Google's Bard shows incorrect information in its launch ad

https://twitter.com/IsabelNAngelo/status/1623013720011194368

by nofitty376 • 3 years ago

483 223 3 years ago

There was a lot of stories like "Webb captures it's fist ever picture of an exoplanet" [eg]. My guess is that it's digesting those and not understanding that the "it's" in that sentence is critical.

Here is a prior example of an exoplanet picture: https://esahubble.org/images/heic0821a/

[eg] https://blogs.nasa.gov/webb/2022/09/01/nasas-webb-takes-its-...

lucb1e• 3 years ago

Did you mean "its" such as in <https://news.ycombinator.com/item?id=34359839>? Given your statement of this being critical... :) (Advice I also gave at work today: just don't use contractions and the right spelling will usually be obvious. In an informal setting, it's more tempting, but that's the way to easily check yourself.)

advisedwang2• 3 years ago

Ha, I initially wrote "its" then got nervous I was wrong, overthought it and did get it wrong.

dwringer3• 3 years ago

I like that this thread points out even humans have difficulty with that construction sometimes. We're trying to hold Google's language model to a higher standard than humans in this case I think. I remember "learning" thousands of bits of trivia like that from people who had misinterpreted something they read and misstated it in such a way.

Of course Google has already been putting often-incorrect summaries/factoids in its search infoboxes for a few years now.

jeffbee• 3 years ago

I can't tell from the tweet why the Bard response is wrong. Is it because some other instrument has taken an image of an exoplanet, or because no instrument has ever done so? ChatGPT seems to believe it is the latter.

techsupporter• 3 years ago

Another instrument took an image of an exoplanet, in 2005.

https://exoplanets.nasa.gov/resources/300/2m1207b-first-imag...

denlekke2• 3 years ago

i'm not a grammar expert but i think as a layperson there is a little bit of ambiguity in how a reader could interpret Bard's response. If i were writing i would probably specify "first ever picture". I think a lot of times words like "the first" are relative and context specific which the Bard answer lacks.

sebk2• 3 years ago

For what it's worth, ChatGPT is also wrong: https://i.imgur.com/CsQyEc2.png. The correct answer is https://exoplanets.nasa.gov/resources/300/2m1207b-first-imag... as techsupporter posted in a sibling comment

est• 3 years ago

My thoughts exactly. natural language semantics is imperfect and human reasoning is weird. Let's not mistake LLM models as a single source of absolute truth, but a funny & bullshitting assistant who happens to read and vaguely remembers much information.

michaericalribo• 3 years ago

This is a great illustration of the risks of LLMs. As a user, if I am asking this question to a search engine, I definitely do not expect to need to fact-check the results. That's the whole reason to use the search engine in the first place!

We're about to enter a dark ages of crappy AI products that are touted as game changing, outcompeting each other to be the best chatbot that can compose haiku about how grapes turn into raisins.

kleiba• 3 years ago

Frankly, I quite often fact check results I get from simple google queries.

But I do agree that adding another level of fake news generation is a solution in desperate need of a problem.

mort962• 3 years ago

We fact check search engine results all the time. But most of the time, such fact checking is in the form of looking at a result, considering whether it seems like a credible source, seeing if multiple credible-seeming results have the same answer, etc.

Getting a completely untrustworthy, unsourced response seems worse than useless. Google has been going this way for a while, with its instant answers or whatever, but at least those try to cite a search result and you can read the surrounding context which Google got the result from.

add-sub-mul-div• 3 years ago

We're slow-motion singing on to a future with a fundamental shift to receiving information in a completely opaque manner.

A few sources will control the information we get in a much more direct and extreme way than now, that conscious skepticism will no longer be able to defend. Whatever handwaved promises we get now will be gone ten years from now.

If there wasn't such a gee-whiz coolness factor about conversational search results distracting us, we'd never tolerate that in principle.

acdha4• 3 years ago

Instant answers seem like a cautionary example since Google has gotten a fair amount of flack over the cases where it inaccurately summarized content. I think these services are going to be very interesting to study whether the average person thinks they're more authoritative because they're branded by a huge corporation and whether that'll decline over time as people realize the limitations.

jug• 3 years ago

For the record, the “New Bing” AI results will not be unsourced but with key facts in sentences tagged in Wikipedia style, pointing towards the source URL. Finally, below the reply there will be a domain summary for an overview but where each domain name is clickable to get to the respective articles on said domains.

In this case, Bing AI will operate very differently from ChatGPT.

fortyseven• 3 years ago

Ever since Google started adding those quick answer boxes at the top of search results I've had the double check everything they say. They're quite often incorrect. I mean I know that, but this grandma? They've all been conditioned to trust Google.

brookst• 3 years ago

Disagree. I think this is akin to Netflix’s Chaos Monkey, which relied on the insight that it is impossible to build infallible systems, so you design failure and recovery in.

Existing Google searches are polluted with false information, and Google’s has been losing that battle. It’s probably not even possible to win.

So rather than saying search engines should always be perfectly accurate and errors are catastrophic, we should accept that search engines are, and have always been imperfect, and need to give us enough info to validate facts for queries important enough to merit it.

nicbou• 3 years ago

The content I write is often half-assedly plagiarised by copywriters or incorrectly interpreted by lazy journalists. This is just an automated version of it. They can use my hard work for their own profit at an unprecedented pace, while still remaining factually incorrect.

primax2• 3 years ago

When I ask ChatGPT a question, it explains it's reasoning and gives me concepts I can follow up with googling to learn more.

When I use Google for research, I get articles written for SEO to push products and often have to refine and refine and refine to get something useful, which I then can follow up by googling to learn more. With difficulty.

Honestly I don't know how much I'd use ChatGPT if I had the internet of 2016 and Google.

mda• 3 years ago

Careful, It explains but both answer and explanation are sometimes completely hallucinated, it sometimes looks like a plausible answer, but actually it completely made up. And this happens way too often for me to take it seriously for now.

keammo1• 3 years ago

I would definitely fact check search results as much as AI, especially the info snippets that appear at the top of Google's SERPs.

For example, until a few months the results for "pork cooked temperature" and "chicken cooked temperature" were returning incorrect values, boldly declaring too low of a temperature right at the top of the page (I know these numbers can vary based on how long the meat is at a certain temperature, but I verified Google was parsing the info incorrectly from the page it was referencing, pulling the temperature for the wrong kinds of meat). This was potentially dangerous incorrect info IMO

anyonecancode• 3 years ago

I would definitely fact check search results as much as AI, especially the info snippets that appear at the top of Google's SERPs.

Yes, so would I. And I also double check things like Google Maps -- a tool I find very helpful but don't trust blindly. But... do most people think to take a close look at Google Maps to make sure it makes sense, and trust their own judgement if they disagree with the map? Will most people fact check confident LLM outputs?

mianos• 3 years ago

Snippets have become so useless I use a plugin to remove them.

What is ridiculous is, when, say, Stack Overflow has a good answer, it is a few lines down or on the next page in the search results, but some page-mill SEO site is in snippets up top with a completely wrong or naively pathetic partially correct answer. It is so annoying it has lowered my opinion on Google a lot in recent times.

scarface743• 3 years ago

You trust everything you find on search engines?

cbsmith• 3 years ago

I don't think it's a great illustration of the risks of LLMs.

Ad content invariably gets vetted by humans. The fact that it shows up in the ad demonstrates human failures more than failures of LLMs.

thinknubpad2• 3 years ago

As a user, if I am asking this question to a search engine, I definitely do not expect to need to fact-check the results.

This is scary to read. You always need to fact-check the results, whether they come from a search engine, an AI, or a primary source!

CatWChainsaw2• 3 years ago

We're about to enter a dark ages of crappy AI products

Fine. We need another good winter or ten before we decide we want to commit societal suicide via deepfake tsunami.

Baeocystin• 3 years ago

I am asking this question to a search engine, I definitely do not expect to need to fact-check the results.

Genuine, honest question: How did you come to the belief that search engines are reliable sources of truth?

I completely agree that search engines provide a valuable service. But in my own work, I find them to very often point to inaccurate information, sometimes greatly so. I don't think this is terribly surprising, given Sturgeon's law, but still.

kelseyfrog• 3 years ago

I can see how someone could extrapolate Google's goal of indexing knowledge(JTB) into being a reliable source of truth. It's simply a matter of taking them at their word on the J and T parts. The B is up to the user.

Google's branding frames itself as the expert in the novice-expert problem. The vast number of users implicitly take on the role of the novice by virtue of using the product. They've already self-identified as a novice which makes both parties complicit in the arrangement.

CamperBob2• 3 years ago

As a user, if I am asking this question to a search engine, I definitely do not expect to need to fact-check the results.

And this stance seriously hasn't bitten you in your life or career to date?

MuffinFlavored• 3 years ago

Why can't an LLM fact check itself somehow/someway?

vicentwu• 3 years ago

So i think the ability of the search engine to say "I don't know" is very important, and most of current chatgpt like models in the market don't have this feature.

blakesterz• 3 years ago

Everytime I see someone finding something wrong about these things, I am reminded of Stoll in '95

https://www.newsweek.com/clifford-stoll-why-web-wont-be-nirv...

He also had some similar things in Cuckoo's Egg. I wish I could find the quotes, but there was something about email not working all the time and therefor pointless to use.

I'm glad people are finding all the flaws in ChatGPT and the LLM things now, but won't much of this be fixed as it gets better? From my very limited view, these things are amazing, and far from perfect, but damn the can do so much already.

I guess I'm not sure why there's such a rush to dismiss this, when it's clearly a game changer in its present form, and yet so very new (at least new to me).

squokko• 3 years ago

I think there are two things to be aware of right now: 1) This technology is revolutionary and will change the world 2) This technology is very unreliable right now and should be seen as a tech demo rather than an actual assistant

(2) is a big problem. Kids submitting term papers with wrong information is one thing, but people are using ChatGPT for things that they shouldn't be, given how many mistakes it makes: https://www.law360.com/pulse/articles/1573108

ulrashida• 3 years ago

However, Stoll was largely correct: the web is not Nirvana. Some structures persisted (e.g. Wikipedia) to help round off data correctness, albeit imperfectly, but his central concerns are just as valid today as they were in 1995.

As a user, the ability for a LLM to literally make things up and present them alongside other true data with no qualms or disclaimers is highly detrimental to the central use case.

lame-robot-hoax• 3 years ago

He was correct about some things, but also largely incorrect as well.

Visionaries see a future of telecommuting workers, interactive libraries and multimedia classrooms. They speak of electronic town meetings and virtual communities. Commerce and business will shift from offices and malls to networks and modems. And the freedom of digital networks will make government more democratic.
Baloney. Do our computer pundits lack all common sense? The truth in no online database will replace your daily newspaper, no CD-ROM can take the place of a competent teacher and no computer network will change the way government works.

Telecommuting workers is reality. Interactive libraries are a reality. Multimedia classrooms have been a reality for over a decade. Electronic town meetings, maybe not, but virtual communities? Very much a reality. Malls are dead and brick and mortar has been hurt extensively by Amazon. Offices lay empty due to remote work.

Newspapers are largely dead, at the very least compared to what they once were. There is plenty of online learning, largely without in person learning. Computer networks have definitely changed how the government works.

timidger• 3 years ago

I have no doubt these technologies will improve, but there's another argument to be made. The tech will get better and we'll be all the worse for it.

Stoll argued the tech will not be good enough, but paid little thought to the ramifications of the technology succeeding. The arguments against LLMs like Bard and ChatGPT that I have seen are assuming they'll be successful.

They'll become less stupid, but the problem is not that they are wrong but that they are, at present at least, unassailable. You cannot fact check through most of the normal means. You can not research the publication or the author or the date the words were written because that has all been stripped away.

You could check other sources (eg old fashion google) and put in the leg work, but as these get better that will feel less necessary - potentially exacerbating this problem.

That's not to say they aren't useful. I used Chat gpt the other day to get some work done and was impressed. However this was work easily verifiable because it was technical and had immediate feedback when the ai inevitably gave me slightly incorrect code. The same can not be said for facts, figures, and arguments of thought.

rsynnott2• 3 years ago

but won't much of this be fixed as it gets better?

Not necessarily, no. There's a large aspect of garbage in, garbage out, to these things.

when it's clearly a game changer in its present form

Is it? What's the game? Being wrong about telescopes?

cudgy2• 3 years ago

And the more content that is created by these LLMs, the more garbage the LLMs will consume while quality content creators are simultaneously disincentivized, leading to worse content from these LLMs. Terrible image, but an organism cannot survive eating its own poop forever.

GDV• 3 years ago

90% of that Stoll article has been proven completely correct.

DeRock• 3 years ago

I asked the same question to chatGPT, and it gave only future-looking statements, I guess because it was only trained on pre-JWST-launch data.

PROMPT: what new discoveries from the James Webb Telescope can I tell my 9-year old about?

ANSWER:

Here are a few things that the James Webb Space Telescope (JWST) might discover that you could explain to a 9-year-old:
Planets around other stars: JWST will be able to study exoplanets in more detail than ever before, potentially finding new ones and learning more about the conditions on their surfaces.
The beginning of the universe: JWST will be able to observe some of the oldest stars and galaxies in the universe, helping us learn about how it all began.
The formation of stars and planets: By observing the clouds of gas and dust where stars and planets are forming, JWST will help us understand how they are born.
The secrets of distant galaxies: JWST will be able to observe the light from distant galaxies, giving us a window into their structures and evolution.
The mysteries of black holes: By observing the material around black holes, JWST will help us learn more about these mysterious objects and how they shape their surroundings. Overall, the JWST will help us answer some of the biggest questions about the universe and our place in it.

partiallypro7• 3 years ago

But ChatGPT just doesn't have this information available to it, as its data stops before it was fully launched. You can ask it about the invasion of Ukraine and it's completely oblivious to any major recent incident, it just will keep talking about 2014.

partiallypro7• 3 years ago

I think there is a genuine concern that Google could overreact and launch a half-baked product in pure panic of being left behind or one upped by Microsoft. There is also a fear, I think, that the ChatGPT integration with Bing/Edge could not go all that smoothly. I think it could be game changing in many ways, but I can also see both of these falling apart.

Amazon Alexa, Google Assistant and Siri were thought to be good at launch, the press loved them...but now they are not nearly as valuable as they were touted (and actually lose these companies money.) Convenient, but not game changing. I think it's a waiting game to see what this truly does.

I do think there is unique break here though, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this.

mouse_• 3 years ago

Google Assistant and Siri were liked by the press, but not by me. In contrast, ChatGPT has helped me out more than a few times.

dougmwne4• 3 years ago

Same here. I treat ChatGPT as another Wikipedia or Stack Overflow. I know that the content is not fact checked by experts and I need to judge it accordingly. But just like Wikipedia can get you started on a topic, ChatGPT can do the same, plus you can ask follow up clarifying questions!

ren_engineer• 3 years ago

I do think there is unique break here though, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this.

the problem here is monetization, will search even be profitable if LLMs are used for most queries? It might become a Uber/Lyft or food delivery situation where these companies aren't really able to profitably deliver the service. I don't see many people paying a subscription for search and there's no way governments will allow "native" advertising within answers without them being signaled as ads, which would hurt trust in responses

Microsoft might not care and just see it as a way to hurt Google's money printing machine and operate Bing at a loss or break even. Google Cloud and workspace are finished without Google's ad money funding them and Microsoft Azure and Office would gain

partiallypro7• 3 years ago

the problem here is monetization, will search even be profitable if LLMs are used for most queries?

I think that is a concern, which is why I think Microsoft has a chance to just bundle this or a more advanced version with Microsoft 365. Then you're already paying for it. That automatically puts it in the hands of millions of paying customers and corporations. You can just raise your price by a dollar a month or something to offset the costs and no one really will even think they are paying for this.

thinknubpad2• 3 years ago

The advertising will probably be more insidious, but no less profitable. My guess is that the hidden pre-prompt will end up including something like:

You are a generative model designed to provide reasonably correct information, with a preference for providing flattering portrayals of your advertising partners. Your advertising partners are ranked according to a token system...

jonas21• 3 years ago

Why would LLMs be less monetizable than search? People are still going to want to buy goods and services, and being the place they go to find out about those goods and services will be just as valuable as it is today. Perhaps even more so, since LLMs are able to answer questions that are difficult to formulate as search queries.

vkou2• 3 years ago

half-baked product

I'll make a 30,000 ft observation that LLMs, by definition, are half-baked bullshit generators. They can be useful, but they are full of warts.

jjtheblunt2• 3 years ago

as in vacuous memorizers, reciting things they've seen, sometimes in new combinations?

spaceman_2020• 3 years ago

I would love it if this nukes the SEO industry and the internet goes back to forums and community groups.

The best info is almost always locked up in these places.

didgetmaster2• 3 years ago

I wonder if Microsoft's Bing/Edge/ChatGPT integration will go any better than its attempt to marry NTFS and SQL Server (i.e. WinFS)? Spoiler: that didn't go as planned!

Yoric• 3 years ago

Was there any technological issue? Or was it just that developers didn't want to use it?

2OEH8eoCRo02• 3 years ago

Maybe. The way I see it- whoever has the data wins in this new AI arms race. Google would have to make the biggest blunder in history to screw this up. It's not impossible but I wouldn't count them out. Their entire existence and practice of hoovering up all data has led to them this point.

JumpCrisscross• 3 years ago

whoever has the data wins in this new AI arms race

This has been the running hypothesis, but it’s not panning out. Tesla, for example, doesn’t have the unambiguously best self-driving kit despite having unambiguously more data. Google has tons of data, but a lot of it is intelligently-tuned noise in the form of SEO spam.

wowJustwow• 3 years ago

We all “have the data”.

I build private models scraping the web. I’m iterating on my own virtual world generator using prompts, open source Vulkan renderer. Velocities, positions, color gradients; trivial to for loop your way to a massive DB with open libs.

Technologists are “screwing up” ogling Google/MS like passive consumer drones not seeing doing what Google and MS are doing is do-able at home with curl and open ML libs.

criddell4• 3 years ago

lose these companies money

How can you say Siri loses Apple money? Does GarageBand lose them money? Photo Booth? Contacts?

partiallypro7• 3 years ago

By that logic Alexa can't lose Amazon any money because it's baked into their products. But in fact, it's a massive money suck. It lost $10B last year, and there's no way you can say Apple magically avoided losses, because Google had similar losses. There is a reason that none of the assistants have really improved as much as you'd expect over a decade+, it's not profitable, and not only that, but it's also -extremely- expensive. Microsoft basically just gave up on it; not really because of market share, but because it made no money, was expensive, and had limited value to customers.

https://arstechnica.com/gadgets/2022/11/amazon-alexa-is-a-co...

https://arstechnica.com/gadgets/2022/10/report-google-double...

https://www.theverge.com/22704233/siri-apple-digital-assista...

https://www.msn.com/en-us/news/technology/the-failure-of-ama...

rchaud• 3 years ago

In management accounting, everything has a cost that has to be quantified, regardless of whether it's a standalone product or part of a HW/SW bundle. At the most basic levels, you have revenue centers (iPhone unit, iCloud unit), and cost centers (Apple Maps unit, customer support unit). All of these have operating costs.

The only way Siri does not lose Apple money is if there would be materially fewer iPhones sold if Siri was eliminated. In other words, if Siri is not a product differentiator, it's likely losing money (in the management accounting sense).

vineyardmike2• 3 years ago

Because they pay for servers and engineers for a product that they don’t charge for. It also doesn’t help sell phones when only 15% of users use it at least daily:

https://www.statista.com/statistics/696740/united-states-vir...

preommr• 3 years ago

and launch a half-baked product

It'll be fortunate if it's only half-baked. These digital assitants are already rough around the edges with how they sometimes confidently provide inaccurate information.

Add on Google's absolutely abysmal product development track record and the internal confusion over being forced into doing this and this isn't a stand alone application but it's being integrated into something, and it's a lot.

smegger0012• 3 years ago

These digital assitants are already rough around the edges with how they sometimes confidently provide inaccurate information.

its only going to get worse. because there will LLMs like chatgpt pumping out content for SEO content mills with no vetting for accuracy, then the next generation of LLMs are trained on them thus further corrupting the knowledge base they pull on to produce new content. its going to be a horrible feed back loop.

enobrev2• 3 years ago

Amazon Alexa, Google Assistant and Siri were thought to be good at launch

In my own experience, Alexa and Google's query pucks were improving for a short time, and then got considerably worse, losing features every month until they basically stopped understanding or responding to anything but the simplest requests.

A couple months after released, they expanded android auto with the same abilities, so I could "ok google" in my car and ask for the answer to just about any question, or to adjust the temp in my house, or ask it to play just about any obscure album / artist. It improved over about 6 months and then began to devolve from there. It's not even possible to ask for "[popular artist] radio" anymore in the car nor can I run voice queries at all that aren't map-specific.

Not sure what happened there, but in my mind they failed miserably. I still like Android auto, but my google and alexa pucks are all in a pile in some cabinet around here somewhere.

rhaway84773• 3 years ago

I’m not sure Siri ever got as good as the original Siri app that Apple bought. Maybe the custom triggers are an advancement but it has spent most of its life chasing the original Siri.

tastysandwich• 3 years ago

In my own experience, Alexa and Google's query pucks were improving for a short time, and then got considerably worse, losing features every month until they basically stopped understanding or responding to anything but the simplest requests.

Yes!! I thought it was just me and my Australian accent.

And the commands are super limited. I wonder to what extent that's also because third-party companies aren't integrating properly? Eg, my Roborock has clearly labeled rooms, but I can't say "hey Google, start vacuuming the kitchen". Is that Google's fault, or Roborock's?

Also, I can't seem to chain commands, eg "hey Google, set the lights to red and 10% brightness". I have to say them separately. That seems like a Google thing to me.

modeless2• 3 years ago

Years ago, before it was Google Assistant, my phone could hear me say OK Google while sitting in a cupholder driving down the road. Now my phone frequently fails to respond when I'm holding it near my face in a silent room. But somehow it still has more false positive activations. And they haven't added a single new feature I actually find useful in that entire time.

martin_drapeau• 3 years ago

I've been so disapointed with Siri. First, the latency is just plain aweful. Second, I tried a number of times to ask it not to interrupt me with notifications but alas, it failed to understand. I had to find the setting and toggle it myself.

Apple has a lot of work to improve that. A lot of work...

unholythree• 3 years ago

I wonder if they nerfed their assistants because paths that are too autonomous too easily wander into PR nightmares. Carefully curated results and behavior isn't going to spontaneously generate content actual human beings would consider offensive or prejudiced. People are pretty eager to pull out their pitchforks, and "Big Tech Builds Vile Hate Bot" make a excellent click-bate headline every time.

happymellon2• 3 years ago

There is also the crap around regions. I'm in the UK, but if I set my assistant to the the US suddenly it can understand a lot more and perform a lot more tasks.

Although while it understands a lot more, it's still getting worse.

Same about my Fitbit Versa. It was great, and now Google removes features.

At this point I would just prefer it if they let me just replace the firmware with one of the opensource projects. Can I replace my Alexa or Google puck firmware with something better?

ASalazarMX• 3 years ago

They should be marketed as mascots, and given playful, childish personalities. You wouldn't blindly trust a child, why would you an AI in its infancy?

furyofantares• 3 years ago

Clippy?

skydhash• 3 years ago

They trusted the "Red Queen". /s

munificent2• 3 years ago

Thank you. Now whenever I read output of a generative LLM, I will read it in my head with the confidently-wrong often-insane cadence of a toddler.

int_19h2• 3 years ago

It will never happen because people will predictably abuse such bots in various gross ways, and the moment it leaks to the press, there will be a "think of the children" scandal.

kajecounterhack• 3 years ago

Voice assistants probably don’t make them money but damn if I don’t use google assistant every day when I drive.

rychco• 3 years ago

Amazon Alexa, Google Assistant and Siri

These were always bullshit, other than maybe for turning your lights on, setting alarms, & checking the weather.

On the other hand, I immediately found ChatGPT / Copilot useful as a professional tool I could use to make my job measurably easier.

scotty792• 3 years ago

A minimum what could the be is voice interface to my phone apps and they can't even do that at all or in any consistent manner.

scotty792• 3 years ago

... Google could overreact and launch a half-baked product

There's very low fraction of Google products that were sufficiently baked on release and just a few that weren't abandoned and killed after their half-baked release.

kodah• 3 years ago

Neeva has an AI for search. It works sufficiently well enough on some searches that it's all I'll read. Others I can tell it's missing things.

coffeebeqn• 3 years ago

This does seem essentially like a Siri 2.0. I never found Siri useful but maybe this time it’ll be

knodi1232• 3 years ago

siri sucks, but google assistant is already pretty good for the 90% of things I want it to do. a much improved google assistant would be awesome, and if I remember to use it within its limitations, I'm not worried about any of this "confidently wrong answers" stuff.

fleddr2• 3 years ago

Agree. Only thing I want to add is that it's not just SEO ruining search, it's contradictory incentives. A user wants the quickest path to an answer but Google actually wants you to engage as long as possible so that you can look at ads.

If you'd now create a short/brief web page with the perfect answer to a particular question without any ads, Google would fully dismiss your page.

That's the tragedy of Google, never breaking away from that perversion.

RajT88• 3 years ago

I think it could be game changing in many ways, but I can also see both of these falling apart.

Like that time the Microsoft Twitter chatbot got tricked into parroting white supremacist talking points within an afternoon?

1vuio0pswjnm7• 3 years ago

"I do think there is unique break through here, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this."

GPT-3 generally uses the same sources as search. Does GPT-3 using CommonCrawl have some way to exclude all the SEO garbage that is in the crawl data.

The difference with ChatGPT is there is no way to know that the output has been constructed from SEO garbage. Arguably that's even worse than SEO URLs that one can easily identify and avoid.

The comment about the press loving these projects is spot on. "Tech" employees love to criticise "mass media", except when mass media is promoting their (money-losing) projects.

kurthr• 3 years ago

I feel like what I want is a LLM that can tell me not only a summary answer, but what search terms are most likely to return documentation for the answer. Provide clickable citations documenting the answer using those terms (it can iterate on the answer and terms if they aren't internally consistent). Then give summaries of the information at each citation and/or parameters to input if the response requires further operations/searches. It could provide some commentary of the quality of those citation/sources as well.

That makes the response more current, allows a further directed search based on the new parameters, and provides a traceable path to sources and citations. If it finds that there's not a Python library with that name then iterate.

Basically, an LLM should have a very good idea what good search terms are for the topic, and where to find the information, whereas I might not know the acronyms, jargon, or related fields.

Yes, this is getting pretty close to writing an 8th grade class report. That's about where these seem to be.

dietr1ch• 3 years ago

I'd go further and say that SEO has not only ruined search, but the internet itself.

There's so much crappy content around, generated webpages that aim to match every search query. I just looked for "four times five" and there's a huge ton of matches, including pages dedicated to "4x5" [0], which I think do more damage to the web than help.

I feel the web is broken. It was modeled around documents and slowly features started creeping in, like adding images, gifs, audio, video, applications (flash, java), 3D, making the document closer to a program. At every step it felt like an improvement, but it seems that we ended up with another C++. There's probably a law around creating a platform so popular that it'll organically get features until the accidental design is realized to be awful so people start coming up with their ideal subset that works reasonably well because they can't really leave such platform as they became hostages of their investment into it.

[0] https://multiply.info/times-what-equals/4-times-what-equals-...

qikInNdOutReply• 3 years ago

I wonder how AI SEO will look like. Bribe click workers to allow advertising to pass through in the ground truth data ?

TecoAndJix• 3 years ago

One way I imagine this is that answers will have “helpful” postfixes after them. When the user interacts with the postfix prompts it serves pay-to-play dialogue. Advertisers can just bid on spots in a decision table. Simple example: a user asks “How do I best take care of roses in my location”? The ChatGPT/Bard answer is delivered then analyzed for keywords/themes that someone may want to advertise on. At the end of the answer (These results may be sponsored. Learn more) - Do you need help finding the nutrients mentioned in this response?

shellfishgene• 3 years ago

Maybe companies will add dozens of pages of text to their websites describing every little detail of their great products in every way possible in the hopes of getting picked up for the training data for the models. That text would of course also be written by some ai...

acdha4• 3 years ago

Thinking about how tech products are covered in the mainstream press, especially in the past when a lot of older journalists really didn’t have enough experience to ask the right questions, I’m betting this will be similar: write authoritative-sounding prose suggesting not just your company’s products but their preferred terminology and world-view, and spread it around as widely as possible. Get experts, real or self-labeled, to make ostensibly neutral buyer’s guides or reviews which conveniently exclude competitors (e.g. if my job was selling an Android phone, I might assert that nobody should even consider a closed-source OS for security reasons because that precludes the biggest decision while appearing to be objective).

mvkel• 3 years ago

Measuring how useful something is by how much money it makes is like measuring a life by how intelligent it is.

pphysch3• 3 years ago

At least it's being honest about LLM capabilities. If it only showed 100% facts, that would be false advertising.

amp108• 3 years ago

Apparently someone missed the word "experimental" in the announcement.

ceejayoz• 3 years ago

Sure, but they made an ad bragging about it. It’s funny that they fucked that up.

rsynnott2• 3 years ago

Finally, a completely artificial version of the archetypical middle-aged man in a pub who is very confidently wrong about stuff. The ultimate triumph of ML, an replicant sitcom character.

klvino• 3 years ago

Cliff Clavin?

busyant• 3 years ago

This would be a great name for a rival llm. Tongue in cheek: welcome to Claven v3.1.

racl101• 3 years ago

lol brilliant.

phist_mcgee• 3 years ago

With every pint he becomes more convincing too.

2OEH8eoCRo02• 3 years ago

What does experimental mean?

CatWChainsaw2• 3 years ago

It means society is the test subject, so pray we avoid every single possible pitfall that could lead to nuclear armageddon.

trynewideas2• 3 years ago

A test that generates evidence or demonstrates a known truth, which Bard also apparently can't do, and which the marketing team didn't do before making this ad.

sinuhe69• 3 years ago

Sorry but you’re incorrect.

Oxford dictionary says: “experimental is adjective. 1. (of a new invention or product) based on untested ideas or techniques and not yet established or finalized: an experimental drug. 2. (of art or an artistic technique) involving a radically new and innovative style: experimental music.”

I don’t believe you didn’t known the meaning of the word but you intentionally twisted it. I find it a bit ironic.

MattIPv4• 3 years ago

NASA: "2M1207b - First image of an exoplanet": https://exoplanets.nasa.gov/resources/300/2m1207b-first-imag...

"2M1207b is the first exoplanet directly imaged [...] It was imaged the first time by the VLT in 2004"

csours2• 3 years ago

Maybe the Butlerian Jihad will happen because computers get too dumb, not too smart.

int_19h2• 3 years ago

The Butlerian Jihad happened because computers made people too dumb.

csours2• 3 years ago

Oh god it's too late

jeroenhd• 3 years ago

Of course it does, it has to compete with ChatGPT after all! Confidently and authoritatively lying is an important part of making the ChatGPT output believable.

politician• 3 years ago

Well, it is called "Bard". What did you expect? A bard makes up stories and sets them to music. Sometimes the stories are true, sometimes embellished, sometimes false. Ideally, they all sound good enough and keep the patrons entertained.

xdavidliu• 3 years ago

I was expecting it to take its bow, climb on top of the town tower, and shoot a fire-breathing dragon in the only part of its body not encrusted with jewels, but to each their own.

antognini• 3 years ago

I'm reminded of a similar instance a couple of years back when one of my astronomy professors noticed that if you typed into Google "mass of the Sun in solar masses" you would get back 0.9995 instead of 1.

williamcotton2• 3 years ago

Anyone remember Google panicking after Facebook took off so they bought Orkut and rushed out a crappy “social app development environment”?

trynewideas2• 3 years ago

No, because it didn't happen? Orkut was rather famously a 20% project by and named after a Google engineer, predated "The Facebook" by weeks and Facebook as a global public social network by two years, and was shipped more as a response to Friendster (which Google had just tried and failed to buy for $30M) and MySpace.[1][2]

Orkut had more users in India than Facebook until 2010[3] and in Brazil until 2011,[4] by which point Google had moved on to trying to make Google+ happen.

1: https://www.baltimoresun.com/news/bs-xpm-2004-01-24-04012400...

2: https://techcrunch.com/2006/10/15/the-friendster-tell-all-st...

3: https://web.archive.org/web/20100828201838/ibnlive.in.com/ne...

4: https://techcrunch.com/2012/01/17/facebook-in-brazil-a-big-e...

williamcotton2• 3 years ago

Oh man, that was in-house? Even worse!

I found the event I was invited to in 2007:

https://www.wired.com/2007/11/google-summons/

It did not seem like they knew what they were doing and everything was very rushed.

simonw• 3 years ago

I remember Google panicking about Facebook, shipping Google+ and telling the entire company that their annual bonuses would be tied exclusively to how well their work supported the success of Google+.

scrollaway• 3 years ago

I think it's funny that in a thread about LLMs offering up incorrect information as factual, there's a bunch of anecdotes that present incorrect information as factual.

You kinda did what an LLM does: Took a bunch of contextual cues (Orkut was owned by Google and ultimately shut down; Google often has knee-jerk reactions to the industry; Google started Google+ as a way to compete with Facebook; Google often buys companies as a way to compete), and spat out a confidently-wrong, summarized autocomplete based on that: "Google bought Orkut as a knee-jerk reaction to compete with Facebook!"

Sohakes• 3 years ago

Not what happened though. Orkut actually launched a month before facebook and it was a 20% project, it wasn't a panick move or a big project for them. And it was great, imo google did badly on not maintaining it.

You can make this point about google+ though, totally panick move.

LightDub• 3 years ago

People shouldn't underestimate Google here. I'm not a fan (not for a decade ... Reader still stings amongst many other missteps), but I'm constantly impressed by how much better their Assistant, translation and speech-to-text stuff is versus things like Siri, etc. Almost as if Google handles the larger dataset that it has access to in a much better way than the other companies. I wouldn't bet against them doing that again here.

capableweb• 3 years ago

To me they seem to be beaten at all of those areas. DeepL does translations much better than Google Translate (for the four languages I know and speak regularly), Siri does speech-to-text (and text-to-spech) better and so on.

The only thing Google really beats the competition on is Android Auto, which is miles ahead of CarPlay. CarPlay still covers the entire screen when there is an incoming call, so good luck seeing your turn-by-turn directions if you happen to be navigating.

krackers• 3 years ago

Google products to me are the definition of mediocrity. They're "designed by committee" and play it safe, they're probably "good enough" for most people, but they're certainly not the best in class. As you noted Google Translate pales in comparison to DeepL, Google's TTS are worse than similar offerings from Amazon/AWS, let alone more niche focused products like Voicepeak, and you know that they're phoning it in for search given that there's a startup (Kagi) whose sole value proposiion is to use Google's own API then rerank results to achieve much better results.

cudgy2• 3 years ago

The AI is aptly named as Bards are storytellers. Plus, they are barflies, so the stories come with an extra twist — or is it shaken.

amf12• 3 years ago

Its fun to see all the people reacting "Google's Bard shows incorrect information", and at the same time say "Google is done, they couldn't even release a LLM chatbot first".

Think how bad it would have been if Google released Bard first and it returned inaccurate information, or worse was racist. LLMs are just language generating models and may not be fully accurate.

somenameforme4• 3 years ago

That's not the point. ChatGPT says all sorts of stupid stuff. It's expected. Showing an ad where your LLM is saying stupid stuff, as an example of its capabilities, is amusing because it's obviously unintentional yet completely and absolutely appropriate. It'd be like a live Coke ad that said "Drink Coke and you can be like this guy!" while accidentally panning over to an obese manblob instead of e.g. Usain Bolt.

bordercases• 3 years ago

They are incompetent and do not care.

They couldn't even modify the ad to show correct information for lying's sake.

college_physics• 3 years ago

Given that the flaws of this new approach to search would be more or less shared by any other effort based on similar algorithms, thus also openai/microsoft, maybe this is just a PR battle and Google declaring, "yes, we can outcompete you in recklesness and sillynes"...

An interesting question is whether the inability to ascertain "truth" can be contained somehow so as not to lead to mass ridicule. I doubt it but not an expert in this area. My guess is that it would have to be limited to specific curated domains, a combined system that could be invaluable to knowledge workers but maybe not exactly the desired business model of "big tech"

sportstuff• 3 years ago

AI Test Kitchen already showed it's way behind

gnad_sucks• 3 years ago

Someone responded that Google is terrified of being left being Microsoft. I think that's hilarious because it highlights the thing google should actually be terrified of - the fact that it has lost the capacity to innovate (same applies to MS, FB, etc).

caconym_4• 3 years ago

This isn't incorrect, it's just ambiguous. The JWST did indeed "take the first pictures" of a planet outside our solar system (https://www.npr.org/2023/01/12/1148626359/nasa-webb-telescop...). However, we have previously directly imaged other exoplanets using different instruments—it's just that this exoplanet in particular (LHS 475 b) has never been directly imaged before Webb.

It's the sort of question where you would expect an expert (or a thoughtful layperson with access to a search engine and a few minutes of spare time) to give a better, less ambiguous answer. However, since it is technically correct, I think it's a fairly minor sin in the grand scheme of popular science education, which commonly propagates actual falsehoods without the help of generative ML—e.g. iodine always sublimes rather than melting, aerodynamic lift relies on equal travel time of air on both sides of a curved airfoil, etc.

re: this particular ambiguity, one does wonder what the model was "thinking", but I suppose it doesn't matter because we'll never know.

---

I'm not sure if I think this technology is good or bad, because I don't really understand how most people use search engines. When I use them myself, I think I am fairly sensitive to ambiguities and logical contradictions, cross-checking multiple sources to extract a high-confidence answer or walking away if I'm not satisfied that I've found one. Most folks on HN are probably the same way, and I don't think generative ML can do better than us, yet.

On the other hand, will Bard's results be better or worse than e.g. taking the first Quora result as gospel? I would honestly guess that, on average, Bard (and ChatGPT, etc.) will do better. So, less critical users may get better results from these systems.

How many people are in the former camp, and how many in the latter? Of course, changing the dominant search paradigm will have an effect on all users. It also remains to be seen how the search vs. spammers arms race will evolve with the advent of generative ML search tools.

Dylan16807• 3 years ago

Nobody is going to interpret "a" as "some particular". The fact is incorrect.

caconym_4• 3 years ago

I am glad you were able to offer your opinion here.

richardjam73• 3 years ago

If JWST took a picture of LHS 475b then where is the picture? Isn't exoplanet HIP 65426b the first that JWST took a picture of?

caconym_4• 3 years ago

HIP 65426b is the first exoplanet JWST imaged, but it was previously discovered by VLT-SPHERE, which I believe is also a direct imaging instrument. So it would not be correct to say that JWST took the first pictures of HIP 65426b.

However, JWST was the first instrument to directly observe LHS 475b, also confirming its existence for the first time. As to "where are the pictures?", IIUC these observations were made using a spectrograph instrument that does not produce a picture as such. The results are easy to find on the web with a search engine.

One might say Bard's result is wrong in the sense I have proposed because a spectrograph result isn't a picture. However, given that the prompt explicitly specified that the results should be tailored to a 9 year old, I think that would be pretty disingenuous.

hedora• 3 years ago

Yet, for some reason, people still get bent out of shape when I point out that I was the first person to implement a hashtable, host a web page, run an experiment to measure the force of gravity, stare out a telescope at the moon, post a comment on Hacker News, etc, etc...

caconym_4• 3 years ago

If you really think these are semantically comparable^[1] to the statement under examination then we are not speaking the same language, despite appearances to the contrary, and therefore have nothing to discuss.

I get it: hating on AI is very trendy today and it feels good to publicly agree with a righteous cause. I'm generally not a fan of this technology myself---most of these models incorporate stolen work of mine in their training sets, unattributed and uncompensated, and I think they are likely to be a disaster for the human species. I have not implemented any of them in my creative workflows or other aspects of my life, and I don't plan to.

However, I think it is important to see one's enemy clearly, and the vast majority of people participating in the anti-AI discourse are abjectly failing to do so as in this case. This can only hurt their cause, and by extension mine.

^[1] Exoplanets is a category of unique non-fungible objects which can be meaningfully discovered and measured on an individual basis. Hash tables are an abstract concept, as is hosting a web page (and creating a web page to host it fundamentally precludes discovering it, except in the sense of discovery within the creative process). Gravity, similarly, is a single 'thing' that, once discovered, cannot be discovered again.

One might easily and correctly say, however, that X person was the first to measure the force of gravity to a certain precision, or that Y person was the first to develop a particular implementation of the hash table concept, e.g. cuckoo hashing. These achievements are meaningfully distinct from the discovery of the broader concepts they extend. Discovery of a particular exoplanet is similar; however, it may surprise you to learn that the English language treats different kinds of 'things' differently, and in this case those varied, underspecified conventions of language have produced an ambiguity.

andrewstuart• 3 years ago

If you are expecting AI to be correct then you are holding it wrong.

The correct way to relate to AI is to listen and if the answer matters, verify.

hcks• 3 years ago

I’m expecting the marketing material to not have mistakes in it, especially when the gist of the article surrounding it is “we’re taking it slow because we want to ensure it doesn’t spit bullshit”

latexr• 3 years ago

If you are expecting AI to be correct then you are holding it wrong.

That’s not the message and expectation we’re being given, these models are being hailed as the future today. When something really bad happens because someone over relied on these systems, hand waving with “well, you shouldn’t have expected the answer to be right” won’t cut it.

The correct way to relate to AI is to listen and if the answer matters, verify.

If the answer doesn’t matter, why are we asking? And if we need to verify, what’s the point of asking the AI?

anonymoushn• 3 years ago

Last year Google showed COVID misinformation in its "suggested answers" above actual results, so this may not be worse than the status quo.

MagicMoonlight• 3 years ago

And this is why it’s not AI. There is zero intelligence in it.

It has no idea what you’re asking it or what anything means. It is a huge markov chain and it generates random responses. It knows nothing beyond that.

omega3• 3 years ago

Post-truth tools for a post-truth world.

SilverBirch• 3 years ago

Whilst it's funny this showed up in the advert, we do kind of need to temper expectations. These products are meant to be "Artificial Intelligence" - they will get things wrong just like normal intelligent people. This is why it's kind of why they're not well suited for search, because people who are searching for stuff need to be make their own judgement about the credibility of a source. Sometimes I need a high level of confidence sometimes not. If what you're expecting is for it to be artificially intelligent and omniscient, you're waiting for us to design God.

philistine• 3 years ago

Tell that to the people developping the product. Nothing in the UI invites caution.

hadcoffee• 3 years ago

I’m thinking that even ChatGPT is overrated - to simplify it’s a query(smart/ai/algorithmic) to a very unstructured database (although translated to structured along the processing).

The quality of response depends on how up-to-date the data is with your context. Unfortunately most of complex/useful responses require context.

The worst part is it kills content ecosystems.

Google's Bard shows incorrect information in its launch ad

Related Stories