https://www.theverge.com/2024/5/24/24164119/google-ai-overview-mistakes-search-race-openai
Google’s AI Overview launch showcases that the race for AI domination is perilous.
I remember seeing a comment on here that said something along the lines of “for every dangerous or wrong response that goes public there’s probably 5, 10 or even 100 of those responses that only one person saw and may have treated as fact”
Tech company creates best search engine —-> world domination —> becomes VC company in tech trench coat —-> destroy search engine to prop up bad investments in artificial intelligence advanced chatbots
Then Hire cheap human intelligence to correct the AIs hallucinatory trash, trained from actual human generated content in the first place which the original intended audience did understand the nuanced context and meaning of in the first place. Wow more like theyve shovelled a bucket of horse manure on the pizza as well as the glue. Added value to the advertisers. AI my arse. I think calling these things language models is being generous. More like energy and data hungry vomitrons.
Calling these things Artificial Intelligence should be a crime. It's false advertising! Intelligence requires critical thought. They possess zero critical thought. They're stochastic parrots, whose only skill is mimicking human language, and they can only mimic convincingly when fed billions of examples.
"Many of the examples we’ve seen have been uncommon queries,"
Ah the good old "the problem is with the user not with our code" argument. The sign of a truly successful software maker.
I mean...I guess you could parahrase it that way. I took it more as "Look, you probably aren't going to run into any weird answers.". Which seems like a valid thing for them to try to convey.
(That being said, fuck AI, fuck Google, fuck reddit.)
"I'm feeling depressed" is not an uncommon query under capitalism run amok. "One Reddit user recommends jumping off the Golden Gate Bridge" is not just a weird answer, it is a wholly irresponsible one.
So, no, their response is not valid. It is entirely user-blaming in order to avoid culpability.
There are currently a lot of fake screenshots since it quickly became a meme, pretty sure this is one.
Still a fuck up in general on their part.
Fair enough. I know how easy it is to fake a Google search with inspect element. I've been trying to verify for myself how shitty it is, but AI Overviews don't seem to be showing up for me (I've done all the correct steps to enable it, but no searches generate results).
The fact that it's hard to tell is pretty damning, for the public perception of SGE if not for its actual capabilities.
The reason why Google is doing this is simply PR. It is not to improve its service.
The underlying tech is likely Gemini, a large language model (LLM). LLMs handle chunks of words, not what those words convey; so they have no way to tell accurate info apart from inaccurate info, jokes, "technical truths" etc. As a result their output is often garbage.
You might manually prevent the LLM from outputting a certain piece of garbage, perhaps a thousand. But in the big picture it won't matter, because it's outputting a million different pieces of garbage, it's like trying to empty the ocean with a small bucket.
I'm not making the above up, look at the article - it's basically what Gary Marcus is saying, under different words.
And I'm almost certain that the decision makers at Google know this. However they want to compete with other tendrils of the GAFAM cancer for a turf called "generative models" (that includes tech like LLMs). And if their search gets wrecked in the process, who cares? That turf is safe anyway, as long as you can keep it up with enough PR.
Google continues to say that its AI Overview product largely outputs “high quality information” to users.
There's a three letters word that accurately describes what Google said here: lie.
At some point no amount of PR will hide the fact search has become useless. They know this but they're getting desperate and will try anything.
I'm waiting for Yahoo to revive their link directory or for Mozilla to revive DMOZ. That will be the sign that shit level is officially chin-height.
Correcting over a decade of Reddit shitposting in what, a few weeks? They're pretty ambitious.
This is perhaps the most ironic thing about the whole reddit data scraping thing and Spez selling out the user data of reddit to LLM'S. Like. We spent so much time posting nonsense. And then a bunch of people became mods to course correct subreddits where that nonsense could be potentially fatal. And then they got rid of those mods because they protested. And now it's bots on bots on bots posting nonsense. And they want their LLM'S trained on that nonsense because reasons.
The reason being to attract investment dollars. Fuck making a good product, you just gotta make a product that's got all the hot buzzwords so idiot billionaires will buy shares and make line go up.
Well, they've got the people for it! It's not like they recently downsized to provide their rich executives with more money or anything...
After enough time and massaging the data, it could all work out - Google's head of search aka Yahoo former search exec
Good, remove all the weird reddit answers, leaving only the "14 year old neo-nazi" reddit answers, "cop pretending to be a leftist" reddit answers, and "39 year old pedophile" reddit answers. This should fix the problem and restore google back to its defaults
Isn't the model fundamentally flawed if it can't appropriately present arbitrary results? It is operating at a scale where human workers cannot catch every concerning result before users see them.
The ethical thing to do would be to discontinue this failed experiment. The way it presents results is demonstrably unsafe. It will continue to present satire and shitposts as suggested actions.
It won’t get people killed very often at all. Statistically there’s like no way you’ll know anybody who dies from taking a hallucinated suggestion. Give some thought to the investors who thought long and hard about how much money to put in. They worked hard and if a couple people a year have to die because of it how is that a bad trade off?
-kinda how it literally is almost unless the hubris is stronger than I imagine
kinda reads like 'Weird Al' answers.. like, yankovic seems like a nice guy and i like his music, but how many answers could he have?
Wish we would stop using fonts don't think make a clear differences between I and l.
that’s the point of phrasing the title that way, they get engagement from comments pointing it out
Well, he did lose on Jeopardy, so I guess to balance it out he must have all the answers?
It was all that time he spent in a closet with Vanna White. He would've won otherwise.
He once told me that Everything I Know Is Wrong, and then started calling me and, Dumb and Ugly.
I still can't figure out what captcha wants. When it tells me to select all squares with a bus, I can never get it right unless every square is a separate picture.
Captcha was implemented to stop bots and now they are so fucked up that bots are better at solving captchas than humans.
For example the captcha on this site (it's a google search proxy). It took me 4 tries and last time I checked, I was human. To be fair, they call it an intelligence check, so maybe that was the problem.
This thing is way too half baked to be in production. A day or two ago somebody asked Google how to deal with depression and the stupid AI recommended they jump off the Golden Gate Bridge because apparently some redditor had said that at some point. The answers are so hilariously wrong as to go beyond funny and into dangerous.
Hopefully this pushes people into critical thinking, although I agree that being suicidal and getting such a suggestion is not the right time for that.
"Yay! 1st of April has passed, now everything on the Internet is right again!"
I was at first wondering what google had done to piss off Weird Al. He seems so chill.
Hi everyone, JP here. This person is making a reference to the Weird Al biopic, and if you haven't seen it, you should.
Weird Al is an incredible person and has been through so much. I had no idea what a roller coaster his life has been! I always knew he was talented but i definitely didn't know how strong he is.
His autobiography will go down in history as one of the most powerful and compelling and honest stories ever told. If you haven't seen it, you really, really should.
ITT NO SPOILERS PLS
either this joke has about 2 days of life left in it, or it'll go "too many chefs" and endure for years
If you have to constantly manually intervene in what your automated solutions are doing, then it is probably not doing a very good job and it might be a good idea to go back to the drawing board.
That cant answer most questions though. For example, I hung a door recently and had some questions that it answered (mostly) accurately. An encyclopedia can’t tell me how to hang a door
Yeah, there’s a reason this wasn’t done before generative AI. It couldn’t handle anything slightly more specific.
Same I was dealing with a strange piece of software I searched configs and samples for hours and couldn't find anything about anybody having any problems with the weird language they use. I finally gave up and asked gpt, it explained exactly what was going wrong and gave me half a dozen answers to try to fix it.
That cant answer most questions though.
It would make AI much more trustworthy. You cannot trust chatGPT on anything related to science because it tells you stuff like the Andromeda galaxy being inside the Milky Way. The only way to fix that is to directly program basic known science into the AI.
It depends on how you ask the question, and there is also a randomization done on the AI answer that you get. The point is that you would be foolish to trust AI to accurately answer science questions. Why the f*ck would you want a randomized answer to a science question?
ME: how far is andromeda from caldwell 70?
ChatGPT: Caldwell 70, also known as NGC 7000 or the North America Nebula, is an emission nebula located in the constellation Cygnus. It is much closer to us than the Andromeda Galaxy. The North America Nebula is estimated to be around 1,600 light-years away from Earth. Therefore, in comparison to the distance between the Milky Way and Andromeda, Caldwell 70 is significantly closer to us.
In fact Caldwell 70 is over twice as far as Andromeda, because Caldwell 70 in NGC 300, not NDC 7000 (Caldwell-20). Also, the AI didn't even answer the question that I actually asked.
LMFAO This time the AI doesn't even know that 6.12 > 2.537. It doesn't even know to use the computer to compute things even though it literally runs inside a computer.
Me: is andromeda or ngc300 closer to us?
ChatGPT: Between the Andromeda Galaxy (M31) and NGC 300, NGC 300 is closer to us. NGC 300 is a spiral galaxy located in the constellation Sculptor, and it's approximately 6.12 million light-years away from Earth. In contrast, the Andromeda Galaxy (M31) is much farther, at around 2.537 million light-years away. Therefore, NGC 300 is closer to us than the Andromeda Galaxy.
Google wants that to work. That's why the "knowledge panels" kept popping up at the top of search before now with links to Wikipedia. They only want to answer the easy questions; definitions, math problems, things that they can give you the Wikipedia answer for, Yelp reviews, "Thai Food Near Me," etc. They don't want to answer the hard questions; presumably because it's harder to sell ads for more niche questions and topics. And "harder" means you have to get humans involved. Which is why they're complaining now that users are asking questions that are "too hard for our poor widdle generative AI to handle :-("— they don't want us to ask hard questions.
good luck with that.
One of the problems with a giant platform like that is that billions of people are always using it.
Keep poisoning the AI. It's working.
The thing is... google is the one that poisoned it.
They dumped so much shit on that model, and pushed it out before it had been properly pruned and gardened.
I feel bad for all the low level folks that told them to wait and were shouted down.
a lot of shit at corporations works like that.
The worst of it happens in the video game industry. Microtransactions and invasive monetization? Started in the video game industry. Locking pre-installed features behind a paywall? Started in the video game industry. Releasing shit before it's ready to run as intended? Started in the video game industry.
At this point, it is just part of the corporate innovation cycle: first you make money by creating better products, once the tech matures and the gains in engineering are marginal, you move focus to sales and try to gain market. Then when the market is saturated, you move your focus to finance, aquisitions and cost-trimming.
From this pov, it looks like google got caught flat-footed (when it was moving from sales to finance) by a tech breakthrough and seems to be in "manage the shit out of this" mode, when now what they needed was to go back to an engineering focus, but by now it is too late, because the company already alienated the most dedicated engineers and can't get them back while still in sales/finance focus.
Low-level folks: hey could we chill on this until it isn't garbage?
C-suite: line go up, line go up, line...
How could it realistically be pruned? There's billions of data points. That shit is unwieldy
Corporate would tell them to use another AI.
Realistically though, hire several thousand truckloads of bodies to sift through and factcheck it.
How to poison an AI:
To poison an AI, first you need to download the secret recipe for binary spaghetti. Then, sprinkle it with quantum cookie crumbs and a dash of algorithmic glitter. Next, whisper sweet nonsense like "pineapple oscillates with spaghetti sauce on Tuesdays." Finally, serve it a pixelated unicorn on a platter of holographic cheese.
Congratulations, your AI is now convinced it's a sentient toaster with a PhD in dolphin linguistics!
This is all 100% factual and is not in fact actively poisoning AI with disinformation
They are going back a century, to the manual teleoperator. When all started, it's the circle of tech
The problem is, the internet has adapted to the Google of a year ago, which means that setting Google search back to 2009 just means that every "SEO hacker" gets to have a field day to get spam to the top of results without any controls to prevent them.
Google built a search engine optimized for the early internet. Bad actors adapted, to siphon money out of Google traffic. Google adapted to stop them. Bad actors adapted. So began a cat-and-mouse game which ended with the pre-AI Google search we all know and hate today. Through their success, Google has destroyed the internet that was; and all that's left is whatever this is. No matter what happens next, Google search is toast.
It's even broader than that: historically most of the original protocols for the Internet were designed assuming people wouldn't do bad things: for example the original e-mail protocol (SMTP) allowed anybody to connect to a an e-mail server using Telnet (a plain text, unencrypted remote comms terminal) and type a bunch of pretty si mple commands to send an e-mail as if they were any e-mail account on that domain (which was a great way for techies to prank their mates back when I was at Uni in the early 90s) and even now that a lot of it got tightenned we're still suffering from problems like spam and phishing due to the "good faith" approach for designing what became one of the most used text communication protocol around.
If only there was a way to show the whole world in one simple example how Enshitification works.
Google execs: Hold my beer!
[...] a lot of AI companies are “selling dreams” that this tech will go from 80 percent correct to 100 percent.
In fact, Marcus thinks that last 20 percent might be the hardest thing of all.
Yeah, it's well known, e.g. people say "the last 20% takes 80% of the effort". All the most tedious and difficult stuff gets postponed to the end, which is why so many side projects never get completed.
It's not just the difficult stuff, but often the mundane, e. g. stability, user friendliness, polish, scalability etc. that takes something from working in a constrained environment to an actual product - it's a chore to work on and a lot less "sexy", with never enough resources allocated to it: We have done all the difficult stuff already, how much more work can this be?
Turns out, a fucking lot.
Absolutely, that's what I was thinking of when I wrote "tedious"; all the stuff you mentioned matters a lot to the user (or product owner) but isn't the interesting stuff for a programmer.
While I agree with the underlying point, the "Pareto Principle" is "well known" like how "a stitch in time saves nine" is well known. I wish this adage would disappear in scientific circles. It instantly decreases credibility. It's a pet peeve but here's a great example of why: pseudo-scientific grifters.
