Google walks back AI search results

M

Marees

2[H]4U
Joined
Sep 28, 2018
Messages
2,265
The tech companies are spending so much heat & energy on GPU driven LLM models but the tech is not ready for primetime yet

There are so many screenshots floating around but I have no way to confirm if those are fake or real.

I saw a screenshot where the AI summarized pineapple on pizza as the best topping !!!

https://edition.cnn.com/2024/05/24/tech/google-search-ai-results-incorrect-fix/index.html

Google’s search overviews are part of the company’s larger push to incorporate its Gemini AI technology across all of its products as it attempts to keep up in the AI arms race with rivals like OpenAI and Meta. But this week’s debacle shows the risk that adding AI – which has a tendency to confidently state false information – could undermine Google’s reputation as the trusted source to search for information online.

GOUwwNhbUAA5C-G.jpeg
 
Not only is it inaccurate, but even if it were,m this is a bad idea.

It is trained on content from 3rd party websites who are then robbed of clicks and resultant ad revenue.

If they drive those out of business, then they will no longer have any content to train their AI models on.

There needs to be an agreed upon method for compensating the sources of training data. I propose designing models that report what references they use in the answers they present, and in turn pay royalties to those sources every time the AI produces an answer that utilizes that reference.

Much how radio stations keep track of and pay royalties to rightsholders every time they play a song.
 
Zarathustra[H] said:
Much how radio stations keep track of and pay royalties to rightsholders every time they play a song.
Click to expand...
Small startup coming with something quite similar for the training data part, artist will be like for song rights participate and be compensated on training.

Has for the click, system like above show the person link for their source (same for binggpt) and it is yet to see if it will bring more traffic on net versus less, for stuff like stack overflow it will be obviously less but for other stuff it will be more.

https://web.archive.org/web/2023071...ch-chatbot-increasing-website-click-throughs/

Once the ai get better maybe it change and it could become a big challenge, a payment to the source is already started for the big (NYtimes, reddit, chora) but ideally it would have a way for evreyone putting something online to get a little something everytime bing-google-facebook-apple answer using them, with people paying to use those service.

Has for training data, now we think a lot of about text found on the internet, soon it will be robot experiencing the world, youtube-tiktok-streamer-camera feed videos, they will never have a situation with no new content to train, they will generate it, from the real and simulated world.
 
I have directly experienced really bad results from this crap just in the last week. Since you can't easily disable it, you encounter it at the top of most searches. Well, if you actually use search for anything serious... like trying to search info for heat treatment of certain metals, or a host of other machining questions pertaining to how to run certain tools or what chemicals to use... it will just flat spit out WRONG answers presenting them as "fact" in the way it grammatically presents it.

You could get yourself or someone else in a world of hurt if you just decided to take that top search result at its word.

And what really frosts me the most about it, is that it has been automatically enabled and forced on us and it really does just present it as if it is an authority on the answer it provides.

This is way past stupid it's flat out real world dangerous and it's on right now and spitting out pure garbage to untold millions of people.
 
Advil said:
You could get yourself or someone else in a world of hurt if you just decided to take that top search result at its word.
Click to expand...
Not in favor of people getting injured by AI, but I'm looking forward to the massive lawsuits when it happens.

The idea of billion-dollar liability lawsuit judgments is kind of insane, but nothing less than that will get companies like Google to back down, I don't think.
 
That's exactly what I'm thinking too. It shouldn't have ever come to this. We shouldn't have to risk our freedom to be mistaken over something like this but it's aggregating ALL the garbage on the internet without any context for a reader to know where the source came from or all the ways a person decides to trust a source or not. It BREAKS the way decisions on information should happen. It doesn't enhance it, it's literally the opposite.
 
"AI optimists argue that we should embrace the hype because of the rapid progress made so far, trusting that it will continue to improve. I really do believe that this technology will continue to get better, but focusing on an idealized future where these technologies are flawless ignores the significant issues they currently face — and allows companies to continue delivering subpar products."


https://www.theverge.com/2024/5/23/24162896/google-ai-overview-hallucinations-glue-in-pizza


One thing about Google’s AI answers is that there’s zero clarity on how the system chooses what to elevate. Should AI answers include material from random Reddit comments? An example: the bullet point about cats licking you to see if you’re fit for consumption was word for word taken from one Reddit post that was clearly a joke.
Click to expand...
445745862_823023152526019_8286744299977395863_n.jpg
 
LLMs in their regular state make little sense for answers widgets since there's no indication of provenance due to how they're trained; could be pulled out its ass (hallucinated), from a joke post or a legit source.

So it's not even an 'overview' of relevant results but just an LLM trained on global crawled text doing LLM things. Google are just in a scramble since Microsoft decided to throw caution to the wind first.
 

Google scrambles to manually remove weird AI answers in search​

The company confirmed it is ‘taking swift action’ to remove some of the AI tool’s bizarre responses.​


Google spokesperson Meghann Farnsworth said in an email to The Verge that the company is “taking swift action” to remove AI Overviews on certain queries “where appropriate under our content policies, and using these examples to develop broader improvements to our systems, some of which have already started to roll out.”


“[These models] are constitutionally incapable of doing sanity checking on their own work, and that’s what’s come to bite this industry in the behind” — Gary Marcus, an AI expert and an emeritus professor of neural science at New York University, told The Verge

https://www.theverge.com/2024/5/24/24164119/google-ai-overview-mistakes-search-race-openai
 
I mean, both of your examples aren't exactly wrong. Suicide will definitely end depression for you. Not sure about your loved ones but that's their problem, not yours. Glue sticking the cheese to pizza should be fine, so long as you get odorless, tasteless kind. I don't really agree with the non-toxic bit though. Toxic glue can give some really nice highs.
 
A speculative blog analysis from last year

What if Generative AI turned out to be a Dud?​

Some possible economic and geopolitical implications​


Microsoft, up for the year by nearly half, perhaps largely on the promise of generative AI, might see a stock slump;

NVIDIA skyrocketing even more, might also fall.

If hallucinations aren’t fixable, generative AI probably isn’t going to make a trillion dollars a year. And if it probably isn’t going to make a trillion dollars a year, it probably isn’t going to have the impact people seem to be expecting. And if it isn’t going to have that impact, maybe we should not be building our world around the premise that it is.


https://garymarcus.substack.com/p/what-if-generative-ai-turned-out
 
Twisted Kidney said:
When you're so excited about automated content theft that you don't even make your system stable.
Click to expand...
I think google is more of an example of being forced too and scared a lot (they are taking a lot of time, Google is by a good amount the most advanced AI in the world with probably the most data, would they have been excited about this it would have launched it years ago but it is a direct competition and disruption to their core business), they would have been fine with the previous statue quo, they had probably the biggest money printing business in the history of mankind.

Okatis said:
LLMs in their regular state make little sense for answers widgets since there's no indication of provenance due to how they're trained; could be pulled out its ass (hallucinated), from a joke post or a legit source.

So it's not even an 'overview' of relevant results but just an LLM trained on global crawled text doing LLM things. Google are just in a scramble since Microsoft decided to throw caution to the wind first.
Click to expand...
Many LLM will give you links for their source, binggpt and that google LLM being 2 of them (all the example here seem to have the link just below) those system can do websearch themselves, they are often wrong too.
 
Last edited:
LukeTbk said:
Many LLM will give you links for their source, binggpt and that google LLM being 2 of them (all the example here seem to have the link just below) those system can do websearch themselves, they are often wrong too.
Click to expand...
The thing is, in these Google examples even if it linked to/cited something it's not a summary of that source but rather a typical aggregate LLM response from everything it's been trained with (see: all the examples where users have found part of the text coming from some joke Reddit post, etc, which are certainly not from any page it may link to).

Ie: the provenance of the text displayed always differs here and is unknown. If instead it were made to indeed summarize a cited page using an LLM (which LLMs can do pretty well) then it'd give a more accurate overview. I have only limited experience with Bing's implementation so can't reasonably judge whether it's effectively doing this or some version of the former and just appending 'citations'.
 
Internally Google is scrambling.
BingGPT and other LLM’s cropping up are eating into their search market.
Googles search market fuels the data to their advertising engines, pair that with a rapid increase in Ad Blockers and tighter security around tracking cookies and advertising ID’s and Google is worried.
Google once famed for the accuracy of its consumer data is starting to loose the confidence of Advertisers themselves.
To compound this Google is facing regulatory backlash for how they “promote” search results and alter searches unannounced to the users making them. Changing queries like “Baby clothes” to a paid advertiser so internally it searches something like “GAP Baby clothes” without changing what is in the search field itself so the user doesn’t know it was altered.
Microsoft and Meta are stepping on Googles toes, and they are throwing figurative shit at the wall trying to catch up.
 
techie81 said:
View attachment 655780
Click to expand...

Apparently this has now been tested (& proved to work):

Google AI said to put glue in pizza so I made a pizza with glue and ate it​

it was kind of OK? I had only a few bites because I was afraid of poisoning myself.

Most importantly: Did the glue keep the cheese from sliding off? You bet it did:

Happy to report NO movement or sliding of the cheese
Click to expand...

https://www.businessinsider.in/tech...ith-glue-and-ate-it/articleshow/110406679.cms

Master-1.jpg
 

Attachments

  • Master.jpg
    Master.jpg
    333.4 KB · Views: 0

Why Google is (probably) stuck giving out AI answers that may or may not be right​


Is Google going to have to backtrack on this?

No, says Google, which argues that the dumb answers it has been generating are few and far between. And that most people don't know or care about search answers that tell people how many rocks to eat.


"In the past, Google was telling you somebody else could answer your question. Now Google is answering your question.
It's the difference between me handing you a map and me giving you directions that will send your car barreling over a cliff."


https://www.businessinsider.in/tech...or-may-not-be-right/articleshow/110407225.cms
 
When mistakes cropped up in Google’s very first demo of Bard in February 2023, shares of Alphabet dropped 7%, wiping $100 billion off the company’s value. On Friday, as more social posts of its latest gaffes went viral, they opened up almost 1%. Wall Street doesn’t seem to care. Does Google?


the more tech companies showcase how much generative AI doesn’t work, the harder it will be for them to prove its usefulness to enterprise customers and consumers alike.


https://www.moneycontrol.com/news/o...-hallucinating-does-anyone-care-12733070.html
 
Lakados said:
Microsoft and Meta are stepping on Googles toes, and they are throwing figurative shit at the wall trying to catch up.
Click to expand...
This feels like a bad mis-step from google

Strangely, they don't seem to care at all that they have trashed the USP of their cash cow
 
Okatis said:
The thing is, in these Google examples even if it linked to/cited something it's not a summary of that source but rather a typical aggregate LLM response from everything it's been trained with (see: all the examples where users have found part of the text coming from some joke Reddit post, etc, which are certainly not from any page it may link to).

Ie: the provenance of the text displayed always differs here and is unknown. If instead it were made to indeed summarize a cited page using an LLM (which LLMs can do pretty well) then it'd give a more accurate overview. I have only limited experience with Bing's implementation so can't reasonably judge whether it's effectively doing this or some version of the former and just appending 'citations'.
Click to expand...
LLM is useful for summarizing the contents of any one website

However to synthesize contents across multiple websites, there should be a weighting algorithm (like google used to have number of links metric for traditional search)

I don't have knowledge in this area, but I hope there is a solution for this mess
 
Okatis said:
Ie: the provenance of the text displayed always differs here and is unknown. If instead it were made to indeed summarize a cited page using an LLM (which LLMs can do pretty well) then it'd give a more accurate overview. I have only limited experience with Bing's implementation so can't reasonably judge whether it's effectively doing this or some version of the former and just appending 'citations'.
Click to expand...
It can link to reddit page yes (at least bingGPT) in the example above google link to some quora answer which is quite a direct reddit competitor.

One possible trick is asking a question about an recent event/discovery that must have happened after its training to see how good it can be a making a websearch and resuming it, say for example, if we ask bingGPT:
what happened in the last maverick game

Even for something that simple, it get a lot of things right (of all the team called maverick it assume correctly we talk about the Dallas vs Minnesota game of yesterday), who win, what was the score, that it was game 3, when-where it was, that Kyrie scored 33 points and Edwars 26, but get stuff wrong like that Doncic contributed 5 points (he had 5 assists so maybe that what it tried to say).

And if you ask it did Doncic really contributed 5 points, it will give you the good answer (33pts) after reflection, which show an relatively easy way a lot of hallucination will get away, once those system are 100 times faster and cheaper to run (if those blackwell numbers were true that could be soon enough, those systems made a 12x cheaper leap the last 12 month or so), in high quality mode it will be able to ask itself 10,000 times question of the: are you sure in a loop for every statement made in the first answer and stop only once the answer have been stable for a while.
 
Last edited:
Marees said:
the more tech companies showcase how much generative AI doesn’t work, the harder it will be for them to prove its usefulness to enterprise customers and consumers alike.
Click to expand...
Google just used generative AI :
https://blog.google/technology/ai/google-deepmind-isomorphic-alphafold-3-ai-model/
https://www.nature.com/articles/s41586-024-07487-w
https://www.pharmavoice.com/news/google-alphafold-3-drug-discovery-pharma-buzz/716496/
The “Nobel Prize-worthy invention” could be worth hundreds of billions commercially — and have a deep impact on drug R&D.

This could be a 100 billions pharma arm of Google. And those examples are generative AI coming up with stuff unknown to humans, for them interpolation and extrapolation and not really different and the idea they can only copy already learned data or interpolate between 2 known points seem to be false.

And that was a strong signal to investor, not many company in the world are sitting on something like this, possibly none and there a feeling google has 2-3 in their sleeve of deep long research projects that end up in new big commercial endeavor.

A tech startup just use generative AI to predict to make a new CRISPR gene editing tools
https://www.fiercebiotech.com/medte...s-and-crispr-open-source-gene-editing-project
https://www.profluent.bio/applications#open-crispr

It seem to be 2 world, one where a laptop can use generative AI to predict the meteo in 6 days using less than 1 minute of a regular laptop power better than a 9 figures super computer using a giant amount of power, achieved to generate accurate protein folding prediction in which and another world in which people make money creating content telling people generative AI doesn't work without ever having to address the last 4 years giant success of it in anyway, simply pointing to some image generation or hallucination in some search result has being deal breaker, meanwhile Tesla is now selling self-driving service on their car.
 
LukeTbk said:
Google just used generative AI :
https://blog.google/technology/ai/google-deepmind-isomorphic-alphafold-3-ai-model/
https://www.nature.com/articles/s41586-024-07487-w
https://www.pharmavoice.com/news/google-alphafold-3-drug-discovery-pharma-buzz/716496/
The “Nobel Prize-worthy invention” could be worth hundreds of billions commercially — and have a deep impact on drug R&D.

This could be a 100 billions pharma arm of Google. And those examples are generative AI coming up with stuff unknown to humans, for them interpolation and extrapolation and not really different and the idea they can only copy already learned data or interpolate between 2 known points seem to be false.

And that was a strong signal to investor, not many company in the world are sitting on something like this, possibly none and there a feeling google has 2-3 in their sleeve of deep long research projects that end up in new big commercial endeavor.

A tech startup just use generative AI to predict to make a new CRISPR gene editing tools
https://www.fiercebiotech.com/medte...s-and-crispr-open-source-gene-editing-project
https://www.profluent.bio/applications#open-crispr

It seem to be 2 world, one where a laptop can use generative AI to predict the meteo in 6 days using less than 1 minute of a regular laptop power better than a 9 figures super computer using a giant amount of power, achieved to generate accurate protein folding prediction in which and another world in which people make money creating content telling people generative AI doesn't work without ever having to address the last 4 years giant success of it in anyway, simply pointing to some image generation or hallucination in some search result has being deal breaker, meanwhile Tesla is now selling self-driving service on their car.
Click to expand...
Google has done stuff with AI such as deepmind/alpha zero that work in a closed environment with clear rules.

However I am deeply skeptical of this LLM stuff, if it requires additional (manual) step to verify the output
 
Marees said:
Google has done stuff with AI such as deepmind/alpha zero that work in a closed environment with clear rules.
Click to expand...
Chess and Go are closed with clear rules, how a protein will fold, how to make a gene editing protein or what the meteo will look like in 5 days at a gps location, I am not sure it counts has being a closed environment one of those is planet earth climate level of openness. They I imagine had a feedback verficiation step with the output that occurred days-week later of the prediction, unlike Chess-Go the system cannot determine by itself if it was right with premade rules to score it right away.

In both case it has clear objective with a really clear high quality data set too, that where it will shine first, but has compute get strong, soon something like basic coding could start to look like that (has it can compile and judge the result of what it created if those part get fast enough).

Marees said:
However I am deeply skeptical of this LLM stuff,
Click to expand...
Still completely unknown (apparently we are still in the phase giving it more computer deliver superb gain phase), but every step of the way it overdelivered of what we thought possible has of now and faster than we thought it would if it could.
 
Marees said:
What if, LLMs have reached the stage of diminishing returns ??
Click to expand...
yes that will always be a question for people outside the loop, but according to microsoft-openAI they currently still see nice improvement with larger compute at the moment:

apabilities-will-continue-to-grow-v0-eshd0gghfw1d1.png


GPT focus has been vaslty on speed and cost, in practical turn, how much better it does at something if you let it spend has much time and money, will start to be necessary in those comp, ability to understand mood from audio and image, for pure text, seem to have got much better at coding in my experience (specially vervsus the first gpt-4 release, this a bit on an issue with the frog in the water strategy, Gpt-4 got better every month since launch, making no sudden transition looking good jump wise):

GNeXwUtaMAInSDP?format=jpg&name=medium.jpg
 
AI just doesn't have the ability to type something out. Do a double take and say to itself, "what a ridiculous response", like I just did before scratching what I was going to post and posted this instead. :cool:
 
You must log in or register to reply here.
Back
Top