Google I/O and the Coming AI Battles
stratechery.com
stratechery.com
Typical big co move.
Also the amount of tweets that suddenly popped up on Twitter with the same template of "ChatGPT is finished! Now that Bard has launched here 10 tips on how to..". It was clearly a Google marketing push ( scummy as always, but Google have done way worse in the past ).
This leads me to believe that Bard is at best OK and has no "AI advantage" over other competitors and it's basically who has the bigger ranch and deeper pockets until everything is dead around them.
As if the spammy AI threads couldn't get worse, they managed to do it.
AI hype guys on Twitter work pro bono. Nobody pays them they just shill whatever the latest thing is.
Google has no central AI strategy because the company has weak leadership and is largely directionless.
The distributed actions of a lot of people does not amount to any concerted plan.
The other AI products at google were just polishing up their old stuff to hastily release. Gemini is their real attempt at catching up to GPT-4.
On the other hand, bard performed much better at a more complex task: I provided fake data spanning six years and a few rows of different categories. I left the sixth year blank. Without prompting it, reformatted the from comma do you limited into a nice table. It also gave a brief summarization and analysis of the trends. When I asked follow-up questions he gave detailed responses that were moderately insightful for anyone that would be familiar with the data domain given a few minutes to review the data set. Finally, I asked it to give a projection of what the values would be for the final column, which I had left blank. It did so, with an explanation of the trend it observed, as well as a description of the linear regression model. It used it and apologized for not being able to give more complex modeling due to the Small data set.
GPT-4 provided answers and analysis that were comparable in quality, but required additional rounds of prompting and clarification.
It obviously remains to be seen how good Gemini will be. I just wanted to clear up the apparent confusion.
Why do you assume not being explicitly paid by Google means not being part of Google's marketing strategy?
Google's whole business model is getting people to do work for free ("produce content") by making them believe they are doing something else.
I would’ve believed the first three were coincidence, but by the time I got to tweet number 4 I was already re-reading Submarine. http://www.paulgraham.com/submarine.html
This is what happens when they are all using the same ab testing and title generation tech.
Google has billions of dollars it doesn’t know what to do with: they can certainly part with a tiny fraction of it and pay some PR firm to save face after the ChatGPT thing.
What often happens is a few paid employees get the ball rolling and everyone else starts doing the hyping for free.
I thought this was the case, until I saw how identical all of the messages were, including the order of the bullet points. It was clearly some form of paid marketing - or fake accounts, except they all had blue ticks (not that that means much these days).
This has always been the case for all corporate fanbois on Twitter. As an ancient example: there was a time when Steve Jobs proclaimed that 3.5" screens were the perfect size, and 5.5"/6" phones were too big(!!). This was followed by suspiciously similar talking points recirculating in the "iPhone Twitter" echo-chamber[0]: including the same doctored image[1] that compared thumb-coverage of a 3.5' iPhone vs the Galaxy Note (IIRC). I don't think Apple was directly paying people to say those things: they simply confirmed their biases and they felt compelled to share the "truth".
0. Trended on Twitter, and even appeared on blogs written by Apple superfans like Marco Arment and Gruber
1. The phones were not to scale and the image was inaccurate in a way that was unfavorable towards the non-iPhone phablet.
Evidence please？
That's a bit like debating whether a gold miner is proficient at alchemy -- if the primary endeavor was actually successful, it would preclude needing to engage in the secondary.
If you had higher-quality generative AI, wouldn't writing marketing tweets, instead of paying for them, be the first thing you used it for?
Here's what happened to anything even decent at Microsoft: some shitty nakedly sociopathic MBA comes in and starts picking it apart for blatantly anticompetitive or monopoly protecting concerns, and the tech people behind it lose all morale and say "fuck it, it's a day job".
Here's what happens to anything at Google: some shitty "product manager" (toootally different that the microsoft monopoly MBAs. Totally.) comes in and says "how do we sell more ads with this". And... well, yeah same thing happens.
""" ChatGPT in Trouble: Bard's Arrival Sends Shockwaves Through the AI Community! ;) """
What you find fun is that Bard gives suggestions that are much less catchy than ChatGPT: So you still need ChatGPT to write trash about ChatGPT
In all my tests so far, this integration fails. The Bing AI worked every time. (Can't believe I just wrote those words, but here we are...)
How do they guarantee hundreds or thousands of likes? Presumably the only reason the algorithm keeps showing them to me is that they’re popular. Otherwise they’d be labeled “sponsored”.
The points are always so bad, too. I asked Bard to summarize one of my repos, and it hallucinated that it was an http server written in Go. My repo didn’t have a single .go file. “Internet-connected AI” is just mistaken.
As to how - many ways, some of them really creative. One that I remember was an account that's reposting stuff that I genuinely liked, some left-wing activist content, and only now and then they would push something different, like a message about how this one time brand X turned out to be crap during a manifa.
[1] https://twitter.com/jeroenvlek/status/1657195209686683648
So neither could generate the plugin, it's just that one of them lied about it.
It was interesting to see that ChatGPT annotated its code with pitfalls to avoid, that it actually didn't avoid in the code.
Why? I mean, why does it sound more plausible to you that it ever could, when we have countless examples of models lying (or rather hallucinating)?
The advanced prompt in question = "create a table from these 3 numbers"
I wouldn’t be surprised that if you looked at the details on those tweets, it’s the same few thousand dudes creating them and liking them and retweeting them.
Couldn't that be said about search as well nowadays? and yet here we are.
C&P "Google just updated its free competitor, Bard."
C&P "ChatGPT has now a big problem"
C&P "Here are 7 incredible things Bard can do (that ChatGPT cannot)"
C&P "ChatGPT is in trouble"
exploding head emoji
I don't think you understand engagement farming. The only objective these hustlers have is to gain more followers, and they will say anything about anything as long as it leads to them standing out in your TL and thus increasing their chance of a follow (and of course, eventually monetization)
Last year they were opining about Web3. In December they were the ones screaming "Google is Finished", then that became generic so now there's a new angle. In 2 months it'll change again. Their words are completely meaningless and bare little relation to reality. The most important thing is to establish the poster as some kind of authority figure who always knows what's going on, so please follow them.
There's clearly a Big Tech turf war going on over AI. All sides are slinging poop/propaganda. I've seen more hysterical and uninteresting anti-Google FUD than ever on this here Orange Site over the YTD. Special mention of the "world is literally ending because Google requested .zip TLD from ICANN" or pre-IO anti-hype "Google CEO earns as much as other highly-paid CEOs".
Superfluous spam that I wish would be filtered, but often times it's produced specifically to get upvoted on HN, which is a dilemma for moderation.
Funny to think that somewhere out there I'm loaded into someone's consciousness as a pawn in the marketing machinery of The Alphabet.
The fast pace of ‘internet time’ means Amazon, AOL, Facebook and Google can follow the trajectory of Ford, AMC, Chrysler, and GM in a fraction of the time. Amazon was a leader in e-commerce once, now they count on someone who subscribed to Prime for the last 15 years to subscribe for the next 15 years. Google search used to be good, but now it is corrupt. Search for a ‘used Honda Fit’ and the first search result is an ad from the one car dealer that owns all the new car dealerships in town for new cars and even though they presumably have used cars in their lot, they’ll try to sell you the same CR-V they’d try to sell you in person. I’d accept any kind of ad from someone who is selling what i am buying but Google aids and abets numerous acts of brand destruction every day.
They know how to walk the line of running a corrupt ads engine and not get anybody arrested but it is an unsolved problem in A.I. ethics to get away with running a corrupt chatbot —- and that is why Google is in so much trouble.
Your query: "Google, show me the best place to eat in manhattan"
"Bard's partners at Manhattan steakhouse are known for their delicious chargrilled steaks. Other options include..."
Sure, Google won't do that today, but they didn't do ads at the top of search for a while... until they did.
Google has a massive valuation because whenever Wall Street demands “M0R M0NEY!” they can just increase the ad load. The thing is they and their investors depend on every penny of ad revenue they get and any innovation that reduces it, even just a little bit, is an absolute disaster for them. That’s why it is a ‘disruptive’ technology to Google…. It eats their lunch right now and the possibility they might be able increase their ad load someday doesn’t make up for losing it. I mean, shareholders can hit the SELL button and run in as much time as it takes me to write this sentence.
did i miss something and they are no longer half a trillion revenue top ecommerce company in the world? who did they get replaced by?
Items you can buy off the shelf at Target for $40 are for sale for $70 at AMZN.
Some people get one day delivery, but other people (me) get five day delivery with Prime, particularly ZIP codes that have an Amazon distribution facility in them. (Everyone I know who gets 1 day delivery has a Cheesecake Factory in their town... But for Christ's sake I order things from Japan on Ebay and get them before I get things from the AMZN warehouse in the next state.) I guess they think I want to watch Rings of Power so bad I'll subscribe to Prime anyway but you know... I don't.
My understanding is that the online store consistently loses money, they are a going concern only because of AWS. But guess what, I've already moved enough spend from AWS to Azure that subjecting me to 5 day delivery cost them much more than it saved.
As for replacements, see every other retailer that sees fulfillment as a way to make a positive impression. Adorama, Best Buy, and Target just to name a few. It's true that many of those have bizzare limited product line ideas (e.g. Best Buy sells Sony camera bodies but has no real selection of lenses but they think I'm stupid enough to buy a Samsung washing machine which is not a good choice unless you want to buy another washing machine in a few weeks) but the product listings are basically honest outside AMZN.
Long term see Temu and other asian retailers. When I look at arXiv I see a large literature in "sequential reccomendation" almost entirely by Chinese and Indian authors. Looks like a "missile gap" for e-commerce and social commerce if you ask me.
It’s like listening to a mad scientist connect totally unrelated (seemingly) events.
E.g., I don’t at all understand the linkage between current events and the schism between the Catholic Church and Protestants denoms.
Maybe I’m just too dumb to follow this much brilliance.
His point is that LLMs are a disruptive technology at the level of the printing press. The power structure most famously disrupted by the printing press was that of the Catholic Church.
EDIT: Yeah this isn't what he's saying. His analogy is more like the major LLM companies like OpenAI or Google are like the Catholic Church which is being disrupted by European regulations which are like the protestants. Or the opposite. He literally says both ways as possibilities. And in this analogy the open models (the refinements of the 'leaked' LLaMa models which can fit on a single GPU) are like the wacky persecuted fringe sects which were exiled to the United States and which he thinks shouldn't be called as either Catholic or Protestant. At the end he says "This is, admittedly, a rather speculative and far-reaching Article" which I have to agree.
The American companies like Google/OpenAI (trying to export AI globally) are on one side, opposing them are Europeans (who want AI to be controlled, licensed, and regulated). Together, these two regimes are the Catholic Church and mainstream Protestantism of AI.
Opposing both these sides are open source AI models. He calls this group radical reformers like the Amish.
Yeah, don’t ask. This much brilliance in way above my pay grade.
Very typical of the author. Except few valuable ideas presented years ago, I haven't read or heard anything worthwhile. Mostly unsubstantiated speculation and “I've seen the big picture” type of ramblings.
Why a mad scientist and not a mad junkie? Or just some clueless language model, artificial or not? I think we tend to make this assumption that things that seem complicated _are_ complicated, and not just random bs, and that assumption no longer holds.
>E.g., I don’t at all understand the linkage between current events and the schism between the Catholic Church and Protestants denoms.
Hadn't read it, but let me guess, is it about how Gutenberg changed the society's interconnection fabric, and it triggered Thirty Year's War, and the same fabric remodeling thing is happening right now? :)
Yes, there is some reference to the evolution of the principle of national sovereignty which came from the disruption of the Catholic Church and how that ties to Europe trying to regulate AI advances coming out of America.
Like you said, he could just as well be an organic LLM.
That's just an attempt to prevent technical colonisation. An analogy would be Japan and China trying to fence off missionaries. Protestantism was about something different: decentralisation. And in that regard there's no difference between Google and their competitors - it's all strictly centralised; nothing Protestant about it.
Having properly Open Source models would be protestant. And EU regulations should help it happen by reducing the incentives for centralised ones.
Because online advertising, particularly search ads, are working less well for advertisers and regulators are pushing for more changes. The vast majority of Google revenue comes from an industry that is facing massive risks.
On top of that I don't think it's clear that "AI" in search will help Google sell advertising or increase revenue. It's also likely to bring them under even more scrutiny from regulators.
In many ways AI is harder to sell ads against, Google is payed when people leave the page via a click on an ad. That doesn't happen when Google is summarising the content. Microsoft/Bing have found a brilliant way to attack Google, build a product that they have to copy that (potentially) decimates ad revenue.
Didn't 0-click search as a trend for google search predate bing's changes?
Not really. The click rate is iirc usually around 5%; I'm not convinced this is much higher than the "noise floor" generated by a human equivalent of drunk racoons. So in practice Google is being paid for displays.
However the judge is still out in regards to how much can AI allow Google to squeeze the profit margins of businesses by offering even better conversion rates. Personally, I don't think it will make a big difference, the AI will not make you buy more. I suspect Google makes more money by having multiple companies fight for the same customer than actually making a conversion.
IMO best case scenario for Google is to keep their search market. Table stakes is keeping search relevant at all.
The inverse is true - people continue using Google products despite being aware that they are spying on them. Doing it outright won't make them suddenly start caring.
It’s what the whole house of cards is built on.
Also for queries asking for product recommendation, where ad price is quite high, they could still plug ads into an LLM and have it write a summary back to you similar to what magazines do with sponsored content.
Main part of Google's business is selling infrastructure for that SEO junk, AdWords et al. Sure, you meant the _other_ SEO junk, but from AI point of view both would be noise to be separated from actual content.
Oh, you optimist!
It's ALWAYS about incentives. Aka money.
Companies will use AI to squeeze garbage (aka ads) in every nook and cranny of the observable universe now. Think of all the thankless tasks that required a human to do them until now.
Guess what? Now that AI can do them 100000x faster and probably 2-5x better.
Our lives are about to be flooded by a stream of -pardon my French-, adshit.
In the context of this thread, I don't think Google could sell more ads with an LLM which is why they sat on this technology. It disrupts their current business, while being more expensive and harder to monetize (charging a subscription to consumer is worse than a monopolized ad business charging hundreds of millions to corporations).
100% Yes. We can (because people are going to ask how to do X in postgres and AI can reply how to do that in postgres with a side-note that this Sponsored product does it out of the box). however, can you sell more ads (by tagging them as "ads" or "Sponsored") - That is where the gray area shows up (IMO).
Obviously, this is still evolving but I guess something along that lines could be done, if one is too serious to monetize it. In my personal opinion, I think right now companies are focussed around capturing market than monetization.
Google has maximized their ad revenue with their current model. You don't have the same ad real estate in your example and that is a problem. Now they get money for impressions on a whole page, and clicks from people trying to find the right fit among the seo. I believe that google wants information filtering but not too much as it needs current estate for ads.
Companies like Google and Meta have been relying on AI to display ads for years now. It’s been central to their strategy.
ChatGPT was just a shiny toy. A very cool, shiny toy, but a toy nonetheless.
OpenAI just implemented something (well) that was developed in other places.
LLMs are indeed very cool and maybe even revolutionary, but their main accomplishment was beating big tech to the punch.
I suspect that those who don't care about privacy or can't afford to pay, will be swarmed with AI-seo results from an ad-supported model, but people will FINALLLY have the choice to pay and avoid this ad-fest. And who knows, maybe when open-source catches up, it will even be affordable enough that ad-tech will shrink considerably.
Yep. Google didn’t want a chatbot telling kids how to cook meth, so they weren’t in a hurry to release. Meanwhile, releasing a chatbot is one of the few cards a much smaller company like OpenAI can play.
What's the last time Google launched a product that everyone really uses?
Also Google Maps, Google Photos and Google Drive have more than 1B users.
[1]: https://appleinsider.com/articles/21/01/27/there-are-1b-ipho...
as for your link. iphones are popular, but it's definitely not true that "everyone" has an iphone, which is my point. android has almost 4 billion active users, for comparison.
History would say otherwise.
Look at how quickly Nokia went south.
Look at how quickly Google took over from the AltaVista search engine.
On the other hand AltaVista was on the early days of the internet where everything was much more turbulent.
Hard to fund loss leaders like Android with say 20% less revenue.
But one thing that is obvious is that having 20% less revenue would not mean Google was unable to work on Android. We know that, because they were able to already fund Android in the distant past of 2021, when their revenue was 20% lower than now. (In fact, they were able to afford Android in 2010, with 90% lower revenue.)
And in terms of ethics, the perception of them being a force for good eroded a lot. When thinking about the typical googler, people around here will sooner think about an exec trying to skip some tax than about a techie doing innovative stuff, because that's what their media coverage is about these days in Europe.
I'll give you that Microsoft is very unlikely to go away either, unless the US govt finally remembers they have antitrust laws.
Go ask 10 non-tech people what they think of Google.
Actually as a factoid, I recently changed jobs, and got handed an iphone for work. People noticed it and commented on it, whereas a pixel would probably have not raised much attention.
Makes one wonder if it's actually representative of the whole gen Z, or just some specific western subset.
It's hard for a software company to compete against the largest ad network ever created.
Despite Bing having a head start, it's unlikely that Google will lose to AI battle to Bing and others. Google's strong brand image and popular properties, such as Gmail, YouTube, Search, Chrome, and Android, used by billions of users daily, give them a significant advantage. It's doubtful that Bing will take over the search market anytime soon. But, who knows? Ha!
Content mills existed long before AI. Clickbait existed long before AI. Fake news and misinformation existed long before AI.
AI will empower things like that, but it isn't at the core of the problem.
As a user, I care that content is truthful, useful, and/or entertaining. I care that it isn't presented in a misleading way, and I care that it doesn't cause harm (hate, misinformation, etc.) - What I don't care about is who or what created it.
I think a bigger step in the right direction for society would be to empower certain groups like journalists in ways that they don't have to compete with all the crap out there.
Furthermore, it seems unenforceable. As AI becomes more sophisticated (if this isn't the case already), it will be virtually impossible to prove mislabelling.
Where is the line?
0: https://support.google.com/mail/answer/9116836?hl=en&co=GENI...
I have written a number of documents recently that are AI assisted but definitely my own work. I use the LLM to help me cross reference topics, clean up some of the language, improve the flow and prepare for the potential follow-up discussions but the arguments and recommendations are still mine. Is this AI or not?
Side note: I just prepared a recommendation for an org change and had the AI argue against it, as well as provide responses. Some were good and some were weak but it was extremely useful and quite fun.
This time, I figured I'll use GPT-4 to help me write the perfect text. I described the details of the event, my relationship to the invitee (inb4: without any kind of PII, of course), the style, tone and context of the desired text, and asked GPT-4 to generate suggestions. I went back and forth with it, asking it to generate more variants of a specific suggestions, then taught it a simple markup for editing and asked to iterate on specific words and sentences, until I was somewhat satisfied with the result.
Then I asked my wife for her idea, and she quickly wrote a little text of her own. Only afterwards, I showed her the best (to me) of GPT-4 texts. We then mixed the sentences from both together, creating a final work that's 50% OG, perfectly clean text written by my wife, and 50% the output of GPT-4 (with me guiding it).
Is that text a work of AI? Or of a human? Or both? Is it even 50% AI and 50% human, given that the AI part were created from my input, and then edited by my telling GPT-4 what to change, and then finally approved by me using my own judgement?
Does a random website sharing invitation templates on-line have a copyright case, if a couple sentences from our text matched something of theirs? What if it wasn't the AI part that matched?
So no, I don't believe there will be regulation on labelling "AI" derived content, but there will be a drive to label and certify "organic content".
People who boast that their company is <insert identity here> owned do that, because their product can't compete otherwise.
Microsoft has little consumer push (no mobile presence) hence they will likely dominate in the business market. Integrating it into Office, private training on company data, etc.
Apple is the big mystery. Supposedly the big consumer-facing competitor to Google but very little is known about what's coming from them. Honestly, I'm quite convinced they're in panic mode as well, they just hide it better.
Irrespective, doesn't matter how the numbers came to be, they're massive opportunities to integrate AI into.
It is only going to accelerate and AI research will increasingly be more closed. There is now no difference between O̶p̶e̶n̶AI.com or Google DeepMind.
Either way it will have to take more than LLMs like 'ChatGPT' to even come close to dethroning Google. [0] There is more to 'AI' than the LLM chatbot hype.
[0] https://www.similarweb.com/blog/insights/ai-news/chatgpt-gro...
I do think all government should require AI to disclose all the material ingested.
You asked for advantages that Google has and the text of 25 million books is one thing. A lot of the text is going to be scarce and so it's something of value Google has and others do not.
Google also has built a massive GIS database from driving their cars around the world. Apple has a smaller database and there are some third parties (like ESRI) that will sell you access to their database. If you want to ask an AI something about your neighborhood, Google has more data than anybody on that front.
Zero to none?
I mean, it's much much much easier for collaborators to produce better (or even competitive) map data than training LLMs, and yet the open alternatives to google maps are hardly ever used, have spotty data for some locations and have made no compelling feature for user's to adopt.
My understanding is that these models are trained using some form of gradient descent. This requires knowledge of the model’s prior state and current training error to adjust parameter values.
With distributed training it’s difficult, to point of impossible, to communicate the state of the model across large distances because these models are so large.
Take my comment with a grain of salt because I’m not an expert in this area. I’d very much enjoy being wrong about this because distributed training would unlock so much potential.
The advantage is that their AI will be integrated into products that already have billions of users. Android users will get used to the default google assistant model so theyre more likely to use it in other instances as well and so on.
Given that Google still gets the vast majority of its revenue from search its hard to see that this doesn’t cause then a huge amount of pain. I suspect they’ve made the mistake of not starting to cannibalize their revenue stream and now it might be too late.
It’s going to be interesting to see how this plays out.
ChatGPT and other modern AI tools are useful specifically because they do what I want when I want to do it. Compare this to Gmail's smart compose: by the time the AI makes a suggestion I'm already half-way through typing that text, so its prediction is completely useless even if it's correct.
(Also, I really enjoy the random historical rambling at the end of some Stratechery articles).
And the safety regulations which made our products less toxic and less shoking
Or in the EU, roaming charges removal, flight reimbursement regulations, etc.
Heck, speaking of flying, the entire airline industry is regulated up the wazoo and a few years ago we had a full year of metal coffins flying through the skies at 10 000m going 950 kmph without any fatalities.
You don't get it! Only big companies can make safe cars cheaply, so it was all a big ploy against small car manufacturers.
I'm being sarcastic in case someone missed it :-)
It’s going to bite them when they need developers to embrace their ai ecosystem.
Their CEO should have been fired years ago.
It could be the case that Pichai is Immelt while Schmidt is the Google Jack Welch.
I find this quote a little bit hard to believe. Are the current models genuinely capable of innovation? I've been productive in using image generation for my album covers, but I haven't found a similar use for Large Language Models yet.