[HN Gopher] Analyzing the websites that are used to train AI cha...
___________________________________________________________________
Analyzing the websites that are used to train AI chatbots
Author : wawayanda
Score : 64 points
Date : 2023-04-19 12:05 UTC (1 days ago)
(HTM) web link (www.washingtonpost.com)
(TXT) w3m dump (www.washingtonpost.com)
| RugnirViking wrote:
| whoa, they claimed deviantart is a text-to-image generator being
| sued by artists? is that really a thing? Last I heard Deviantart
| was really against it
| Nicksil wrote:
| https://news.ycombinator.com/item?id=33567202
| archon1410 wrote:
| > more than 72,000 instances of "swastika," one of the banned
| terms from the list.
|
| Isn't swastika being a common sacred symbol in Eastern and
| specially Dharmic religions common knowledge? This, alongwith
| most of the religious websites being Christian, Islamic and so on
| is quite concerning.
| bufo wrote:
| This article is so misleading. First there's nothing secret about
| all this. Also these chat bots don't just "sound" smart, they are
| providing value to millions of users.
| dang wrote:
| OK, but the article itself is (or at least may be?) more
| interesting than the misleading title. Let's focus on that
| instead. We've edited the title now.
|
| From https://news.ycombinator.com/newsguidelines.html: "
| _Please don 't pick the most provocative thing in an article or
| post to complain about in the thread. Find something
| interesting to respond to instead._"
| Turing_Machine wrote:
| > First there's nothing secret about all this.
|
| Right. It's going to come as a shock to exactly no one that
| (e.g.) Wikipedia was used as a source.
|
| Then there's the fear-mongering about using public databases
| (e.g., voter registration lists) in "unforeseen ways". While
| that may be true, they don't seem to be concerned about non-AI
| actors using these databases in "unforeseen ways".
|
| Pretty much pure clickbait.
| pier25 wrote:
| It's disheartening to see AI companies profiting from free
| content I've created online over years and years. StackOverflow,
| forums, Github, even my personal crappy website is there.
| geoduck14 wrote:
| Can we mess with ChatGPT by posting wrong answers on SO?
| tyronehed wrote:
| [dead]
| newaccount74 wrote:
| When asking ChatGPT about programming questions, it feels pretty
| similar to asking questions on stack overflow, except you get
| answers instantly. I assume that Q&A sites must be among the most
| important training data.
| ChatGTP wrote:
| Also gists and GitHub issues, they would've had near perfect
| data sets of everything they had permission to touch at GitHub.
|
| I saw something on Twitter, where someone was offered a job at
| OpenAI to write code for common coding problems and provide
| very verbose reasoning and comments. Can't find it now..
| Entinel wrote:
| I am shocked that Discord is not on this list. I would bet my
| mortgage that Discord would sell their chats as a dataset to
| train on if the price was right.
| EGreg wrote:
| I often get downvoted for suggesting this, but why not simply
| host all your chats, videoconferencing etc. on your own
| website? Make it a gated community and only post little teasers
| to the likes of Instagram and YouTube and TikTok and Twitter
| and Facebook. Let them come to YOUR site to participate and
| abide by YOUR terms. Why make a Facebook Group or a Discord
| server?
|
| Why do so many of us, including the biggest influencers in the
| world, spend so much money only to post all of our content on
| the Big Tech's sites and infrastructure? You give all your
| social capital to them too, and they turn around and sell it
| back to you (access to your own followers) and can deplatform
| you. They encourage you to pay them to attract eyeballs away
| from other videos, and vice versa, in a zero-sum game on their
| own platform.
|
| They now increasingly even control our conversations and our
| democracy. Why not have these conversations on our own forums,
| powered by our own software, the way we do on HN?
|
| We gave away all our social capital and our content, to these
| platforms. And then they turn around and train big data on it
| internally, and scrape it for their AIs.
|
| People have complained for years yet do nothing about it:
| https://readwrite.com/mark-cuban-facebooks-sponsored-posts-a...
|
| I ask again: why not simply host it ourselves?
| eropple wrote:
| As it happens, I am in the process of building something
| pretty close to what you describe. Here is an extremely
| partial list of reasons people don't do it.
|
| - Written content: because network effects are real. Also
| because building a blog in 2023 that doesn't have awful look-
| and-feel or seem like a spam site is remarkably hard.
| Substack is full of chuds, but you can firewall off from them
| most of the time and at least you can _write things_ instead
| of _writing code to write things_. WordPress is maybe the
| least bad option today if you want interactivity (which I
| would assume you do, what with the "community" bit and all),
| and getting things like SSO (for playing nice with all those
| other things in that gated community) costs money for
| dubiously-supported plugins that claim to do things they
| might not actually do. Ghost has a ton of weird limitations
| that are often claimed to be by design, so unless you want a
| plain white (or, if you're bold, black) site that's Just Text
| and maybe a few images, it's not great either. Static sites--
| community, no bueno.
|
| - Podcasts: this one is _not_ "because network effects are
| real", this one's just "because open-source podcast hosting
| software basically doesn't exist and what does is bad at
| doing the job." Meanwhile, anchor.fm is free.
|
| - Chats: because network effects are real. And because the
| options are mostly unpleasant. IRC is difficult and mobile-
| unfriendly. Mattermost is buggy. Zulip is pretty good, but
| now you're a sysadmin and most people on Discord just want to
| shitpost about Pokemon. (I am not doing this, either, because
| there are limited hours in the day.)
|
| - Videoconferencing: because network effects are real. (You
| may be detecting a theme here!) _And_ because self-hosting
| Jitsi is an enormous pain in the rear. And, again, now you
| 're a sysadmin, and most people on Discord just want to
| shitpost about Pokemon. (Might do this one, just for
| latency's sake when recording stuff. Haven't decided yet. It
| is hard. It is also pretty expensive to run.)
|
| - Video content: because network effects are real. Also
| because I don't own a CDN, and because YouTube is $free
| dollars per minute watched. Even my former employer, whose
| stack I quite like, is not $free dollars per minute watched--
| nor should they be! They're targeting businesses. (Needless
| to say: definitely not doing this one.)
|
| Gated communities are great, but _somebody_ needs to make
| achievable these things for normal people, and they 're
| probably going to want money for it. And while I've written
| plenty of open source software, the stuff I'm building won't
| really be portable to somebody else because I'm not making
| decisions for others' consumption. I'm making decisions for
| exactly my (relatively high) tolerance for system
| administration and exactly my (relatively high) ability to
| fix things when they break. I'm going to open-source parts of
| it, but I can _only_ open-source parts, and they 'll still
| depend on my CMS of choice, etc. etc. unless I spend time not
| building my thing and instead building
| abstraction/indirection layers for hypothetical other people.
|
| If you want people to do similar, maybe it should be you who
| goes and builds the path to doing it?
| Turing_Machine wrote:
| > I ask again: why not simply host it ourselves?
|
| Because the average person and the average HN user are
| different sets of people. The typical person who wants to
| start a forum on, say, trout fishing or stamp collecting has
| only a vague notion of a server _is_ , much less how to set
| one up and maintain it. Even if they did, they'd have little
| or no interest in doing those things. It's far easier for
| them to just make a Facebook group or whatever.
| JohnFen wrote:
| But if the HN crowd runs their own servers, then the non-HN
| crowd could use them. Everybody wins.
| grepfru_it wrote:
| The intersection of people who can host it themselves vs
| those who host chats, video conferences, etc events is very
| very small. Once they start trying to figure out the SSH host
| keys and TLS firewalls, they will become very quickly
| overwhelmed and out of their depth. They would likely turn to
| an expert at this point, but it has become far cheaper and
| easier to just turn to a third-party service to handle this.
| The very service you ask them to dismiss for the troubles of
| hosting themselves.
|
| On the other end of the spectrum, I have dedicated systems to
| hosting my own chats, emails, source code, and more. My user
| count is super low. Seems that not everyone wants to come to
| my copy-cat Gitea server when "all the cool kids" are on
| GitHub
| EGreg wrote:
| Forget the SSH keys and TLS firewalls. What open source
| software would you actually use, to start to match Facebook
| Groups or Discord Forums in features?
| grepfru_it wrote:
| Flarum and (Mattermost or RocketChat)
| tsukikage wrote:
| Because we're not content with just a private little soapbox
| in our private little space talking to the few hipsters we
| can convince to drop by; we are vain and want to talk to the
| crowds, and the crowds are all over /there/.
| tivert wrote:
| > Why do so many of us, including the biggest influencers in
| the world, spend so much money only to post all of our
| content on the Big Tech's sites and infrastructure? You give
| all your social capital to them too, and they turn around and
| sell it back to you (access to your own followers) and can
| deplatform you. They encourage you to pay them to attract
| eyeballs away from other videos, and vice versa, in a zero-
| sum game on their own platform.
|
| Because we're lazy and Big Tech solves the immediate, short
| term problem.
| dpkirchner wrote:
| > I ask again: why not simply host it ourselves?
|
| Cost, ongoing maintenance, convenience, existing apps are
| good enough, existing infrastructure is good enough, I could
| probably go on for a while.
| operator-name wrote:
| Self hosting is hard.
|
| Self hosting with channels, voice chats and permissions is
| harder. Discord also had one unified account to join
| different channels, private message, etc.
| postdb wrote:
| Glad someone mentioned this!!! It is one of good reason to
| self-host! And only send sneak peek to social media sites.
| Stallman has a recommendation similar to that on what to post
| to FB.
|
| Only use social media as a
| notification/publishing/advertising system, post important
| stuff on your own site.
| EGreg wrote:
| Can you link to the Stallman suggestion you referenced
| please?
| postdb wrote:
| https://stallman.org/facebook-presence.html
| tiberious726 wrote:
| Agreed, I'm shocked he suggested interacting with
| Facebook at all. Their site uses proprietary JavaScript
| postdb wrote:
| Of course, I should clarify, it is his recommendation if
| organisations who need a facebook presence, as many small
| businesses or event organizers struggle to not have a
| FB/instagram presence.
| EGreg wrote:
| Can you Link to where he says this recommendation ?
| fnordpiglet wrote:
| Because they act as a portal and distribution service. The
| same reason freelance journalists don't print their own
| newspapers but instead publish via either a syndicate or
| directly through papers. The paper has the brand,
| distribution, ability to monetize and pay journalists for
| their content, layout staff, etc. Similarly say YouTube is
| YouTube in the same way the NYT is the NYT - people go to
| YouTube for content, it provides hosting and distribution as
| well as ways to promote your content, monetize it, they
| provide a UX and technical infrastructure, etc.
|
| There is _nothing_ stopping _anyone_ in any RFC or other
| technical standard from hosting their own content on _equal
| footing_ with YouTube. People _do_ host stuff on their own
| machines and sites. You're just not directly aware of it
| because it's basically impossible to find relative to the
| ease of content discovery on YouTube. No one is getting rich
| or even paying their bills because on self hosted because
| beyond having no funnel monetization of content is actually
| kinda hard.
|
| Case in point, politics aside, Parler and Truth Social were
| basically formed based off similar thoughts. Why do we let
| these big companies make content moderation policies? I'd
| like a place where I can be abusive and racist or whatever
| without fear of censorship. I'm after all entitled to hold
| those views and say them. But, it turns out, it's hard to
| provide the quality of service of say Facebook on the premise
| of absolute freedom - especially when it's so toxic most
| people are turned off. This gets to the point: you can host
| yourself, but if you want to build a "free YouTube" you
| probably need to ask why did YouTube, a for profit entity,
| take the positions it has in the market? Because the market
| rewards them for making a family friendly middle ground
| platform that provides easy access to personalized content
| and a simple experience for content creators to distribute
| and monetize their content. Do that on your own systems, and
| you'll probably do alright - except one thing. They have the
| most users. Acquiring users, or in news papers, subscribers,
| is not easy or cheap.
|
| So, that's why things are the way they are. But you're
| absolutely able to self host. Just don't expect everyone to
| show up at your random website.
| EGreg wrote:
| Absolutely able to self-host? What open source software
| would I use to match the features of, say, Facebook Groups
| or a Discord Channel?
| gomox wrote:
| When ChatGPT first came out, I was surprised at the depth of
| information it had about region-specific market sizing in our
| (relatively niche) industry.
|
| Turns out the model had ingested a pay-to-read article that cost
| upwards of $2000 [0] and was quoting the figures in it and
| referencing it directly (i.e. attributing the info to the article
| in question).
|
| I knew about the paper but had never purchased it. I was actually
| surprised I could access the data through ChatGPT.
|
| A few days later the same information was gone. I assumed someone
| might have decided to keep it on the low regarding having
| ingested all of these sources of restricted access content. I now
| get a generic blurb about the same question.
|
| The opacity of it all made me a little worried. What similar
| models exist, trained on actually private information, for more
| nefarious purposes?
|
| [0] https://www.marketresearch.com/Infiniti-Research-
| Limited-v26...
| michaelbuckbee wrote:
| One of my friends was trying to figure out "how" ChatGPT knew
| about tried something I hadn't thought out, just ask it "do you
| know paper xyz by author abc?" and in fact it had.
| soco wrote:
| Just saying, if ChatGPT says it knows something it doesn't
| necessarily mean it knows something, or that said thing is
| even real.
| richbell wrote:
| I once asked ChatGPT if it knew about an esoteric
| programming language. It said "yes!" and then proceeded to
| quote information and code snippets for Python.
| Mizza wrote:
| I was hoping this article would be about that. It's not. It's a
| waste of time spent pearl clutching about how the authors find
| some small percentage of the internet/Common Crawl "troubling."
| I don't know what they expected.
| PaulHoule wrote:
| If you want a model to, say, never draw pornography or write
| a Hitler speech, you don't do that by excluding "bad" content
| from the foundation model but rather you tackle that in later
| stages of training, particularly in this phase
|
| https://en.wikipedia.org/wiki/Reinforcement_learning_from_hu.
| ..
|
| Having some of that in the foundation model actually helps
| the model "understand" what it is it is not supposed to do.
| The whole point of the foundation model is that it
| generalizes from the examples you show it later to similar
| examples it saw during "pre-training" and a certain about of
| offensive content will help it learn what it is offensive
| much quickly later.
| Sunhold wrote:
| Unless you actually checked the article, ChatGPT likely
| hallucinated the figures and picked a plausible-sounding
| source. Most likely, nothing changed when you tried a few days
| later. The responses are stochastic due to the temperature
| setting. If anything did change, it was probably an update to
| reduce hallucination.
| gomox wrote:
| I never saw the actual article, but the quoted figures were
| not just plausible and reasonable but also internally
| consistent throughout several exchanges. It produced exact
| figures for key (i.e. largest) submarkets and growth
| estimates for smaller ones, in the same way that I would
| expect the original content to be structured.
|
| I would bet with a 95%+ confidence that it parroted the
| actual contents. $2000 for that extra 5% tho.
| aaron695 wrote:
| [dead]
| ipaddr wrote:
| That might have come from someone else making it freely
| available
| BiteCode_dev wrote:
| Imagine how powerful an ai would be if we would give it access
| to all the info behind firewalls.
| tsukikage wrote:
| While we're at it, imagine how awesome an ordinary search
| engine could be if we would give it access to all the info
| behind firewalls.
| htss2013 wrote:
| It states that if an LLM is good at the LSAT, it must have
| trained on many practice LSAT tests. How do we know this? The
| test measures things like reading comprehension and logical
| reasoning. LLMs have reasoning ability with novel material, and
| there are plenty of ways to train logical reasoning outside of
| LSAT test prep materials.
| mdrzn wrote:
| https://archive.ph/kNEyR
___________________________________________________________________
(page generated 2023-04-20 23:01 UTC)