I wonder if there will be a human information/knowledge equivalent of low-background steel (pre-WWII/nukes). Data from before a certain point won't be 'contaminated' with LLM stuff, but it'll be everywhere after that.
My inner conspiracy theorist can't help wonder if the continued reduction in search usefulness isn't part of an ongoing deliberate disempowerment of everyday people - but my rational side says it's merely an unfortunate emergent behaviour of the systems we've built.
"Oh, you want a guide to writing your own loss function for Tensorflow? Here's an FAQ that could have existed on comp.lang.python3.tensorflow"
Browsing today is like: “You ask for a spaghetti recipe and the page tell you the whole history of civilization.”
Allegro (big polish auction/e-commerce site) in their mobile app will unconditionally rewrite the search terms instead of showing you no results
Also, note to self to collect my favourite recipes in markdown files from now on.
As time goes on, even the amount of text I am putting out get trimmed down. Make the words count, don't count the words.
At a minimum, you'd have to validate them by confirming existence in the Wayback Machine.
Otherwise agreed that those are indeed high-signal documents. Increasing reliance on integrated educational software means that even such things as online syllabi are increasingly rare.
.edu domains can be had for any otherwise eligible "U.S.-based postsecondary institutions" per Educause: <https://net.educause.edu/eligibility.htm>
Pages at extant domains might variously be available to undergraduate or graduate students, faculty, staff, and adjuncts. Those might either directly host emulative material or be convinced or compromised into hosting content.
If there's one thing that the Internet's history to date has proved, its that perverse incentives lead to perverse consequences.
- Enroll or be hired at an eligible institution. There are literally thousands of these.
- Bribe or compromise someone enrolled or hired at an eligible institution.
- Create a de novo eligible institution. For-profit colleges are not uncommon.
Someone motivated by profit or advantage would likely find virtually any of these options quite straightforward.
I'm ... somewhat pained that this needs to be spelled out.
You don’t generally get the kind of personal website being discussed to my understanding.
> Bribe or compromise someone enrolled or hired at an eligible institution
Finding a professor willing to stake their job and reputation for such a blatantly immoral scam seems hard.
> Create a de novo eligible institution
Is it actually easy for a regular person to create their own college?
I find the snarky finish to your comment obnoxious.
- start their own country and call it Edunistan
- bribe ICANN to take over the .edu TLD
- open a university in the new country
- spend 15 years earning a PhD at that university
- reserve ~/name and start posting LLM generated content
It seems to me it relies mostly on discounting just how much we've already had to deal with this same problem in humans over the millenia.
The problem of proliferation of bad information might be getting worse, but this isn't native to generative AI. The entire informational ecosystem has to deal with this. GPTs compound the issue, but as far as I can tell, no where near what social media has forced us to deal with.
A human's lie is different than an AI's hallucination, since it's still based on (distorting) the truth, whereas the hallucination is based on an invented reality (yes I know it's applied statistics and there's no true model of the world in there, but it can report as if there is)
LLMs are no different in this respect.
LLMs on the other hand are amazing and prolific liars and can produce a lot of bullshit for a price that’s effectively free - in fact it’s cheaper to create a LLM that’s inaccurate than one that’s … less inaccurate. The truth to lie ratio on the internet is about to take a huge hit. LLMs really are a Pandora’s Box.
Humans always assemble information according to a standard of truth. It is a big part of how humans learn. No method is perfect but the human method results in fewer routine hallucinations.
I'd love you to produce data to back this up.
My guess is that you are wrong, on the basis of how often I discover that I'm full of shit and how often I discover other people are full of shit.
Humans are built for being wrong just as much as being right. We wouldn't have such complicated institutions and social structures built around controlling for those symmetric capacities if it weren't the case.
Even then, we find ourselves surrounded and overcome by falsehoods of our own design.
Embodiment and multi modal AI will likely provide such filters or limits to AI in which it can derive truth.
with AI this limit is all but removed
all the human generated bullshit ever created will soon be dwarfed by what AI can vomit out in an hour
LLMs changes all that. They can produce content on a massive scale that can drown out everything else. It doesn’t help the fact that more inaccurate LLMs are cheaper and easier to create and run - think about all the ChatGPT3 level LLMs vs ChatGPT4 level LLMs; guess which ones SEOs will gravitate towards.
> Because everyone has at least tried to lie - not necessarily for malicious reasons, e.g. white lies - a few times in their lives and know it’s not easy coming up with plausible sounding lies.
How do you know that?
Common knowledge.
Regardless, the amount of false information on the internet used to be limited by the amount of people producing it. Some of it is deliberately produced misinformation. Some of it are just mistakes due to carelessness, others due to willful negligence - example of the latter would be content farms who make their money off flooding search engine result pages and pushing ads in your face when you visit their website; who couldn’t care less what if you got what you came for.
Despite all that the internet was still useful. There are enough people posting accurate information such that the signal to noise ratio is good enough.
LLMs will change all that. They have the potential to flood the internet with hallucinated falsehood. And because bad LLMs are cheaper and easy to create and operate, those will be the majority of LLMs used by aforementioned content farms.
They won't make up a lie when telling the truth is far easier and when they do have to lie, it's a slight bit more effort.
But that's a moot point with LLMs ...
If/When they use LLMs, they will pick the cheapest LLM possible which will produce the most garbage - they might not even bother keeping it up to date; why bother when hallucinated BS sound just as convincing.
It has already started to happen.
In doing so the species would improve critical thinking skills which can be applied to all information regardless of source. Which, I agree, was often BS to begin with. But in theory would be more difficult to skirt on by without notice if humanity upgraded their critical thinking.
That soft data could have never been trusted, rhe information that can be verified (calculations etc.) seems safe from LLM
What I'm responding to is the strong tendency to discount our very long history of dealing with factually incorrect information and the ascertainment of truth from sources both dubious and trustworthy.
Entire institutions are set up in order to handle these very real problems, the set of which currently dwarfs the problem of hallucinations in GPT.
From a social perspective, non-GPT falsehoods are even more insidious, because we are inclined to trust and believe those whom we like and are like us.
Again, people are in the habit of discounting just how much we are wrong in our everyday lives. The hallucinations therefore appear more singular than they actually are.
There is a small sub-plot about how he had to give a fake persona credibility on the untrusted network in order to be able to leverage a creating a fake account on the trusted network.
Explained: <https://www.explainxkcd.com/wiki/index.php/635:_Locke_and_De...>
Reality has largely demonstrated that far more thoughtless propaganda of the Big Lie, Firehose of Bullshit (or Falsehood), associated with Russia, floods of irrelevance which tend to bury more significant stories, favoured by China, and outrage / hot-button topics, which are common in US-centric media, though a timeless technique.
Memes and simple messages attract attention and spread. Complex narratives and analyses ... not so much.
But yes, voices that deserve no attention whatsover have dominated the media landscape of the past decade or so. Not that this is entirely novel.
But yeah, maybe the idea that you can even 1% trust random content on the Internet without having a source doesn't really make sense if you think about it IMHO. Either you do this web of trust, coming from a well know real world source, or be Wikipedia-like with linked reliable sources for the viewer to check.
By the way, wasn't this how Google ranked pages back in the day? Ranking pages that get linked to higher? And even before that there were P2P web rings.
https://boingboing.net/2023/10/05/ai-search-chatbots-output-...
That being said how many people write blogs with grammerly or chatgpt these days. The temptation to use these technologies all the time is too strong for even self preservation of your own (writers) voice.
My sense is that you use this technology you might be happy with the results at first but on later review you just notice something off in some sentences and maybe it just doesn’t flow right. I’m not convinced that it will replace writers jobs yet. Especially when you want to create something authentic and unique.
I don't know about that. I have played with ChatGPT/Copilot/etc enough to know what they're capable of doing. But the thing is, I enjoy programming. I enjoy breaking down a problem and solving it with code. I enjoy crafting elegant code. So I don't use AI even though I'm fully aware it could save me hours on projects. Why? Because I enjoy those hours very much.
Why am I telling you all this? Because I suspect many writers are the same and personal blogs are their canvas. They enjoy communicating. They enjoy crafting articles. They might have AI proof-read them, but they won't let them write everything. So, to me, there is hope that personal blogs will maintain their human element, as opposed to news websites or tabloids or learning platforms.
Enjoy this luxury while it lasts. Based on what I have seen in performance review committees for software developers, your peers who drive results faster than you do because they use AI will be rewarded more and will be more likely to survive rounds of layoffs when they inevitably happen.
If the job changes that drastically I'll just have to quit and find something else.
I use ChatGPT all the time to suggest how I could make sure something isn't passive aggressive. It'll point out parts that aggression and suggested changes. It can be for a short slack message, or a many paragraph message.
Does this mean we'd end up with a finite set of verified human only data?
Would people start going through all kinds of offline archives via AI-gapped means, trying to uncover and document new sources of human input?
But who's gonna pay for it?
Assuming that semi-convincing misinformation spreads everywhere, people will finally have to find the original source of a certain statement, verify their "knowledge supply chain", and maybe use logic to evaluate every single statement made.