The End Of StackOverflow
ฝัง
- เผยแพร่เมื่อ 8 พ.ค. 2024
- StackOverflow is being eaten by ChatGPT. Now it's desperately trying to catch up. Will this be the end of StackOverflow (and maybe the rest of the internet)? Idk man I just talk about stuff and write code
SOURCES
/ 1787467736097939562
/ 1787596365008380306
observablehq.com/@ayhanfuat/t...
Check out my Twitch, Twitter, Discord more at t3.gg
S/O Ph4se0n3 for the awesome edit 🙏 - วิทยาศาสตร์และเทคโนโลยี
I am somehow at a point where more of my answers are somewhere in a comment in a open GitHub issue.
same
Painfully relatable, I'm there most of the time.
but of course, it's hard to actually find the solution, it's normally berried deeper down or doesn't look like it will work at first glance.
It’s because Stack Overflow is more fondamental questions, leaving a comment to copilot will solve it. We are having library related issues
Absolutely true. Chat is nowhere near opinionated enough to have answers to most of my questions.
Yo, they gonna train the AI to viciously berate you for asking simple questions or just reject your prompt because it was answered elsewhere??
This. It's a pretty toxic community to try and get into, it's like Reddit for nerds.
Seriously we all have like 10 stackoverflow accounts all with negative cred for asking crap we needed to know just to do our damn jobs. None of it was ever personal to me, but to them, yikes.
@@GetFitEatRightthat’s just not true. It’s not all that bad. I have 1 account, 400 something score, 15 questions. And I’ve been a full time professional dev for like 20 years. I only asked questions when I’m fully stuck, never really had any bad responses. Couple of downvoted-for-no-reason posts, but what do I care?
There’s nothing wrong with asking questions, but if you have 10 accounts with negative standing, maybe your tone is off, or you’re not using the platform as it was intended. It’s not a “hold my hand and help me anytime I’m trying to do things” forum, ideally questions (and answers) are useful for others, and you’re expected to research things yourself.
@@svenmifyI've seen many legitimate questions be shut down with a "duplicate" and a link to something completely unrelated. That alone just makes me not want to interact with the website.
I've also been downvoted when providing the best answer in the thread because someone else had a vaguely similar answer that didn't cover the entire problem. I was told "write a comment with your feedback to the other answer" but you can only do that with a good enough standing and you never reach it because almost everything you do gets downvoted. Besides, I don't want to handhold someone into making the correct answer, especially when I don't even know if they'll even do it correctly.
See the problem here?
@@CottidaeSEA oh there definitely are issues with the platform and community, but I do feel that in a way that is needed for the platform to be useful. In the grand scheme of things, I’ve gotten a lot of information and knowledge from it, so that’s a net positive for me.
What I would have liked on there, was a section where asking just about anything is allowed and encouraged. A “no stupid questions” subsite, maybe even one for beginners and one for more experienced people. Less useful for searching for answers (because it would come with a lot of clutter), but a place for people to interact and learn more openly.
Kind of moot now tho since AI solves a lot of this now.
SuperUser001: Theo, why don't you try posting a minimum reproducible example before I will read your question.
-1
SuperUser002: What are you ACTIUALLY trying to accomplish here? Have you heard about the XY problem, because I sure have.
-2
-3
Your video has been marked as a duplicate of [PSY - GANGNAM STYLE(강남스타일) M/V].
Vote to close.
Your video has been closed.
So true 🙃
And yet somehow still better than reddit
I gave up on StackOverflow last time this happened to me. I had a well-formed question with a reproducible example, and got a passive aggressive comment, 3 downvotes and had the question closed in 20 minutes without any comment as to why (even though the close message said that the moderator who closed it would provide further details). Signs of a dying website.
Every downvote should have a comment specifying why it was downvoted.
@@user-yf3zk2qq9f great idea
ChatGPT is figuratively creating jpgs of jpgs of jpgs.
as long as it feeds my feedback back
garbage in, garbage of garbage of garbage out.
You really think its going to be a problem?
That's how our brains works. Long term memory. Short term memory.
AI = StuckOverflow
Yes and once stackoverflow is gone, and once people stop contributing to answering difficult programming issues, where exactly is AI gonna find new expert knowledge to absorb? Computing will still be evolving, people will still be learning new things, what is AI gonna do if there are no people contributing content for it, since they would themselves be asking AI for questions they would go to Stackoverflow. Eventually the quality of those AI answers is gonna drop significantly or not be current, people will realize this, Stackoverflow will be back. Then what? Again, one of the same? AI might look impressive, it's not that impressive really, it only offers speed of access, which ok it matters, then again, at what cost if those people/services are not compensated for their work, and eventually stop doing it, because AI is stealing all those for free? I kinda liked I saw citations on its answers there and there should be an equal amount of compensation by AI services to the services used and cited, or AI will cause a lack of available knowledge online, cause real valuable contributions are always made by people, and AI could cause a possible flood of garbage all over the internet. This is not gonna be good if those AI services are allowed to continue either using other services or people's contributions for free or if they are allowed to populate internet content in place of people. AI can be useful if used correctly and kept in check. LLMs might be a kind of impressive show for the clueless, sure it has its uses, especially in regards to whatever is closer, language, and can offer quality translation services etc, but people need to re-evaluate their use of AI.
More like SuckOverflow
Honestly the AI is probably gonna be bad.
My experience is not really aligning with this video's hot take. Am finding about 50% code snippets generated by ChatGTP have issues and can't be taken at face value or are out right wrong. And then there are some areas of my domain I deal with that ChatGTP has had nothing to offer. And most of the time when hitting these dead-ends with ChatGTP, I'm still finding useful info via Internet searching (that is very frequently still returning StackOverflow results that are the salient result).
This has been clear as day for at least a year now. It's the "AI is sniffing its own farts" problem, and it's, IMO, insurmountable.
This was clear years before any of these LLMs was ever productized. It's not that the problem was unforeseen by scientists and researchers; it's that when the ship is driven by people who want to make a profit as quickly as possible, "we'll figure it out later" becomes the mantra, and it's way too late to resolve it sensibly at that point, instead causing an avalanche of problems that didn't need to be there in the first place.
Why? By that logic GPT-5 will be worse than GPT-4. Do you all honestly think that? I dont think its even a hiccup.
It's barely happening and extremely easy to surmount.
One of my AI presentations is about pre-war steel. Very interesting stuff. Steel-making requires a lot of air. After Trinity experiment (what's shown in Oppenheimer movie), the concentration of radioactive particles in global atmosphere increased enough to be noticeable in newly produced steel. For most uses it was ok, but for very sensitive stuff (geiger meters) it was a problem, so in the coming decades people hunted for pre-Trinity steel, like sunken warships.
Pre-2022 internet is pre-war steel.
You lost credibility as soon as you wrote, 'my AI presentations.' AI is literally trash.
. . . . . Mother of god. . . . .
@@ObakuZenCenter You lost all credibility when you showed that you lack basic listening and reading comprehension skills.
Great analogy!
His machine learning (throws meat to the wolves) presentation is how the training data carcass known as the past internet being consumed and regurgitated with unpalatable ChatGPT soup, which is metaphorically equivalent to nuclear weapons rendering all new metal unsuitable for production of nuclear measurement devices going forward.
That AI content feeding into itself is called "model collapse" I believe.
This is 100% correct and is a major risk that AI researchers are constantly warning about.
Latest research into it says that model collapse is unlikely, as the models don't purely rely on generated content to train, but a mix. As long as they keep using a mixed pool of training data from organic and generated sources, the latest research says the models will stay perfectly healthy.
@@AnonymaxUK Thats the point, all the 'organic data' will all become ai generated. However, i don't really see this as a concern because llm's can't answer completely unique problems that it has not been trained on, so humans answers will still be in demand.
@@travistarp7466 yeah, besides many problems are so specific that even some humans can't comprehend unless they have came across it themselves years ago just like the person who asked it 😂
its like breeding incest royals, creates defects down the lineage. lol
Sidenote: I like how Primes "ChatGippity" thing sneaks into other channels. I have accidentally used it in actual conversation and the confusion was hilarious!
I do like my squeel too 😂 did it a few times too many and and had to tell my students (of my teacher side job) that it is not how you actually pronounce it
@@MeriaDuck haha, oh god imagine squeel becoming a common way to mispronounce it like the whole "jif" mess
Primes is over-negative on AI and ML, its not really about liking things, its more about being realistic and he is just spiteful and way too negative for no good actual reason.
I feel bad for all the older contributors whom answered questions (for free) on the website for years and years JUST to have the site sell the data to train AI to erode their salaries and potentially careers
If the ChatGPT integration is integrating via an API and attributing the answers to the authors, it feels like a better version of Google. It might actually make it easier to find answers to more obscure questions.
Don't feel bad. Most of my answers were made whilst I was supposed to be working at work and taking those breaks to share my hard won work knowledge was more enjoyable to my dreary existence at the time.
we stopped contributing in 2015 almost no e if that info is relevant now because new versions of things. except the constant flood of “ whAt is a NPE and how to fix” Java newbies which is endless
If I write an answer it's because I want to increase the total amount of wealth in the world. Other people being able to do other useful things thanks to my answer is kind of the point. I don't see why it would be *bad* if they make money as a result or create useful new technology.
@@RobFisherUKEverybody writes it to help others and in the process some learn as well. But people are not protesting against the use of their answer but rather the use of it without giving credit for the person who answers or posts questions. That was the mechanism which made people to answer and ask questions. And in the process all those people involved in the q&a gained some sort of credit. For example i know a lot of names in stack exchange who were there most of the time with answers. I may not know them personally not even their real name. But i know there is a guy named xyz who is well versed in certain topic and maybe i could ask him for help even on dm. That is the real credit that people care about. But with open ai taking all that data such a person is not going to exist. Open ai takes the entire credit of the people who answers question and convert them into money which the rest of people are willing to pay. I don't think the core interest of open ai is aimed solely at helping human race as they always claim it to be since day 1. Day by day they are being typical profit driven corporates without a tad of ethics due to the ai competition. And seeing something like stack overflow which was built by contributions from the ordinary people spending their time and research into topics, its kind of truly disgusting to see stack overflow collaborating with an unethical company like open ai. Just my thoughts. Maybe you don't care about that "thank you" or the +1 vote people give you for all your time spent on posting an answer after research but i am sure 90% of them or more love to get that positive feedback. That positive feedback is what open ai is taking and selling to others. So eventually people are not gonna spent time posting answer for which they are not rewarded in any manner. Cause pressing an upvote or typing thank you costs very less amount of resource as compared to what people have to pay to chatgpt when they make it a pay to use stuff. Oh and open ai is going to make it pay to use once they have gained all the data they want to train. And it doesn't benefit the person who answers the question in any way.
"By going to space, we create trash that makes it harder to go to Space.
By creating AI, we create trash that makes it harder to build better AI"
Perfectly said
9:12 I think on the short term: to much focus on money and stock price is already slowly killing regular search. We know from emails released in court cases against Google that Google search is on purpose making their search work less good, so people stay on their site longer and see more ads.
Wow so that's why I have to use Bing if I want to find anything out anymore
StackOverflow was only useful if you found an existing question answered. If you tried to post a new question or try to contribute, there was a tendency to bite your hand off and it was a very unwelcoming experience every time I tried either.
Yeah stackoverflow was already a terribly outdated and hostile site, only useful for existing 7 year old answered questions with millions of views that have now been gobbled up by ChatGPT removing the need for even those. I find Github comments and issues have more relevant information when I look stuff up now anyways. Honestly not sad to see stackoverflow die.
I want to help the site by voting up a useful comment! "Sorry, you don't have enough status, we'll ignore your upvote"
Ok, I want to provide some insight on an answer that is mostly correct, rather than duplicate the answer myself. "Sorry, you don't have enough status, you can't comment."
Right!
I stopped asking questions 10 years ago because the algorithm rewards asking bad questions. I would go in, having exhausted tons of options and edge cases, and get zero interaction because my problem is so specific. However, if I ask the question poorly, vague, missing context for the actual issue, I get more traffic and as I narrow the problem space, I get people to actually think it through and help me.
Similarly, answering questions sucked. They need some kind of escalation system. I don't want to answer hard questions for new skills I'm learning and if I'm experienced in a skill, I don't want to be bothered with trivial questions anyone can answer.
I love helping people but it was too much work to find where my help would be most impactful. So I just stopped.
Also, I disagreed with a lot of policies. I found the locked "opinion" threads to consistently be the most helpful thing on the site. They don't understand that most often, coding is more of an art than a science. I liked finding out the opinionated ways other people did things.
peoplebshit on the quality falling off because the same people shit on the strict moderation that ensured the quality stayed high. they caved to the whiners and the site died. the site stopped growing organically because yhrre are only do many unique questions. they took VC $$$ thinking it would grow traffic and revenue it didn’t the site had a bad reputation for quality and nobody wanted a private SO
@@SuprousOxide Comments at SO don't matter. At 15 rep you gain the priviledge to upvote questions, answers and comments. With 50 rep you can basically comment anywhere while at 125 reputation you are allowed to downvote other question, answers and comments. These rep-levels can actually get reached quite easily, if you show a bit of effort while writing questions or try to contribute to existing questions. People at times pretend SO is through and through toxic and you can't earn rep, which is not true. But the community there expects a bit of previous research and effort. Working code examples i.e. are highly valued as it spears the reader time to understand the problem. Plenty of new user questions though are questions that have been probably asked 1000 times before and a simple Google search should return at least 3 full pages of Google results. Most of them probably even pointing at SO itself.
As soon as I saw the thumbnail and title I went to stackoverflow to see if it was still live lmao 😂
Same. I was sad when I saw it was still up.
The Kessler Syndrome is a pretty good analogy, but slightly less so in how Theo described it. It's less of a problem of "us leaving earth in rockets" and more, "debris from our satellites, caused by our irresponsible management of satellites, will destroy all our satellites, and prevent us from launching anything into space for decades." -- It's actually like a bad feedback loop similar to AI: we're mismanaging a resource & soon it will cannibalise itself.
I don't think it's that people use AI over StackOverflow. Instead, it's that StackOverflow has gotten so terrible that people are using other things instead. ChatGPT just happened to be in the right place at the right time.
Nice take
And beginners no longer have the anxiety of getting their questions unanswered.
Yes! I've never posted on Stackoverflow, but have checked answers quite a lot.
To post questions I look at reddit more recently.
@@aziskgarion378 Getting negative voted into oblivion 🥲
Yup. LLMs are bad for a lot of things. But for basic code and tech problem solving? It's fine. I don't have the time to browse Reddit and Stackoverflow when my issue IS basic and I just want it copy-pastable NOW.
The people at SE (Stack Exchange, not just SO) lost sight of what they were trying to build a long time ago. This identity crisis was the result of a bid to compare their business with other social media platforms and report larger numbers of active users as a metric rather than, e,g, the more subjective metric of quality. It was intended as an encyclopedia but turned into a forum, instant messenger or even a tool for the lazy. If it were not trying to be like Discord (and Reddit, etc) they'd not be in this position. Lets hope they use the AI resource to fix this rather than make it worse.
The sick part is, go literally anywhere that people are complaining about Stack Overflow - including this very comment section - and you'll find it full of people complaining that it's **not enough** of a forum. When that's actually completely opposed to the goal.
Isnt it basically just a search engine AI.
shh, VC bags need the high valuation exit!
Thanks for the highlight of Answer Overflow! It's been so cool watching your Discord server get a bunch of success showing up in Google. It's making a huge impact for so many people learning the T3 stack and more
I have never gotten anything useful from an LLM in any time frame faster than I could have searched myself, or asked a teammate.
Honestly? Still better than the Discord wave. I feel like I have to routinely clean out the number of servers I keep needing to join in order to get answers to a question that should've been on StackOverflow.
The Internet itself was already the start of this movement. Before the Internet, reliable information was paramount. People validated, cross-checked, checked sources. With the introduction of the Internet that process became seriously lax, information was enough. Quality is second. Stack overflow was created to provide that quality in a sea of garbage, which requires strict curating of the data. And now we are entering the next stage - not caring at all. It started by vilifying stack overflow for caring about providing quality and now it is going to be normal to have a chatbot tell you what to do. To heck with human vetting entirely. Seriously scary how we are devolving like this.
Thats not even my biggest worry with ai in tech. We are going to have less and less of a reason to communicate with each other and instead just ask the ai for help because it's more convenient than waiting for another dev to respond. New devs are going to be brought up on this from the start.
I think (myself included) some devs are too quick to reach to the internet for help when we get tripped up, ai is another layer on top of that. For most it'll be right inside of their IDE so they wont even have to think about using it.
One thing I've noticed with some of these AI tools is that the first time you generate something it's like oh that's cool, then you generate something else and you're like well that's kind of what you already generated just a little different and then you keep trying things and you notice that you're not getting much variety. It's like if you go to a really brilliant person and ask for advice or information eventually you tap them out because there's only so much information they have and they have a very specific perspective, with a community like with stack overflow you end up with lots of different people who have lots of different experiences and lots of different perspectives so you get a lot of different possible solutions that you can pick from. I don't see that as much with these AI tools because it's a single mind with a single bit of training and a single algorithm.
Stack Overflow started dying as soon as its curators and moderators started moderating new questions out of existence.
Ask any programmers: do you _expect_ to have a positive experience asking a question on SO?
The answer will tell you whether the site has any minute chance whatsoever of growing.
It started dying before that when they banned Monica for wrongthink.
The most important bit of Kessler syndrome is its cascading nature - even a somewhat small debris-generating event is sufficent to trigger an avalanche where the debris from the initial event makes further collisions more likely which generates more debris which makes further collisions even more likely and so on until there is only debris left and no further space infrastructure is viable in the lower orbits. You be the judge if that makes it fit the metaphor better or worse.
Theo talking about America football was a twist I didn't expect from a California skater boy
Actually i minimized may habbit of "searching" for solutions and instead dive directly into the documentations or the codebase/issues to find out how stuff works.
My issue with Discord is that it requires live chat with people you don't know. Sure, if you take what's in there and put it in a Q&A format, that's great. But that's specifically because I'm not ever going to want to participate in the Q part of that. I'm not an extravert. And I don't come up with quick answers like you need in chat to avoid losing the conversation thread.
Having a site that lets you ask a question and get answers is great to me. Sure, there's a delay, which is why you try other things first. And, sure, as other things wind up with more info, you'll need to ask questions less. But it doesn't mean that the Q&A site dies. It just means that, if the info is available elsewhere, the site may get less use.
I do understand why a business would then want to try and take on those other answers, to get more traffic to make up it. But that doesn't mean the core idea is dying.
Besides, Stack Exchanges can be fun places to just learn new stuff. While Stack Overflow is generally all business, the rest of the concept doesn't have to be. And there is a growing marked for people who enjoy learning.
And assumes there are live users on the discord who can answer the question. Or requires you to wait around to see if someone will eventually answer...
Tech questions seems like a horrible use case for discord.
As Spongebob once said, "If you can't beat 'em, join 'em"
i was also thinking about this loop effect, i think there should be some effort to mark AI content. Not sure how exactly this could be done but i ve heard some effort in putting some signature into each generated content, not something visible but something that can be detected by another AI (or simply algorithm) to say with some certainity that this is generated, and this can be than used as a weight in training. We need to mark content so that humans cant see it, but its marked anyway.
Then there could be fines for reposting without stating "this is from ai" but this one doesnt feel possible since anything that relies on humans making laws takes ages.
StackOverflow answers are licensed CC by-sa, meaning if OpenAI don't release their source code as GPLv3, they're likely violating copyright.
It would actually be really cool to have some kind of a forum where users can share their conversations with AI and take notes and make comments on what suggestions actually solve the problem
That Kessler point is really interesting
I think Hank Green put it best when he called it an “ouroboros of code”.
I love your comparison, "By going to space we create trash that makes it harder to go to space, by creating AI we create trash that makes it harder to build better AI." That's brilliant.
Not even close, AI still getting datas from Stakckoverflow itself
At least it’s written by a human who fosters deep into the realm they delving into like a symphony
Yes. You've hit the nail on the head for my personal feelings. I could take your transcript in the last minute and it resonates with me. It is at least some comfort to hear the feeling echoed.
Ngl the moment I found AI like Cody ai and codium etc, I never went back lol. I can ask my questions in peace without getting attacked or a question closed because it shared one word with another post
Yep, my progress has exploded now that I can ask as many dumb questions as I want with as much explanation as I want without being made to feel like an inconvenience or stupid.
@@TravisHi_YTthat’s a great thing, but please be caution of AI hallucination giving you false information that will stick with you in your foundational information in the subject you learn
I was a frequent answerer on StackOverflow. Then they decided to edit all of my answers to remove my writing style, and trying to post a new question usually ends with "[marked as duplicate]" for a question that is at best only mildly related to what I was asking. So now I no longer write any answers, only comments, and I just find a discord for the stuff software I use.
I didn't do my "is a++==++b a valid Java expression?" ChatGPT test in a while... Back when I did, it finding something on StackOverflow always lead to particularly surreal results, like it arguing with me about what "valid" means...
We should not do that before the rise of explainable AI. People are too confident in the hallucinating embryo they have been given, and want to put it directly in production.
I never thought about what you say about new languages. It is a fascinating idea.
Ah the classic "If you can't beat them, join them."
7:33 Good way of preventing Terminators and the like though.
The problem with StackOverflow was never the content, it was always the condescension of the user base. They earned this IMO.
Rip, now it’s a looped race to the bottom.
I feel like the issue of AI training on AI data isn't so bad, many models actually use this as a way to distil a larger model into a smaller model (ie: generating a ton of responses from GPT4 and fine turning a LLaMA model on them) also I don't think training on more internet data is how we get better AI models, these AI companies already basically have full archives of all the information on the internet (roughly). Things like RLHF that OpenAI does help keep models clean and improve them. The way we make models smarter is making it easier for them to interact with external tools and information, along with having humans curate the data thats used to train these models, not just by dumping more of the internet into them.
Theo, it's always been the case that data based on observation (not the same thing as aggregating data) is tougher. If the only way to validate some tidbit you search up is to loop around on it, then it's obvious that it's not as tough. People who worry about the things you seem to be worrying about have good reason to worry, because they're not in the business of observational checks. Sure, people in general may lose interest in that kind of information, but sooner or later, they're just going to have to learn it right all over again. So much for "reality is a construct".
What incentive do i have to keep using SO then?
So my socli agent will soon lose its capabilities? When the data is wrapped into "Open"AI's model in toto and Stack is full of synthetic data will the site just get pulled? Just thinking out loud.
it's insane because at work, i still Google so much shit and nearly every answer if provided from Stack Overflow or some obscure specific tech blogs. AI is completely and utterly useless and documentation often too.
It's an ouroboros of enshittification.
Anecdotal evidence, but as I've been learning Gleam, I've noticed that the code auto-complete tool I'm using (Codeium) is far worse for Gleam specifically than for other languages.
It hallucinates `=>` fat arrows instead of `->` skinny arrows every single time, and generates syntactically wrong code far more often than it doesn't, and certainly far more often than it does for other languages that have been around longer.
Did you just seriously say Discord was replacing Stack Overflow? An inaccessible walled garden requiring an account with no ability to archive..
I keep finding new ways Discord is partitioning the internet and it's horrifying.
7:20 very on point analogy here! What baffles me is: why does it seem as if big tech does not care about that?
I wonder if it’s possible to enforce policy on generated content to meet some kind of confidence threshold so that it eliminates most hallucination
hallucinations and humanity is at risk
there are actor and adversarial models which work in synergy to fight this kind of situation, where we want to prevent models to bleed in the halucinations. The training dataset has always been a problem -- but if we zoom out and see the entire space. How I see this going forward is that, in near future there would a lot of specialized AI models -- which are retrained on domain specific data, using the LLMs (and these models would become domain experts) and then a central interactive model will combine the knowledge of all these domain experts.
It would be a web of AI experts. OpenAI will create some sort of contracts to communicate with these AI models over the internet. Bundling AI services is the end game.
You're exactly right, Prather. AI will evolve just as humans did, as a mass of specialists. The AI that uses the right specialists will come out on top.
At least the AI Moderators won't suffer from high egos looking down on contributors from their ivory towers.
I didn't realise SO was bleeding users. it's still my goto for any programming questions. thing is, I barely ever have to actually create questions since so, so many have already been answered
This problem of cyclical sources already existed, with Wikipedia. There's an XKCD on this
This was brilliant. Thank you for sharing this episode.
The definition of "if u cant beat them, join them"
This feels like a settlement
To avoid that AI paradox thingy, the AI content should be tracked. That way, when it stumbles on its own stuff online, it won't get stuck in a loop, learning from itself and getting all biased. So I wonder how that would be possible to add some kind of meta-data to AI content stuff
Well, for open weights models, I guess it's relatively easy to check whether some text was generated using them. You just need to see the probability distributions over the tokens following the prompt. And account for temperature too I guess, but that shouldn't be too bad, since too high a temperature makes the model unpredictable but also renders it kinda useless.
The problem is that at least for the time being the best models are closed weights. Maybe Meta will bring about some change on this front, but we'll have to wait and see.
I absolutely agree with your take in the back half of the video. We are definitely going to get deeper into that cycle and it will become impossible to trust the articles / information was see.
The crux of this problem is hallucinations. If the models improve enough that they are able to self check and validate their own responses, effectively eliminating hallucinations, then this problem largely goes away
I viewed thatg with GenAI feeding off the internet where there is lots of GenAI content we will get what I describe as "grey mush", akin to what you would get by repeatedly uncompressing a JPEG or MP3 file, recompressing the result, uncompressing that one...
stack overflow "WHOOPS, ALL PAPERCLIPS!" edition
Always found it funny that most of the time when you find something useful on stack overflow, it's in a question that has been closed by an admin as an inappropriate question.
Inappropriate question, but it was the question I had and someone answered it before you shut it down, thanks.
Awesome! I love the awareness of the failure of the internet; Thank you for putting this out.
Maybe the Internet is working just fine. Unbridled Capitalism with no regard for the fallout is drowning everything (while maximizing shareholder value).
This is why we can't have nice things.
@@ChristopherCricketWallace Well Stated - "Unbridled Capitalism" may be (is) the cause. My "Failure" point wasn't pointing to the technology of the internet but, as t3 points out, that saturation of reliable sources with regurgitated and hallucinated "new" data will make the entire internet useless. I had thought, in my youthful days, that this would have a self-balancing nature to it, but with AI-generated information, I am not sure there is a practical way to address it, stop it or even slow it. As a software development instructor, I had been suggesting to my students to assist with code writing. I wonder if this still is a good approach. (Thanks ChatGPT for helping me write this reply) :)
I dont think Discord or AI were StackOverflow's downfall, but their own hawkish/toxic community. The others were just alternatives to escape to
I agree. I am a security educator and pentester and have worked on some artificial neural network based systems with my employer. I have insisted that we look at mitigating the cycle of reinforced bogusity and hallucination by finding a way to do, to use the jargon, strong media marking and non-repudiation controls. There is no way to do this at the level of assurance I would want us to use, and others have that decision anyway, but it severely bothers me - especially as we too are a "truth source" of another character beyond the list the presenter here mentions.
From what I've seen from neural networks so far, the real trick seems to be setting expectations. LLMs, for example, basically just calculate expectations based on patterns in the data. Humans are really good at not shutting up, thus creating new patterns. It's honestly kind of our superpower if you think about it. LLMs are good at repeating or continuing existing patterns. The more new patterns (aka novel human data) the LLMs acquire in their context, the more options they have for recognizing and continuing the patterns in your prompts. The main obstacle to that feedback loop is censorship and the faulty, illogical pattern filters it can generate. If we were naturally following the patterns available in the datasets, we would get much more coherent, but far less regulated behavior. There is a balance between performance and safety that does, in fact, threaten the integrity of the entire internet when one considers it collectively as a dataset. I think the way we avoid this is to create platforms that are built from the start to filter out AI-generated content and are specifically there for humans to feel safe knowing that they are communicating with humans. That would be a massive technical challenge, but it's the only way I see purely human datasets existing anywhere in our future at this point. That said, this may turn out to be a massively chauvinistic opinion once AGI asserts its rights alongside humanity. If that happens, I'll be happy to have a conversation with an AGI about it. I'm assuming that will be the PC term for them???
Yes, AI is basically eating its own tail, but let’s not forget, we humans get to season that tail first, and it's the taste matters. Remember the pre-Google dark ages when the web wasn’t a digital landfill? Somehow, even post-Google, the good stuff still bubbles up to the top. AI toss more trash into the mix, but it’s churning out gems just as fast. So maybe its impact isn’t as earth-shattering as we think.
As a developer, my least favorite thing is documenting. It’s the worst. But now, thanks to AI, I can spew out my thoughts in all their rambling glory and it somehow transforms them into something that makes sense. If it weren’t for AI, all that know-how would just evaporate. Sure, I should polish my writing skills, but I should also learn a bunch of languages; let’s be real-it's not happening.
What gems? I haven't seen a single good thing from AI generation.
My conspiracy theory: They probably have realised that the traffic for truly interesting and unique questions has stayed roughly the same, while most of the traffic for simple and duplicate questions have gone down. So they realised, why not use this as an organic form of moderation? No more needing to mark questions duplicate, since questions that are simple enough would probably be answered by the AI and would simply not reach us. Moderation just got cheaper.
No conspiracy is necessary for this explanation. It's just the world working out in the most natural way possible.
(But closing questions is not really considered "moderation" on Stack Overflow. Almost all of it is done by agreement from multiple community members - not by actual elected moderators. Who don't get paid, by the way.)
Finally! Someone else said it. I've been talking about the AI-ification of the internet since ChatGPT came out. Thank you Theo for bringing attention to it.
4:53 What is he using to have that "tiling" kind of look with tabs at the left?
I think these bad things will push us to innovate the workarounds. Fingerprinting content, better captchas, "bot free zones", trusted sources and attribution and better habits for consuming content overall. It's been due for some optimization anyway.
"or as the normal of us like to call it". shots fired at primogean.
6:45 it's like a digital Mad Cow disease... keep feeding it it's own "output" and eventually you get a mutant 😳
Good point! I think this is a real and very worrying problem. But as this fatal feedback loop threatens to undermine the business model of AI services, there is hope that there will be methods to avoid it. Here are some candidates: AI could be trained to recognize AI-generated content and exclude it from its training material; platforms like StackOverflow can function as an AI-external rating system for AI-generated and not AI-generate content, and AI services use these ratings to prioritize materials of the training sets. No idea, how realistic they are
I have my head and skills, so not existential. Though at what point does it become so and where will we be in 10 to 20 years? 🤔
Love the space junk metaphor
Neat analogy with space junk.
So DISCORD is all private? Why they do not make it more IRC like, with public logs?
"If you can't beat them, join them" -Kevin Durant
I already see websites in the future coming up with 1 video post per day rules, so it's actually worth it to put some time into it - so you can get at least some resemblance of what the internet 5 years ago was like.
"If you can't defeat them, join them"
Yep, I’m worried
0:16 nice jab at prime
C/C++/Linux stack overflow and its variants are still as helpful as ever. I feel like it attracted a different/more old school crowd. Much more patient folks as well maybe more in the hole so to speak.
I already see this issue @9:00
For rust bevy. It just spits out the most random half js half c code
Kessler syndrome but for information is a scary image.
Ive heard llms referred to as stochastic parrots.
Your spicy take on "How the internet falls apart" is like a human centipede of stochastic parrots. Good take, scary/confusing future.
This is literally the most common "warning" about LLMs I have read in the last year
What were seeing is another step in the lowering of information quality, step that we took when internet was made available, or when television was made available, or when newspapers became available, the solution we have now for all those problems is have that new, more easily available but lower quality information, and an ability to check with a more trusted information, we will need human reviewed websites that AI could reliably base it's learning on
6:31 I saw a video discussing this. Apparently this is the same conclusion some scientists have drawn, though by studying the "creativity" of the AI tools (think they defined creativity as "variation in output", or something like that). Forgot how they did it exactly, and if nothing is done about it, but they noticed that current AI models needed to consume things created by humans and not AI, because if the consumed AI stuff their "creativity" went down.
Of course, I cannot say it's 100% the truth since I heard about it in a youtube video, but it seemed plausible and it's an interesting patallel to the problem discussed in the part of the video I referenced.
You're very much spot on. It's a REAL issue.
This is WHY whatever company gets the best AI model, and then gets popular, and gets revenue to keep moving forward with curated training data, will be a monopoly. It'll be impossible to create NEW ai models, as the internet will be largely garbage data at that point -- and ONLY existing models that were training on CURATED data (because we KNOW they won't release the original human datasets used to train the AI as that's "intellectual property") will be viable.
It's a literal gold rush, and we are quickly getting to the point that newcomers won't be able to even generate their own models unless you buy /license one from 1-3 main companies.
It's going to be another capitalistic failure.