omitted.

I think bootstrapping documentation with LLM output is a great practice. It's a wiki, people can update it from a baseline, just as long as they can see what was LLM generated to know that it shouldn't be taken as absolute truth. The hardest part of good documentation is getting started. Once there are docs in place it's usually much easier to revise and correct than it would have been to write correctly by hand the…

Maybe the generated text could be a slightly different colour until it's verified. But you'd have to make sure there's no easy way of verifying everything mindlessly without having read it.

Technique I've found helpful personally: get the LLM to generate text in small chunks (e.g. a paragraph at a time). After generating each chunk, it is immediately reviewed by a human, who can edit it manually to correct any mistakes, ask the LLM to try again, or prompt the LLM to make specified changes. When the human is satisfied with that chunk, it is saved, and we move on to the next one.

Sometimes, the output the LLM generates is correct and I'm just approving it. Other times, it is mostly right, and I can easily identify and correct its errors. Yet other times, it is totally wrong, but often typing out why it is wrong is a good start to actually generating correct text. Often (but not always), the kinds of false assumptions which LLMs make are similar to those a human reader would make, so stuff like "A and B sound very similar but, in the context of this system, actually have completely different meanings" is a useful addition to documentation anyway.

The worst case scenario is the LLM generates something which is subtlety wrong, and the human review fails to pick up on the subtle error. But, that's something which can happen even with no LLMs involved at all. It isn't uncommon for people to make subtle errors in documents they write (often because they are misremembering something) and for those subtle errors not to be picked up during the review process. I'm not convinced the odds of this happening with an LLM-assisted workflow are significantly greater than with a purely human workflow.

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#93

Post by elicksaur »

lukev wrote: post #27

I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.) Even if they do it well and truthfully (which they don't) current LLMs can only summarize, digest, and restate. There is no non-transient value add. LLMs may have a place to help que…

How could it be true that the content generated would have value only if it is not persisted?

If it doesn’t have value for being saved and published, why would it have value for the person viewing it ephemerally?

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#94

Post by skissane »

petercooper wrote: post #58
lukev wrote: post #27

I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.) Even if they do it well and truthfully (which they don't) current LLMs can only summarize, digest, and restate. There is no non-transient value add. LLMs may have a place to help que…

current LLMs can only summarize, digest, and restate. There is no non-transient value add. Though, at a stretch, Wikipedia itself could be considered based around summarization, digesting, and restating/citing things said elsewhere, given its policy of verifiability: "Even if you are sure something is true, it must have been previously published in a reliable source before you can add it." Now, LLMs aren't well known…

I remember one time, I wanted to update a Wikipedia article with some more recent developments, but was having trouble working out the best way of wording them. I found a newspaper article discussing those developments, so I scribbled down some notes summarising the article's content, and then asked ChatGPT to reword them for me more eloquently. And its output was good enough for me to paste in to the Wikipedia article with only minor adjustments, along with a cite to the newspaper article. I'm sure I'm not the only person to have done something like that.

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#95

Post by xbmcuser »

lukev wrote: post #27

I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.) Even if they do it well and truthfully (which they don't) current LLMs can only summarize, digest, and restate. There is no non-transient value add. LLMs may have a place to help que…

I would love for something like this to be attached to LibGen where it reads the millions of scientific papers. As in my opinion human knowledge today is more than what a group of people can handle let alone individuals. Their is lot of domain specific knowledge that would translate and be used in other domains but unless by chance a human with speciality in both domains sees it will not get ported or assimilated in the 2nd domain.

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#97

Post by visarga »

__loam wrote: post #83
visarga wrote: post #70

Earlier quoted context omitted.

> current LLMs can only summarize, digest, and restate. There is no non-transient value add. No, you're wrong. LLMs create new experiences after deployment, either by assisting humans, or by solving tasks they can validate, such as code or game play. In fact any deployed LLM gets to be embedded in a larger system - a chat room, a code running environment, a game, a simulation, a robot or inside a company - it can lea…

https://twitter.com/itsandrewgao/status/1689634145717379074?... Yeah this is really having a positive impact on scientific discovery.

Your link only shows what unscrupulous people would do.

Here is a LLM with search doing competitive level coding:

https://deepmind.google/discover/blog/competitive-programmin...

and in general, applying evolutionary methods on top:

https://scholar.google.com/scholar?cites=1264402983985539857...

The explanation is simple - either learn from past experience which is human text for now, or learn from present time experience which comes from the environment. The environment is where LLMs can do novel things.

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#98

Post by echo8899 »

whitehexagon wrote: post #11

Hmm something about this title containing the word 'research' disturbs me. I associate that word with rigorous scientific methods that leads to fact based knowledge or maybe some new hypothesis, not some LLM hallucinating sources, references, quotes and all the other garbage they spit out when challenged over a point of fact. Horrifying to think peeps might turn towards these tools for factual information.

Is there any better word in your mind instead of "research"?

Re: Storm: LLM system that researches a topic and generates full-length wiki article

#100

Post by vineyardmike »

elicksaur wrote: post #93
lukev wrote: post #27

I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.) Even if they do it well and truthfully (which they don't) current LLMs can only summarize, digest, and restate. There is no non-transient value add. LLMs may have a place to help que…

How could it be true that the content generated would have value only if it is not persisted? If it doesn’t have value for being saved and published, why would it have value for the person viewing it ephemerally?

Obviously LLMs aren't deterministic, but they generally can be modeled as a simple, stateless, function: Given text, generate text. Depending on your views towards wasting energy, why store something that can be "trivially" re-generated?

One reason to NOT want to persist an LLM output is to avoid contaminating the internet with LLM text that could be confused for human text. This is useful for historical reasons (what did people say/think) and for future LLM training purposes.

In the future, there might be other reasons to not persist LLM-generated text. If you store the inputs instead of outputs then it can be "replayed" with a new LLM. I can generate some text via an LLM running on my phone, but if I later wanted a higher-quality output, it'd be easier to re-generate the data on my laptop later (or the cloud, obviously). One related idea that comes to mind is personalization - let the LLM use language each reader is familiar with, or focusing on different bits of info to different people. Maybe if an LLM knows I have a PhD in a certain subject, it'd summarize a research paper differently than it'd summarize it for my neighbor who has different preexisting knowledge.

Post reply on HN