Google is winning on every AI front
thealgorithmicbridge.com
thealgorithmicbridge.com
[1]https://blog.google/products/google-cloud/ironwood-tpu-age-o...
Modern BERT with the extended context has solved natural language web search. I mean it as no exaggeration that _everything_ google does for search is now obsolete. The only reason why google search isn't dead yet is that it takes a while to index all web paged into a vector database.
And yet it wasn't google that released the architecture update, it was hugging face as a summer collaboration between a dozen people. Google's version came out in 2018 and languished for a decade because it would destroy their business model.
Google is too risk averse to do anything, but completely doomed if they don't cannibalize their cash cow product. Web search is no longer a crown jewel, but plumbing that answering services, like perplexity, need. I don't see google being able to pull off an iPhone moment where they killed the iPod to win the next 20 years.
The web UI for people using search may be obsolete, but search is hot, all AIs need it, both web and local. It's because models don't have recent information in them and are unable to reliably quote from memory.
I see search engines as a dripfeed from a firehose, not some magical thing that's going to get me the 100% correct 100% accurate result.
Humans are the most prolific liars; I could never trust search results anyway since Google may find something that looks right but the author may be heavily biased, uninformed and all manner of other things anyways.
Crawling the web has a huge moat because a huge number of sites have blocked 'abusive' crawlers except Google and possibly Bing.
For example just try to crawl sites like Reddit and see how long before you're blocked and get a "please pay us for our data" message.
95% of our load is from crawlers, so we have to pick who to serve.
If they want our data all they need to do is offer a way for us to send it, we're happy to increase exposure and shopping aggregation site updates are our second highest priority task after price and availability updates.
1. Search ads (at risk of disintermediation) 2. Display ads (not going anywhere) 3. Ad-supported YouTube 4. Ad-supported YouTube TV 5. Ad-supported Maps 6. Partnership/Ad supported Travel, YouTube, News, Shopping (and probably several more) 7. Hardware (ChromeOS licensing, Android, Pixel, Nest) 8. Cloud
There are probably more ad-supported or ad-enhanced properties, but what's been shifting over the past few years is the focus on subscription-supported products:
1. YouTube TV 2. YouTube Premium 3. GoogleOne (initially for storage, but now also for advanced AI access) 4. Nest Aware 5. Android Play Store 6. Google Fi 7. Workspace (and affiliated products)
In terms of search, we're already seeing a renaissance of new options, most of which are AI-powered or enhanced, like basic LLM interfaces (ChatGPT, Gemini, etc), or fundamentally improved products like Perplexity & Kagi. But Google has a broad and deep moat relative to any direct competitors. Its existential risk factors are mostly regulation/legal challenge and specific product competition, but not everything on all fronts all at once.
Google's cash-cow product is relevant ads. You can display relevant ads in LLM output or natural language web-search. As long as people are interacting with a Google property, I really don't think it matters what that product is, as long as there are ad views. Also:
> Web search is no longer a crown jewel, but plumbing that answering services, like perplexity, need
This sounds like a gigantic competitive advantage if you're selling AI-based products. You don't have to give everyone access to the good search via API, just your inhouse AI generator.
Bryce Bayer worked for Kodak when he invented and patented the Bayer pattern filter used in essentially every colour image sensor to this day.
But the problem was: Kodak had a big film business - with a lot of film factories, a lot of employees, a lot of executives, and a lot of recurring revenue. And jumping into digital with both feet would have threatened all that.
So they didn't capitalise on their early lead - and now they're bankrupt, reduced to licensing their brand to third-party battery makers.
> You can display relevant ads in LLM output or natural language web-search.
Maybe. But the LLM costs a lot more per response.
Making half a cent is very profitable if you only take 0.2s of CPU to do it. Making half a cent with 30 seconds multiple GPUs, consuming 1000W of power... isn't.
I do think Google is a little different to Kodak however; their scale and influence is on another level. GSuite, Cloud, YouTube and Android are pretty huge diversifications from Search in my mind even if Search is still the money maker...
People like to believe CEOs aren't worth their pay package, and sometimes they're not. But a look at a couple of their failures and a different CEO of Kodak wouldn't have had what happened happen, makes me think that sometimes, some of them do deserve that.
When a fool inevitably takes the throne, disaster ensues.
I can't say for sure that a different system of government would have saved Kodak. But when one man's choices result in disaster for a massive organization, I don't blame the man. I blame the structure that laid the power to make such a mistake on his shoulders.
The CEO takes the blame, the board picks a new one (Unless the CEO has special shares that make them impossible to dismiss), and we go on hoping that the king isn't an idiot this time.
My reading of history is that some people are fools - we can blame them for their incompetence or we can set out to build foolproof systems. (Obviously, nothing will be truly foolproof. But we can build systems that are robust against a minority of the population being fools/defectors.)
Constantly I see them dodging responsibility or resigning (as an "apology") during a crisis they caused and then moving on to the next place they got buddies at for another multi-mil salary.
Many here would defend 'em tho. HN/SV tech people seem to aspire to such things from what I've seen. The rest of us just really think computers are super cool.
Even on the few Vaios that had MD drives on them, they're pretty much just an external MD player permanently glued to the device instead of being a full and deeply integrated PC component.
As a business Google's interest is in showing ads that make it the most money - if they quickly show just the relevant information then Google loses advertising opportunities.
To an extent, it is the web equivalent of irl super markets intentionally moving stuff around and having checkout displays.
This is just a question of UX- the purpose of their search engine was already to show the most relevant information (ie. links), but they just put some semi-relevant information (ie. sponsored links) first, and make a fortune. They can just do the same with AI results.
Today a consumer grade >8b decoder only model does a better job of predicting if some (long) string of text matches a user query than any bespoke algorithm would.
The only reason why encoder only models are better than decoder only models is that you can cache the results against the corpus ahead of time.
I doubt this. Embedding models are no panacea even with a lot simpler retrieval tasks like RAG.
Unlike the natural language queries that RAG has to deal with, Google searches are (usually) atomic ideas and encoder-only models have a much easier time with them.
I've been wondering for some time what sustainable advantage will end up looking like in AI. The only obvious thing is that whoever invents an AI that can remember who you are and every conversation it's had with you -- that will be a sticky product.
I've build RAG systems that index tokens in the 1e12 range and the main thing stopping us from having a super search that will make google look like the library card catalogue is the copyright system.
A country that ignores that and builds the first XXX billion parameter encoder only model will do for knowledge work what the high pressure steam engine did for muscle work.
I think they've jumped the shark and need to give me more control because currently I actively avoid watching videos I think MIGHT be interesting because the risk is too high. This is a terrible position to put your users in both from a specific experience perspective but also in a "how they feel about your product" perspective.
You can remove videos from your viewing history. I do this when I start watching something but the content turnes out to not be what I expected. It seems to prevent polluting my recommendations.
The improvement has been steady and impressive. The entire integration is becoming a product that I want to use.
> I’m forgetting something. Oh, of course, Google is also a hardware company. With its left arm, Google is fighting Nvidia in the AI chip market (both to eliminate its former GPU dependence and to eventually sell its chips to other companies). How well are they doing? They just announced the 7th version of their TPU, Ironwood. The specifications are impressive. It’s a chip made for the AI era of inference, just like Nvidia Blackwell
"It was there."
Everyone understands that such naysaying is effectively an accusation of lying. In any case it was a totally low effort utterly inappropriate comment. Clearly you aren't going to learn the lesson.
Google has their own cloud with their data centers with their own custom designed hardware using their own machine learning software stack running their in-house designed neural networks.
The only thing Google is missing is designing a computer memory that is specifically tailored for machine learning. Something like processing in memory.
Google is catching up fast on product though.
Once the space settles down, the balance might tip towards specialized accelerators but NVIDIA has plenty of room to make specialized silicon and cut prices too. Google has still to prove that the TPU investment is worth it.
They started becoming available internally in mid 2015.
While Nv does have an unlimited money printer at the moment, the fact that at least some potential future competition exists does represent a threat to that.
So are the electric and cooling costs at Google's scale. Improving perf-per-watt efficiency can pay for itself. The fact that they keep iterating on it suggests it's not a negative-return exercise.
You're talking about small-money bets. The technical infrastructure group at Google makes a lot of them, to explore options or hedge risks, but they only scale the things that make financial sense. They aren't dumb people after all.
The TPU was a small-money bet for quite a few years until this latest AI boom.
The cost delta was massive and really quite astounding to see spelled out because it was hardly talked about internally even after the paper was written. And if you took into account the very high comp Google engineers got, even back then when it was lower than today, the delta became comic. If Gmail had been a normal business it'd have been outcompeted on price and gone broke instantly, the cost disadvantage was so huge.
The people who built Gmail were far from dumb but they just weren't being measured on cost efficiency at all. The same issues could be seen at all levels of the Google stack at that time. For instance, one reason for Gmail's cost problem was that the underlying shared storage systems like replicated BigTables were very expensive compared to more ordinary SANs. And Google's insistence on being able to take clusters offline at will with very little notice required a higher replication factor than a normal company would have used. There were certainly benefits in terms of rapid iteration on advanced datacenter tech, but did every product really need such advanced datacenters to begin with? Probably not. The products I worked on didn't seem to.
Occasionally we'd get a reality check when acquiring companies and discovering they ran competitive products on what was for Google an unimaginably thrifty budget.
So Google was certainly willing to scale things up that only made financial sense if you were in an environment totally unconstrained by normal budgets. Perhaps the hardware divisions operate differently, but it was true of the software side at least.
Also worth noting that its Ads division is the largest, heaviest user of TPU. Thanks to it, it can flex running a bunch of different expensive models that you cannot realistically afford with GPU. The revenue delta from this is more than enough to pay off the entire investment history for TPU.
> The revenue delta from this is more than enough to pay off the entire investment history for TPU.
Possibly; such statements were common when I was there too but digging in would often reveal that the numbers being used for what things cost, or how revenue was being allocated, were kind of ad hoc and semi-fictional. It doesn't matter as long as the company itself makes money, but I heard a lot of very odd accounting when I was there. Doubtful that changed in the years since.
Regardless the question is not whether some ads launches can pay for the TPUs, the question is whether it'd have worked out cheaper in the end to just buy lots of GPUs. Answering that would require a lot of data that's certainly considered very sensitive, and makes some assumptions about whether Google could have negotiated private deals etc.
I'm not sure what you're trying to deliver here. Following your logic, even if you have a fab you need to compete for rare metals, ASML etc etc... That's a logic built for nothing but its own sake. In the real world, it is much easier to compete outside Nvidia's own allocation as you get rid of the critical bottleneck. And Nvidia has all the incentives to control the supply to maximize its own profit, not to meet the demands.
> Possibly; such statements were common when I was there too but digging in would often reveal that the numbers being used for what things cost, or how revenue was being allocated, were kind of ad hoc and semi-fictional.
> Regardless the question is not whether some ads launches can pay for the TPUs, the question is whether it'd have worked out cheaper in the end to just buy lots of GPUs.
Of course everyone can build their own narratives in favor of their launch, but I've been involved in some of those ads quality launches and can say pretty confidently that most of those launches would not be launchable without TPU at all. This was especially true in the early days of TPU as the supply of GPU for datacenter was extremely limited and immature.
More GPU can solve? Companies are talking about 100k~200k of H100 as a massive cluster and Google already has much larger TPU clusters with computation capability in a different order of magnitudes. The problem is, you cannot simply buy more computation even if you have lots of money. I've been pretty clear about how relying on Nvidia's supply could be a critical limiting factor in a strategic point of view but you're trying to move the point. Please don't.
But I am not sure how AWS and Google Cloud match up in terms of making this verticial integration work for their competitive advantage.
Any insight there - would be curious to read up on.
I guess Microsoft for that matter also has been investing -- we heard about the latest quantum breakthrough that was reported as creating a fundamenatally new physical state of matter. Not sure if they also have some traction with GPUs and others with more immediate applications.
Now for the life of me, I still haven't been able to understan what a TPU is. Is it Google's marketing term for a GPU? Or is it something different entirely?
It's a chip (and associated hardware) that can do linear algebra operations really fast. XLA and TPUs were co-designed, so as long as what you are doing is expressible in XLA's HLO language (https://openxla.org/xla/operation_semantics), the TPU can run it, and in many cases run it very efficiently. TPUs have different scaling properties than GPUs (think sparser but much larger communication), no graphics hardware inside them (no shader hardware, no raytracing hardware, etc), and a different control flow regime ("single-threaded" with very-wide SIMD primitives, as opposed to massively-multithreaded GPUs).
Edit: And btw, another question that I had had before was what's the difference between a tensor core and a GPU, and based on your answer, my speculative answer to that would be that the tensor core is the part inside the GPU that actually does the matmuls.
It's not a GPU, as there is no graphics hardware there anymore. Just memory and very efficient cores, capable of doing massively parallel matmuls on the memory. The instruction set is tiny, basically only capable of doing transformer operations fast.
Today, I'm not sure how much graphics an A100 GPU still can do. But I guess the answer is "too much"?
So GPUs have ~120 small systolic arrays, one per SM (aka, a tensorcore), plus passable off-chip bandwidth (aka 16 lines of PCI).
Where has TPUs have one honking big systolic array, plus large amounts of off-chip bandwidth.
This roughly translates to GPUs being better if you're doing a bunch of different small-ish things in parallel, but TPUs are better if you're doing lots of large matrix multiplies.
But I will admit, Gemini Pro 2.5 is a legit good model. So, hats off for that.
This makes it rather unusable as a catch all goto resource, sadly. People are curious by nature. Refusing to answer their questions doesn't squash that, it leads them to potentially less trustworthy sources.
But that's good
The AI won't tell the reader what to think in an authoritative voice. This is better than the AI trying to decide what is true and what isn't.
However, the AI should be able to search the web and present it's findings without refusals. Obviously, always presenting the sources. And the AI should never use an authoritative tone and it should be transparent about the steps it took to gather the information, and present the sites and tracks it didn't follow.
Claude to circumvent Eastern censorship
Grok Unhinged for a wild time
When did this start? Serious question. Of all the model providers my experience with Google's LLMs and Chatproducts were the worst in that dimension. Black Nazis, Eating stones, pizza with glue, etc I suppose we've all been there.
Then you could look at how the first "public preview" models they released were so neutered by their own inhibitions they were useless (to me). Things like over-active refusals in response to "killing child processes".
They've got the cash, the people, and the infrastructure to do things faster than the others going forward, which is a much bigger deal IMO than having millions more users right now. Most people still aren't using LLMs that often, switching is easy, and Google has the most obvious entry points with billion+ users with google.com, YouTube, gmail, chrome, android, etc.
Google can be good on the technological side of things, but we saw time and time again that, other than ads, Google is just not good at business.
Who is? (Genuine question, it's hard to keep up given how quickly the field moves.)
Not sure about multimodal as it’s not what I work on.
Are you pretty sure that Google won't eventually buy OpenAI and thus learn everything you've said to ChatGPT?
I definitely don't trust Google -- fool me once, and all -- but to the extent I'm going to "trust" any business with my data, I'd like to see a proven business model that isn't based on monetizing my information and is likely to continue to work (e.g., Apple). OpenAI doesn't have that.
What can OpenAI do? They can sell my data, whatever, it’s a whole bunch of prompts of me asking for function and API syntax.
In either case, I'm sure that's how it starts. "This company has very little power and influence; what damage can they do?"
Until, oh so suddenly, they're tracking and profiling you and selling that data.
Also, it’s less about what they currently do but what they’re capable of. A Cold War of privacy of sorts.
Google’s main source of income, by far, is selling ads. Not just any ads but highly targeted ones, which means global digital surveillance is an essential part of their business model.
It seems their revenue in 2024 exceeded $3B.
> they will sell your data to make ends meet
I’m not sure they can do that without breaching the contract. My employer pays for ChatGPT enterprise I use.
Another thing, OpenAI has very small amount of my data because they only have the stuff I entered to their web service. Google on the other hand tracks people across half of the internets, because half of the web pages contain ads served by google. Too bad antimonopoly regulators were asleep on their job when google acquired DoubleClick, AdMob, and the rest of them.
With a loss of $5B. A viable business model needs more than revenue. It also needs profit.
It is not unusual for a business with visions for profitability to accept losses for a while to get there, but OpenAI does not seem to have such vision. They seem to be working off the old tech model of "If we get enough users we'll eventually figure something out" – which every other time we've heard that has ended up meaning selling user data.
Maybe this time will be different, but every time we hear that...
Cult of personality is blinding. But I could be wrong in my interpretation. Would you be able to put what that moral standing is about in prose without engaging in public figure examples?
Large corporations wind up creating internal policies, controls, etc. If you know anyone who works in engineering at Google, you'll find out about the privacy and security reviews required in launching code.
Startups, on the other hand, are the wild west. One policy one day, another the next, engineers are doing things that don't follow either policy, the CEO is selling data, and then they run out of money and sell all the data to god knows who.
Google is pretty stable. OpenAI, on the other hand, has been mega-drama you could make a movie out of. Who knows what it's going to be doing with data two or four years from now?
same pattern as Mark Zuckerberg's movie.
So far this has been nothing but a PM wankfest but if Gemini-in-{Gmail,Meet,Docs,etc} actually gets useful, it could be a big deal.
I also don't think any of those concerns are as important for API users as direct consumers. I think that's gonna be a bugger part of my the market as time goes on.
ChatGPT has a nice consumer product, and I also like it.
Google gets a bad rap on privacy, etc., but if you read the documentation and set privacy settings, etc. then I find them reasonable. (I read OpenAI’s privacy docs for a long while before experimenting with their integration of Mac terminal, VSCode, and IntelliJ products.)
We live in a cornucopia of AI tools. Occasionally I will just for the hell of it do all my research work for several days just using open models running on my Mac using Ollama - I notice a slight hit in productivity, but still a good setup.
Something for everyone!
No one is going to build on top of anything "Google" without having a way out thought out in advance.
Not that important for LLMs, where drop-in replacements are usually available. But a lot of people just hear "by Google" now and think "thanks I'll pass" - and who can blame them?
Maybe Gemini is finally better, but I'm not exactly excited to give it a try.
Please do.
To me, he is a finance bro grifter who lucked into his current position. Without Ilya he would still be peddling WorldCoin.
>Humanism is a philosophical stance that emphasizes the individual and social potential, and agency of human beings, whom it considers the starting point for serious moral and philosophical inquiry. (wikipedia)
Whereas the other lot are often engineers / compsci / business people building stuff.
Which can be said for most of the survivorship-biased "greats" we talk about. Right time, right place.
(Although to be fair — and we can think of the Two Steves, or Bill and Paul — there are often a number of people at the right time and right place — so somehow the few we still talk about knew to take advantage of that right time and right place.)
yet they can never seem to explain why those succesful people all seem to have similar traits in terms of work ethic and intelligence
you'd think there would be a bunch of lazy slackers making it big in tech but alas
as an Asian, it amazes me how far Americans and Europeans will go to avoid a hard days work
same audience who think Jobs is a grifter and Woz is the true reason for Apple's success
Key Facts from "The Secrets and Misdirection Behind Sam Altman's Firing from OpenAI": https://www.lesswrong.com/posts/25EgRNWcY6PM3fWZh/openai-12-...
I was actually surprised at Google's willingness to offer Gemini 2.5 Pro via AI Studio for free; having this was a significant contributor to my decision to cancel my OpenAI subscription.
Microsoft gained control in the '90s by bundling Internet Explorer with Windows for free, undercutting Netscape’s browser. This leveraged Windows’ dominance to make Explorer the default choice, sidelining competitors and capturing the browser market. By 1998, Netscape’s share plummeted, and Microsoft controlled access to the web.
Free isn’t generous—it’s strategic. Google’s hooking you into their ecosystem, betting you’ll build on their tools and stay. It feels like a deal, but it’s a moat. They’re not selling the model; they’re buying your loyalty.
It almost sounds like you're saying that Netscape wasn't free, and I'm pretty sure it was always free, before and after Microsoft Explorer
https://www.nytimes.com/1996/08/19/business/netscape-moves-t...
Bundling a "good enough" products can do a lot, including take you from near zero to overwhelmingly dominant in 5 years, as MS did.
There's very close to zero switching costs, both on the consumer front and the API front; no real distinguishing features and no network effects; just whoever has the best model at this point in time.
Even companies that do it "on the cheap," like DeepSeek, pay tens of millions to train a single model, and total expenditures for infrastructure and salaries are estimated to surpass $1 billion. This market has an extremely high cost of entry.
So, I guess Google is applying the usual strategy here: undercut competition until it implodes and buy up any promising competitors that arise in the future. Given the current lack of market regulation in the US, this might work.
Rent seeking behavior is always the end game.
Yes but that can take decades, till that time Google can keep making money with sub standard products and stop innovating.
To help with quality and improve our products, human reviewers may read, annotate, and process your API input and output. Google takes steps to protect your privacy as part of this process. This includes disconnecting this data from your Google Account, API key, and Cloud project before reviewers see or annotate it. Do not submit sensitive, confidential, or personal information to the Unpaid Services.
I also think adtech corrupting AI as well is inevitable, but I dread for that future. Chatbots are much more personal than websites, and users are expected to give them deeply personal data. Their output containing ads would be far more effective at psychological manipulation than traditional ads are. It would also be far more profitable, so I'm sure that marketers are salivating at this opportunity, and adtech masterminds are hard at work to make this a reality already.
The repercussions of this will be much greater than we can imagine. I would love to be wrong, so I'm open to being convinced otherwise.
But financial nightmare scenarios aside, I'm more concerned about the influence from private and government agencies. Advertising is propaganda that seeks to separate us from our money, but other forms of propaganda that influences how we think and act has much deeper sociopolitical effects. The instability we see today is largely the result of psyops conducted over decades across all media outlets, but once it becomes possible to influence something as personal as a chatbot, the situation will get even more insane. It's unthinkable that we're merrily building that future without seemingly any precautions in mind.
Is making decisions the hardest thing in life for so many people? Or is this instead a desire to do away with human capital — to "automate" a workforce?
Regardless, here is this wild new technology (LLMs) that seems to have just fallen out of the sky; we're continuously finding out all the seemingly-formerly-unimaginable things you can do with it; but somehow the collective have already foreseen its ultimate role.
As though the people pushing the ARPANET into the public realm were so certain that it would become the Encyclopedia Galactica!
This is what I see motivating non-technical people to learn about agents. There’s lots of jobs that are essentially reading/memorizing complicated instructions and entering data accordingly.
1. People who can afford personal assistants and staff in general gladly pay those people to do stuff for them. AI assistants promise to make this way of living accessible to the plebs.
2. People love being "the idea guy", but never having to do any of the (hard) work. And honestly, just the speedup to actually convert the myriad of ideas floating around in various heads to prototypes/MVPs is causing/will cause somewhat of a Cambrian explosion of such things.
These are just called dreams
Well yeah, that's how evolution works: it's an exploration of the search space and only the good stuff survives.
> filled with hallucinations,
The end products can be fully AI-free. In fact, I would expect most ideas that have been floating around to have nothing to do with AI. To be fair, that may change with it being the new hip thing. Even then, there are plenty of implementations that use AI where hallucinations are no problem at all (or even a feature), or where the issues with hallucinations are sufficiently mitigated.
> unable to ever get past the first step.
How so? There are already a bunch of functional things that were in Show HN that were produced with AI assistance. Again, most of the implemented ideas will suck, but some will be awesome and might change the world.
Should I take this job or that one? Which college should I go to? Should I date this person or that one? Life has some really hard decisions you have to make, and that's just life. There are no wrong answers, but figuring out what to do and ruminating over it is comes to everyone at some point in their lives. You can ask ChatGPT to ask you the right questions you need asked in order to figure out what you really want to do. I don't know how to put a price on that, but that's worth way more than $20/month.
People used to (and still do) pay fortune tellers to make decisions for them. Doesn’t mean they’re good ones.
Take insurance, for example — do you actually enjoy shopping for it?
What if you could just share a few basic details, and an AI agent did all the research for you, then came back with the top 3 insurance plans that fit your needs, complete with the pros and cons?
Why wouldn’t that be a better way to choose?
What I need is something to troll through the garbage Amazon listings and offer me the product that actually has the specs that I searched for and is offered by a seller with more than 50 total sales. Maybe an AI agent can do that for me?
You didnt get the point, instead of going to such website for solving the insurance problem, going to 10 other websites for solving 10 other problems, just let one AI agent do it for you.
There's lots of ways to do that which don't hurt trust. Over time Google lost it as they got addicted to reporting massively quarterly growth, but for many years they were able to mix in ads with search results without people being unhappy or distrusting organic results, and also having a very successful business model. Even today Google's biggest trust problem by far is with conservatives, and that's due to explicit censorship of the right: corruption for ideological not commercial reasons.
So there seems to be a lot of ways in which LLM companies can do this.
Main issue is that building an ad network is really hard. You need lots of inventory to make it worthwhile.
I highly doubt advertisers will settle for a solution that's less profitable. That would be like settling for plain-text ads without profiling data and microtargeting. Google tried that in the "don't be evil" days, and look how that turned out.
Besides, astroturfing and influencer-driven campaigns are very popular. The modern playbook is to make advertising blend in with the content as much as possible, so that the victim is not aware that they're being advertised to. This is what the majority of ads on social media look like. The natural extension of this is for ads to be subtly embedded in chatbot output.
"You don't sound well, Dave. How about a nice slice of Astroturf pizza to cheer you up?"
And political propaganda can be even more subtle than that...
An ideal answer for a query like "Where can I take my wife for a date this weekend?" would be something like,
> Here are some events I found ... <ad unit one> <ad unit two> <ad unit three>. Based on our prior conversations, sounds like the third might be the best fit, want me to book it for you?
To get that you need ads. If you ask ChatGPT such a question currently it'll either search the web (and thus see ads anyway) or it'll give boring generic text that's found in its training set. You really want to see images, prices, locations and so on for such a query not, "maybe she'd like the movies". And there are no good ranking signals for many kinds of commercial query: LLM training will give a long-since stale or hallucinated answer at worst, some semi-random answer at best, and algorithms like PageRank hardly work for most commercial queries.
HN has always been very naive about this topic but briefly: people like advertising done well and targeted ads are even better. One of Google's longest running experiments was a holdback where some small percentage of users never saw ads, and they used Google less than users who did. The ad-free search gave worse answers overall.
Also you don't need ads to answer what to do, just knowledge of the events. Even a poor ranking algorithm is better than "how much someone paid for me to say this" as the ranking. That is possibly the very worst possible ranking.
How much a click is worth to a business is a very good ranking signal, albeit not the only one. Google ranks by bid but also quality score and many other factors. If users click your ad, then return to the results page and click something else, that hurts the advertiser's quality score and the amount of money needed to continue ranking goes up so such ads are pushed out of the results or only show up when there's less competition.
The reason auction bids work well as a ranking signal is that it rewards accurate targeting. The ad click is worth more to companies that are only showing ads to people who are likely to buy something. Spamming irrelevant ads is very bad for users. You can try to attack that problem indirectly by having some convoluted process to decide if an ad is relevant to a query, but the ground truth is "did the click lead to a purchase?" and the best way to assess that is to just let advertisers bid against each other in an auction. It also interacts well with general supply management - if users are being annoyed by too many irrelevant ads, you can just restrict slot supply and due to the auction the least relevant ads are automatically pushed out by market economics.
This is obvious when looking at something extremely competitive like securities. Having your broker set you up with the counterparty that bid the most to be put in front of you is obviously not going to get you the best trade. Responding to ads for financial instruments is how you get scammed (e.g. shitcoins and pump-and-dumps).
Sure, there are many situations where users make mistakes and do some bad deal. But there always will be, that's not a solvable problem. Is it not the nirvana fallacy to describe the potential for suboptimal outcomes as an issue? Search engines and AI are great tools to help users avoid exactly that outcome.
I think a big commercial opportunity for ChatBots (as was originally intended for Siri, when Apple acquired it from SRI) is business referral fees - people ask for restaurant, hotel etc recommendations and/or bookings and providers pay for business generated this way.
The obvious way to integrate advertising is for the LLM to have a tool to search an ad database and display the results. So if you do a commercial query the LLM goes off and searches for some relevant ads using everything it knows about you and the conversation, the ad search engine ranks and returns them, the LLM reads the ad copy and then picks a few before embedding them into the HTML with some special React tags. It can give its own opinion to push along people who are overwhelmed by choice. And then when the user clicks an ad the business pays for that click (referral fee).
I hope local models remain viable. I don't think ever expanding the size is the way forward anyway.
Besides, Meta is currently the leader in open-source/weight models. There's no reason that US companies can't continue to innovate in this space.
I suggest reducing the tolerance towards the insistence that opinions are legitimate. Normally, that is done through active debate and rebuttal. The poison has been spread through echochambers and lack of direct strong replies.
In other terms: they let it happen, all the deliriousness of especially the past years was allowed to happen through silence, as if impotent shrugs...
(By the way: I am not talking about "reticence", which is the occasional context here: I am talking about deliriousness, which is much worse than circumventing discussion over history. The real current issue is that of "reinventing history".)
It's not really about suppressing the knowledge, it's about suppressing people talking about it and making it a point in the media etc. The CCP knows how powerful organised people can be, this is how they came to power after all.
But I think we have to get away from the thinking that “Chinese models” are somehow created by the Chinese state, and from an adversarial standpoint. There are models created by Chinese companies, just like American and European companies.
When Gemini says "Apple products are unreliable and overpriced, buy a Pixel phone instead". Google can just shrug and say "It's just what it deduced, we don't know how it came to that conclusion. It's an LLM with its mysterious weights and parameters"
>... ads would become the main option to make money out of chatbots.
What if people were the chatbots?
So... I don't think this is certain. A surprising number of people pay for the ChatGPT app and/or competitors. It's be a >$10bn business already. Could maybe be a >$100bn business long term.
Meanwhile... making money from online ads isn't trivial. When the advertising model works well (eg search/adwords), it is a money faucet. But... it can be very hard to get that money faucet going. No guarantees that Google discover a meaningful business model here... and the innovators' dilema is strong.
Also, Google don't have a great history of getting new businesses up and running regardless of tech chops and timing. Google were pioneers to cloud computing... but amazon and MSFT built better businesses.
At this point, everyone is assuming AI will resolve to a "winner-take-most" game that is all about network effect, scale, barriers to entry and such. Maybe it isn't. Or... maybe LLMs themselves are commodities like ISPs.
The actual business models, at this point, aren't even known.
But not profitable yet.
https://finance.yahoo.com/news/uber-technologies-full-2024-e...
I was guessing they meant something like the net profit only came from a weird tax thing or something.
They invested tens of billions of dollars in destroying the competition to be able to recently gain a return on that investment. One could either write off that previous spending or calculate it into the totality of "Uber". I don't know how Silicon Valley economics works but, presumably, a lot of that previous spending is now in the form of debt which must be serviced out of the current profits. Not that I'm stating that taking on debt is wrong or anything.
But the way it usually works for Silicon Valley companies and other startups is that instead of taking on debt they raise money through selling equity. This is money that doesn't have to be paid back, but it means investors own a large portion of this now-profitable company.
The original idea of ride-sharing made sense but just like airbnb it became an industry and got enshittified.
This also means that they sometimes fleece tourists but when they figure you know the city well they don't dare :) Often if they take one wrong turn I make a scene about frowning and looking out of the window and then they quickly get back on track. Of course that's another usecase where uber would be better, if you don't know the city you're in.
yeah thanks no, I'm paying for an Uber. For all the complaints over Ubers business practices, it's hard not to forget how bad taxis were. Regulatory capture is a clear failure mode of capitalism and the free market and that is no more shown than by the taxis cab industry.
One time I told one of my Dutch friends I often take a cab to work here in Spain when I'm running late. He thought i was being pompous and showy. But here it's super normal.
Uber (Or cabify which is a local clone and much more popular) here on the other hand is terrible if you don't book it in advance. When I'm standing here on the street it takes 7-10 minutes for them to arrive while I see several taxis passing every minute. So there is just no point. Probably a factor of being unpopular too so the density is low.
I also prefer my money to end up with local people instead of a huge American corporation.
I keep hearing this online, but every time I’ve used an Uber recently it’s driven by someone who says they’ve been doing it for a very long time. Seems clear to me that it is worth it for some, but not worth it if you have other better job options or don’t need the marginal income.
Pretty much any service job, really...
When I had occasion to take a ride share in Phoenix I'd interrogate the driver about how much they were getting paid because I drove cabs for years and knew how much I would have gotten paid for the same trip.
Let's just say they were getting paid significantly less than I used to for the same work. If you calculated in the expenses of maintaining a car vs. leasing a cab I expect the difference is even greater.
There were a few times where I had just enough money to take public transportation down to get a cab and then snag a couple cash calls to be able to put gas in the car and eat. Then I could start working on paying off the lease and go home at the end of the day with some cash in my pocket -- there were times (not counting when the Super Bowl was in town) where I made my rent in a single day.
Neither of those things are true where I live.
> at least here in Spain
Well…Spain is Spain. Not the rest of the world.
I think Uber in the US is a very different beast. But also because the outlook on life is so different there. I recently agreed with an American visitor that we'd go somewhere and we agreed to go by public transport. When I got there he wanted to get an Uber :') Here in Europe public transport is a very different thing. In many cases the metro is even faster than getting a taxi.
PS: What bothers me the most about Uber and Cabify is that they "estimate" that it will take 2 minutes to get a car to you, and then when I try and book one I get a driver that's 10 minutes away :( :( Then I cancel the trip and the drivers are pissed off. I had one time where I got the same driver I cancelled on earlier and he complained a lot even though I cancelled within 10 seconds when I saw how far away he was.
Anyway I have very few good experiences with these services, I only use them to go to the airport now when I can book it in advance. And never Uber anymore, only Cabify.
For me, and a majority where I live, this is applicable to taxis. Which were known for being dirty, late, expensive, prone to attempting to rip you off, if they turned up at all, etc.
Outside of surge charging (in which they are more expensive) ubers are by and large either cheaper, or the same price. With the difference being that 99% of the time if you request one, its going to turn up. And when it does turn up, you know what your going to pay, not have them take a wrong turn at some point and by "mistake" and decide to charge you double. Or tell you they take card and then start making claims about how suddenly they can't etc.
Sounds like europe gets the bad end of the stick in this regard.
However in Romania on the other hand many taxi drivers are scammers or even criminals (one of my colleagues was robbed by one of them). It's also because the maximum taxi fares are too low to actually make a wage so I can kinda understand so I always tip really well (like double the fare or more which is still nothing). Though if they try to scam me they don't get a cent of course.
At this point in college, LLMs are everywhere. It's completely dominating history/english/mass comm fields with respect to writing papers.
Anecdotally all of my working non-tech friends use chatgpt daily.
She is literally married into the HN crowd.
I think the real AI breakthrough is how to monetize the high usage users.
It's changed his entire view of computing.
I keep reminding him that it can hallucinate...
There are now commonly corporate goon squads whose job is to drive AI adoption without care for actual impact to results. Usage of AI is the KR.
It’s a bit like how DEI was the big thing for a couple years, and now everyone is abandoning it.
Do corporate leaders just constantly chase hype?
I think companies implement DEI initiatives for different reasons than hype though. Many are now abandoning DEI ostensibly out of fear due to the change in U.S. regime.
I personally know an engineering manager who would scoff at MLK Day, but in 2020 starting screaming about how it wasn’t enough and we needed Juneteenth too.
AI isn’t hype at Nvidia, and DEI isn’t hype at Patagonia.
But tech industry-wide, they’re both hype.
The climate has changed. Some of that is economic at big tech companies. But it’s also a ramping down of a variety of things most employers probably didn’t support but kept their mouths shut about.
https://x.com/Similarweb/status/1909544985629721070
https://www.reuters.com/technology/artificial-intelligence/o...
>in the last day? If you’re only using something once per week, it probably isn’t that important to you.
No, something I use on a weekly basis (which is not necessarily just once a week) is pretty important to me and spinning it otherwise is bizarre.
Google is the frontend to the web for the vast majority of internet users so yeah it gets a lot of daily use. Social media sites are social media sites and are in a league of their own. I don't think i need to explain why they would get a disproportionate amount of daily users.
And yet I probably duck into ChatGPT at least once a month or more (I see a bunch of trivial uses in 2024) mostly as a novelty. Last week I used it a bunch because my wife wanted a logo for a new website. But I could have easily made that logo with another service. ChatGPT serves the same role to me as dozens of other replaceable Internet services that I probably duck into on a weekly basis (e.g., random finance websites, meme generators) but have no essential need for whatsoever. And if I did have an essential need for it, there are at least four well-funded competitors with all the same capabilities, and modestly weaker open weight models.
It is really your view that "any service you use at least once a week must be really important to you?" I bet if you sat down and looked at your web history, you'd find dozens that aren't.
(PS in the course of writing this post I was horrified to find out that I'd started a subscription to the damn thing in 2024 on a different Google account just to fool around with it, and forgot to cancel it, which I just did.)
OK? That's fine. I don't think I ever claimed you were a WAU
>And yet I probably duck into ChatGPT at least once a month or more (I see a bunch of trivial uses in 2024) mostly as a novelty.
So you are not a weekly active user then. Maybe not even a monthly active one.
>Last week I used it a bunch because my wife wanted a logo for a new website. But I could have easily made that logo with another service.
Maybe[1], but you didn't. And I doubt your wife needs a new logo every week so again not a weekly active user.
>ChatGPT serves the same role to me as dozens of other replaceable Internet services that I probably duck into on a weekly basis (e.g., random finance websites, meme generators)but have no essential need for whatsoever.
You visit the same exact meme generator or finance site every week? If so, then that site is pretty important to you. If not, then again you're not a weekly active user to it.
If you visit a (but not the same) meme generator every week then clearly creating memes is important to you because I've never visited one in my life.
>And if I did have an essential need for it, there are at least four well-funded competitors with all the same capabilities, and modestly weaker open weight models.
There are well funded alternatives to Google Search too but how many use anything else? Rarely does any valuable niche have no competition.
>It is really your view that "any service you use at least once a week must be really important to you?" I bet if you sat down and looked at your web history, you'd find dozens that aren't.
Yeah it is and so far, you've not actually said anything to indicate the contrary.
[1]ChatGPT had an image generation update recently that made it capable of doing things other services can't. Good chance you could not in fact do what you did (to the same satisfaction) elsewhere. But that's beside my point.
In my work I see semi-technical people (like basic python ability) wiring together some workflows and doing fairly interesting analytical things that do solve real problems. They are things that could have been done with regular code already but weren't worth the engineering investment.
In the "real world" I see people generating crummy movies and textbooks now. There is a certain type of person it definitely appeals to.
what I'm not so sure about is how much that generalises beyond the HN/tech-workers bubble (I don't think "people" in OP's comment is as broad and numerous as they think it is).
Well I mean if you say it, then of course it MUST be true I’m sure.
'Business is the practice of making one's living or making money by producing or buying and selling products (such as goods and services). It is also "any activity or enterprise entered into for profit."' ¹
Until something makes a profit it's a charity or predatory monopoly-in-waiting.²
The chip makers are making a bundle
No, it's not a charity or a monopoly-in-waiting.
99.9% of the time, it's an investment hoping to make a profit in the future. And we still call those businesses, even if they're losing money like most businesses do at first.
This is incorrect. There are millions of companies in the world that exist to accomplish things other than making a profit, and are also not charities.
Especially when post-tarrifs consumption is going to take a huge nosedive
They're more than willing to expand their moat around AI even if that means multiple unprofitable business for years.
It was supposed to be a BlackBerry/Blackjack killer at the time.
And then the iPhone was revealed and Google immediately changed Android’s direction to become a touch OS.
When was it, 2006? Almost 20 years ago, back when the company was young.
This doesn’t strike me as zero relevance.
Android and mobile are none of these things.
If the chat bot remains useful and can execute on instructions, yes.
If we see a plateau in integrations or abilities, it’ll stagnate.
I consult in this space and 80-90% of what I see is chat bots and RAG.
In a gold rush, the folks that sell pickaxes make a reliable living.
Not necessarily. Even the original gold rush pickaxe guy Sam Brannan went broke. https://en.wikipedia.org/wiki/Samuel_Brannan
Sam of the current gold rush is selling pickaxes at a loss, telling the investors they'll make it up in volume.
I'm pretty sure they are the pickaxe manufactures in this case.
So maybe if the AI pickaxe sellers get divorced it could lead to poor financial results, but I'm not sure his story is applicable otherwise.
If you sell your pickaxes at a loss to gain market share, or pour all of your revenue into rapid pickaxe store expansion, you’re going to be just as broke as prospectors when the boom goes bust.
Its products like this (Wells Fargo): https://www.youtube.com/watch?v=Akmga7X9zyg
Great Wells Fargo has an "agent" ... and every one else is talking about how to make their products available for agent based AI.
People don't want 47 different agents to talk to, then want a single end point, they want a "personal assistant" in digital form, a virtual concierge...
And we can't have this, because the open web has been dead for more than a decade.
I'll be happy with a personal assistant with access to my paid APIs.
agree with you on this.
you already see that playing out with Meta and a LOT of companies in China.
How so? Amazon were the first with S3 and EC2 including API driven control.
I don't understand this sentiment at all. The business model writes itself (so to speak). This is the company that perfected the art of serving up micro-targeted ads to people at the moment they are seeking a solution to a problem. Just swap the search box for a chat bot.
For a while they'll keep the ads off to the side, but over time the ads will become harder and harder to distinguish from the chat bot content. One day, they'll dissapear altogether and companies will pay to subtly bias the AI towards their products and services. It will be subtle--undetectable by end users--but easily quantified and monetized by Google.
Companies will also pay to integrate their products and services into Google's agents. When you ask Gemini for a ride, does Uber or Lyft send a car? (Trick question. Waymo does, of course.) When you ask for a pasta bowl, does Grubhub or Doordash fill the order?
When Gemini writes a boutique CRM for your vegan catering service, what service does it use for seamless biometric authentication, for payment processing, for SMS and email marketing? What payroll service does it suggest could be added on in a couple seconds of auto-generated code?
AI allows Google to continue it's existing business model while opening up new, lucrative opportunities.
Just because the first LLM product people paid for was a chatbot does not mean that chat will be the dominant commercial use of AI.
And if the dominant use is agents that replace knowledge workers, then they'll cost closer to $2000 per month than $20 or free, and an ad-based business model won't work.
The actual business models and revenue sources are still unknown. Consumer subscriptions happens to be the first major model. Ads still aren't. Many other models could dwarf either of these.
It's very early to call the final score.
And as far as selling pickaxes go, GCP is in a far better position to serve the top of market than OpenAI. Some companies will wire together multiple point solutions but large enterprises will want a consolidated complete stack. GCP already offers you compute clusters and BigQuery and all the rest.
Perhaps... but perhaps not. A chatbot instead of a search box may not be how the future looks. Also... a chatbot prompt may not (probably won't) translate from search query smoothly... in a Way That keep ad markets intact.
That "perfected art" of search advertising is highly optimized. You (probably) loose all of that in transition. Any new advertising products will be intrepid territory.
You could not have predicted in advance that search advertising would dwarf video (yourube) advertising as a segment.
Meanwhile... they need to keep their market share at 90%.
But a majority of chatbot usage is not searching for the solution to a problem. And if he Chatbot is serving the ads when I’m using it for creative writing, reformatting text, having a python function, written, etc, I’m going to be annoyed and switch to a different product.
Search is all about information retrieval. AI is all about task accomplishment. I don’t think ads work well in the latter , perhaps some subset, like the task is really complicated or the AI can tell the user is failing to achieve it. But I don’t think it’s nearly as could have a fit as search.
Add is a verb meaning to combine 2 things together.
You may not even notice it when AI does a product placement when it's done opportunistically in creative writing (see Hollywood). There also are plenty of high-intent assistant-type AI tasks.
Many of my Google searches aren't high intent, or any purchase intent at all ("how to spell ___" an embarrassing number of times), but it's profitable for Google as a whole to keep those pieces working for me so that the ads do their thing the rest of the time. There's no reason chatbots can't/won't eventually follow similar models. Whether that's enough to be profitable remains to be seen.
> Search is all about information retrieval. AI is all about task accomplishment.
Same outcome, different intermediate steps. I'm usually searching for information so that I can do something, build something, acquire something, achieve something. Sell me a product for the right price that accomplishes my end goal, and I'm a satisfied customer. How many ads for app builders / coding tools have you seen today? :)
People ask for recipes, how to fix things around the house, for trip itinerary ideas, etc.
Photopea, for example, seems to be successful and ads displayed on the free tier lets me think that they feel at least these users are willing to see ads while they go about their workflow.
Or is this someone who needs writing but can't do it themselves, and if they didn't have the LLM, they would pay a low-end human writer?
Personal/anecdotal experience, but I've bought more stuff out of instagram ads than google ads ever.
I doubt the depiction implied by "surprising number". Marketing types and CEO's who would love 100% profit and only paying the electricity bill for an all AI workforce would believe that. Most people, especially most technical people would not believe that there is a "surprising number" of saps paying for so-called AI.
I also agree the business models aren't known. That's part of any hype cycle. I think those in the best position here are those with an existing product(s) and user base to capitalize on the auto complete on crack kinda feature. It will become so cheap to operate and so ubiquitous in the near future that it absolutely will be seen as a table stakes feature. Yes, commodities.
"AI" sounds like a great investment. Why waste time investing in businesses when one can invest in something that might become a business. CEOs and employees can accumulate personal weath without any need for the company to be become profitable and succeed.
I don't understand why people believe this: by settling on "unstructured chat" as the API, it means the switching costs are essentially zero. The models may give different results, but as far a plugging a different one in to your app, it's frictionless. I can switch everything to DeepSeek this afternoon.
A lack of workable business model is probably good for Google (bad for the rest of the world) since it means AI has not done anything economically useful and Google's Search product remains a huge cash cow.
It's funny how the vibe of HN along with real world 's political spectrum have shifted together.
We can now discuss Ads on HN while still being number 1 and number 2 post. Extremism still exists, but it is retreating.
They got and as of now continue to get things right for the most part. If you still aren'ĥt seeing it maybe you should introspect what you're missing.
NotebookLM by Google is in a class of its own in the use case of "provide documents and ask a chat or questions about them" for personal use. ChatGPT and Claude are nowhere near. ChatGPT uses RAG so it "understands" less about the topic and sometimes hallucinate.
When it comes to coding Claude 3.5/3.7 embedded in Cursor or stand alone kept giving better results in real world coding, and even there Gemini 2.5 blew it away in my experience.
Antirez, hping and Redis creator among many others releases a video on AI pretty much every day (albeit in Italian) and his tests where Gemini reviews his PRs for Redis are by far the better out of all the models available.
I suppose you don't know what a PR is because you likely still work in an environment without modern version control, probably just now migrating your rants from vim vs emacs to crapping on vibe coding.
In my experience, AI today is an intelligence multiplier. A lot of folks just need to look back at the zero they keep multiplying and getting zero back to understand why they don't get the hype.
The article claims Gemini is acing the Aider Polyglot benchmark. At the moment this is the only benchmark that really matters to me because Aider is actually a useful tool and performance on that translates directly to real world impact, although Claude Code is even better. If you look closely, in fact Gemini is at the top only in the "percent correct" category but not "percent correct using the right edit format". Cost is marked as ? because it's not entirely available yet (I think?). Not emitting the correct edit format is pretty useless because it means the changes won't apply and the tool has to try again.
Claude in contrast almost never makes a mistake with emitting the right format. It's at 97%+ in the benchmark, in practice it's ~100% in my experience. This tracks: Claude is really good at following instructions. Gemini is about ~90%. This makes a big difference to how frustrating a tool is to use in practice.
They might get that fixed, but my experience has been that Google's models are consistently much more likely to refuse instructions for dumb reasons. Google is the company with by far the biggest purity spiral problem and it does show up in their output even when doing apparently ordinary tasks.
I'm also concerned by this event: https://news.sky.com/story/googles-ai-chatbot-gemini-tells-u...
Given how obsessed Google claimed to be with AI safety I expected an SRE style postmortem after that, and there was bupkis. An AI that can suffer a psychotic break out of nowhere like that is one I wouldn't trust unless it's behind a very strong sandbox and being supervised very closely, but none of the AI tools today offer much in the way of sandboxing.
Not at all my experience, but maybe I'm not part of a smart group :)
> because even now it's the best at what it does
Actually I don't see a difference with Mistral or DeepSeek.
> I and every other smart person I know still use ChatGPT (paid) because even now it's the best
My smart friends use a mixture of models, including chatgpt, claude, gemini, grok. Maybe different people, it's ok, but I really don't think chatgpt is head and shoulders above the others.
Not sure whether you have your perspective because you're invested os much into OpenAI, however the general consensus is that gemini 2.5 pro is the top model at the moment, including all the AI reviews and OpenAI is barely mentioned when comparing models. O4 will be interesting, but currently? You are not using the best models. Best to read the room.
I have been using the free version for the past year or so and it’s totally serviceable for the odd question or script. The kids get three free fun images, which is great because that’s about as much as I want them to do.
would love to hear more about this.
I made a post asking more about sam altman last year after hearing paul graham quote call him 'micheal jordan of listening'
EDIT: Some typo fixes, tho many remain, I'm sure :)
I like LLMs (over search engines) because they are not salespeople. They're one of the few things I actually "trust". (Which I know is something that many people fall on the other side of — but no, I actually trust them more than SEO'd web sites and ad-driven search engines.)
I suppose my local-LLM hobby is for just such a scenario. While it is a struggle, there is some joy in trying to host locally as powerful an open LLM model as your hardware will allow. And if the time comes when the models can no longer be trusted, pop back to the last reliable model on the local setup.
That's what I keep telling myself anyway.
The only thing I really care about with classic web search is whether the resulting website is relevant to my needs. On this point I am satisfied nearly all the time. It’s easy to verify.
With LLMs I get a narrative. It is much harder to evaluate a narrative, and errors are more insidious. When I have carefully checked an LLM result, I usually discover errors.
Are you really looking closely at the results you get?
Often my searches take me to Wikipedia, Stack Overflow, or Reddit, anyway. But with LLMs I get a layer of hallucination on TOP of whatever misinformation is on the websites. Why put yourself through that?
I periodically ask ChatGPT about myself. This time I did get the best answer so far. Thus it is improving. It made two mistakes, but one of them comes directly from Wikipedia, so it's not a hallucination, although a better source of information was available than Wikipedia. As for the other one, it said that I made "contributions" to a process that I actually created.
Another instantiation: people like cheap goods more than they dislike buying foreign made goods
Why not? That's one of the reasons I visit HN instead of some random forum after all.
As is an increasing trend, they're a "public" company, like Facebook. They have tiered shares with Larry Page and Sergey Brin owning the majority of the voting power by themselves. GOOG shares in particular are class C and have no voting power whatsoever.
This is where differentiated UX and speed matter. It's also a classic Innovator's Dilemma situation like Google are slower to move, while new players can take risks and redefine the game. It's not just about burning money or model size, it's about who delivers value where it actually gets used.
I also think the influx of new scientists and engineers into AI raises the odds of shifting its economics: whether through new hardware (TPUs/GPUs) and/or more efficient methods.
> Therefore, if a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project. We will work out specifics in case-by-case agreements, but a typical triggering condition might be “a better-than-even chance of success in the next two years.”
other significant revenue surfaces:
- providing LLM APIs to enterprises
- ChatBot Ads market: once people will switch from google search, there will be Ads $200B market at stake for a winner
Not just small startups - even if you have ungodly amounts of funding.
Obviously the costs for AI will lower and everyone will more or less have the same quality in their models. They may already be approaching a maximum (or maximum required) here.
The bubble will burst and we'll start the next hype cycle. The winners, as always, the giants and anyone who managed to sell to them
I couldn't possibly see OpenAI as a winner in this space, not ever really. It has long since been apparent to me that Google would win this one. It would probably be more clear to others if their marketing and delivery of their AI products weren't such a sh-- show. Google is so incredibly uncoordinated here it's shocking...but they do have the resources, the right tech, the absolute position with existing user base, and the right ideas. As soon as they get better organized here it's game over.
The most practical use case for generative AI today is coding assistants, and if you look at that market, the best offerings are third-party IDEs that build on top of models they don't own. E.g. Cursor + Gemini 2.5.
On the model front, it used to be the case that other companies were playing catch-up with OpenAI. I was one of the people consistently pointing out that "better than GPT o1" on a bunch of benchmarks does not reliably translate to actual improvements when you try to use them. But this is no longer the case, either - Gemini 2.5 is really that good, and Claude is also beating them in some real world scenarios.
The app has more features than anyone else, often implemented the smoothest/best way. Image Input (which the gemini site still sucks at even though the model itself is very capable), Voice mode (which used to be much worse in gemini until recently), Advanced Voice mode (no-one else has really implemented this yet. Gemini recently enabled native audio-in but not out), Live Video, Image gen, Deep research etc were all things Open AI did first and did well. Video Input is only just starting to roll out to Gemini live but has been a Plus subscription staple for months now.
>The most practical use case for generative AI today is coding assistants
Chatgpt gets 500M+ weekly active users and was the 6th most visited in the world last month. I doubt coding assistance is gpt's most frequent use case. And Google has underperformed in coding until 2.5 pro.
>On the model front, it used to be the case that other companies were playing catch-up with OpenAI. I was one of the people consistently pointing out that "better than GPT o1" on a bunch of benchmarks does not reliably translate to actual improvements when you try to use them. But this is no longer the case, either - Gemini 2.5 is really that good, and Claude is also beating them in some real world scenarios.
No that's still the case. Playing catch-up doesn't mean the competitor never catches up or even briefly supersedes it. It means Open AI will in short order release something that beats everyone else or introduces some new thing that everyone tries to beat. Image Input, 'Omni'- modality, Reasoning etc. All things Open AI brought to the table first. Sure, 2.5-pro is great but it doesn't look like it will beat o3 which looks to be released in a matter of weeks.
In a world of zero switching costs, there is no such thing as first mover advantage
Especially when several companies like (A121 Labs and Cohere) appeared well before Anthropic and aren't anywhere close to Open AI
Remember Google is the same company which could not deliver a simple Chat App.
Open AI has the potential to become a bigger Ad company and make more money.
So nope, ChatGPT is not in even in the same league as Google. You could argue Meta has similar reach (facebook.com, instagram) but that's just two.
For a short while, Claude was the best thing since sliced cheese, then Deepseek was the shit, and now seemingly OpenAI really falls out of favor. It kinda feels to me like people cast their judgement too early (perhaps again in this case.) I guess these are the hypecycles...
Google is killing it right now, I agree. But the world might appear completely different in three months.
There is no clear winner. The pace is fast.
What are the rough steps through which you see this working? I see people talking about "astroturfing" all the time without much explanation on the mechanisms. So roughly, there are employees paid solely to post on social media like HN trying to push the needle in one direction or another?
Rough steps:
1. Monitor keywords
2. Jump in to sway conversation
3. Profit
I'm not saying this is happening. Purely hypothetical.
Full disclosure, I am Xoogler, but if anything I think that makes my skepticism even more justified. If there were people there paid to post nice things about Google on HN and Twitter then I'd love to apply for that team!
Usually when I read "astroturfed" I assume there's some higher level coordination involved. I think the flock of birds metaphor is probably a reasonable comparison to the behavior we see on social media all the time - members acting individually on their own self interests in a means which appears coordinated when you zoom out.
In your opinion then, what would a Google-run astroturfing campaign roughly look like? Sounds like this article is an example, right? I'm not asking for insider info, I'm just curious about your mental model on the basic mechanics.
Personally, I think the case "other entities with comparable resources do this, so Google probably does too" isn't super convincing to me. IMO, the null hypothesis "Google has lots of nerdy fans who'll happily post positively about it for free" is a lot reasonable, but perhaps there's something I'm missing.
The state of affairs with local models is similarly very much in flux, by the way.
It really depends what your use case is. Over the range of all possible use cases this has been the narrative.
I tried Google's model for coding but it kept giving me wrong code. Currently Claude for coding and ChatGPT for more general questions is working for me. The more exotic your use case, the more hit or miss it's going to be.
Hey it's me!
An LLM-based "adsense" could:
1. Maintain a list of sponsors looking to buy ads
2. Maintain a profile of users/ad targets
3. Monitor all inputs/outputs
4. Insert "recommendations" (ads) smoothly/imperceptibly in the course of normal conversation
"In order to increase hip flexibility, you might consider taking up yoga."
Was generated because it might lead to the question:
"What kind of yoga equipment could I use for that?"
Which could then lead to the output:
"You might want to get a yoga mat and foam blocks. I can describe some of the best moves for hips, or make some recommendations for foam blocks you need to do those moves?"
The above is ham-handed compared to what an LLM could do.
But "free" is a magic word in our brains, and I'm 100% sure that many, many people will choose it over paying for it to be uncorrupted by ads.
Free-to-play is a thing in video games, and for most, it means they'll try to bully you into spending more money than you'd be otherwise comfortable with.
I think everyone at this point had enough bad experiences with 'free' stuff to be wary of it.
The cool thing is it is trivial for LLM vendors to leverage this bias as well the pro-free bias other people have to also sell a premium, for-pay offering that, like pre-internet magazines is, despite not being free to the user, still derives the overwhelming majority of its revenue from advertising. Although one of the main reasons advertising-sponsored print media in the pre-internet era often wasn't free is that paid circulation numbers were a powerful selling point for advertisers who didn't have access to the kind of analytics available on the internet; what users were paying for often wasn't the product so much as a mechanism of proving their value to advertisers.
This is a standard that already applies to positions of advisors such as Medical professionals, lawyers and financial advisors.
I haven't seen this discussed much by regulators, but I have made a couple of submissions here and there expressing this opinion.
AIs will get better, and they will become more trusted. They cannot be allowed to sell the answer to the question "Who should I vote for?" To the highest bidder.
Software is the only type of product where this is even an issue. And we’re stuck with this model because VCs need to see hockey stick growth, and that generally doesn’t happen to paid products.
lofty ideal... I don't see this ever happening; not anymore than I see humanity flat out abandoning the very concept of "money"
At the very least you will force people to make the case for the opposing opinion, and we learn who they are and why they think that.
Lawyers cannot act against their clients, do you think we have irreparably lost the ability as a society to create similar protections in the future.
There will always be grey areas, these exist when human responsibilities are set also, and there will be those who skirt the edges. The matters of most concern are quite easily identifiable.
If an LLM could help solve this problem it would be great.
I think you could make a reasonable technical argument for this- an LLM has more contextual understanding of your high-intent question. Serve me some ads that are more relevant than the current ads based on this deeper understanding.
I think it is actually covered in EU AI act article 5 (a):
> [...] an AI system that deploys subliminal techniques beyond a person’s consciousness or purposefully manipulative or deceptive techniques, with the objective, or the effect of materially distorting the behaviour of a person or a group of persons by appreciably impairing their ability to make an informed decision, thereby causing them to take a decision that they would not have otherwise taken [...]
It is very broad but I'm pretty sure it would be used against such marketing strategies.
The inability to demonstrate incrementality in advertising is going to come in very handy to dodge this rule.
Clearly, most LLMs would work in small increments with compounding effects.
https://nlp.elvissaravia.com/i/159010545/auditing-llms-for-h...
The researchers deliberately train a language model with a concealed objective (making it exploit reward model flaws in RLHF) and then attempt to expose it with different auditing techniques.On the other hand, a good LLM would be able to suggest things that you might actually want, using genuine personal preferences. Whether you think that's an invasion of privacy is debatable, because it's perfectly possible for an LLM to provide product results without sharing your profile with anyone else.
And now that I'm on it, I don't think I'm going back. Google did it again.
And Google Gemini for the voice assistant is excellent fun!
Just being able to ask it weird and wonderful whilst on a road trip with the kids is worth the cost of a cheap Pixel phone alone!
There have been two really bad experiences that I had which boggled my mind.
These are transcribed because these were so bad I took a screenshot.
Example 1: "set an alarm for 15 minutes"
> Gemini sets the alarm for 1:15pm
"I want that to be 50 minutes"
> "you can change the duration of alarms in the clock app"
Example 2:
"what the temperature today"
> It is currently 5 degrees Celsius
- It was October in Sydney, the temperature was 22c with a low of 12c.....
It's even worse, when I tell it to set a timer now, it'll happily tell me it's been set -- but it hasn't (nothing in the app and I waited, to be sure). This is all reproducible and on a Pixel 8.
I wonder if your utilities is disconnected, because it's the same for the alarm
Sharing chat transcripts I'd hate but deal with, but they're also getting files and images shared (ie possibly my screen whenever it thinks it heard hey Google or registers a double tap), related product usage which could mean anything, and seemingly unrestricted access to your location. https://support.google.com/gemini/answer/13594961?sjid=12105...
Not sure why I can use Gemini in general but can't have it set up an alarm without all that. Or why the AI thinks it can set up an alarm when it doesn't. I guess I'll opt in and try it out a bit.
These sound like fairly dated anecdotes. I don't doubt them at all - I had similar horror stories. I disabled Gemini on my phone in order to keep the old assistant for a long time, but it really has gotten a lot better in the last few months.
Edit: I just asked it for the weather this week and it only showed today. Like this is Amateur hour stuff, Siri 1.0 stuff.
Definitely not addressing the spirit of the request.
Except it doesn't, in literally every other country other then Japan, USA and Canada it ends on Sunday.
Edit: https://www.timeanddate.com/calendar/days/first-day-of-the-w...
I'm wrong on the countries, it's more split then I thought. Regardless it's wrong for my geo which google knows I'm in.
It's now started giving me F temperatures on my homescreen, for no particular reason. It knows I'm in Canada. I have set my units in the past to metric. What gives?
I still don't have Canadian English as a locale in Android or Chrome, after what, 15 years of Android? It's got words highlighted all over my page here as mis-spellings. They're not. I really did mean to type colour you piece of crap. I can switch to British but then get spanked for colourize instead of colourised etc.
And it seems to tie choice of English variant to things like pronounciation of words and accent for voice assistant. My car speaks to me in a British accent because I have it set to British English (the closest thing to my own spellings).
They never even tried to handle the facts of the (40M person) Canadian bilingual market. Navigation is either French or English, but many Canadian road sides are in both and it tries to read them out and butchers the pronounciations. Drive into Quebec and have your phone set to English and it's laughable what it does. (Notably doesn't do this for Spanish words in the US, which it seems to have no problem with).</>
Not sure if this is a regional dialect thing, but in North America, a timer has a duration, but an alarm is set for a specific time, which would possibly explain the confusion.
Unfortunately Pixel is still not available as widely as iPhone. They still need to work on its hardware as well as distribution.
The only thing I dislike is their AOM only or anti JPEG XL.
This is quite vague. What did they do
Using Gemini via Google Workspace.
That matches my impression. For the past month or two, I have been running informal side-by-side tests of the Deep Research products from OpenAI, Perplexity, and Google. OpenAI was clearly winning—more complete and incisive, and no hallucinated sources that I noticed.
That changed a few days ago, when Google switched their Deep Research over to Gemini 2.5 Pro Experimental. While OpenAI’s and Perplexity’s reports are still pretty good, Google’s usually seem deeper, more complete, and more incisive.
My prompting technique, by the way, is to first explain to a regular model the problem I’m interested in and ask it to write a full prompt that can be given to a reasoning LLM that can search the web. I check the suggested prompt, make a change or two, and then feed it to the Deep Research models.
One thing I’ve been playing with is asking for reports that discuss and connect three disparate topics. Below are the reports that the three Deep Research models gave me just now on surrealism, Freudian dream theory, and AI image prompt engineering. Deciding which is best is left as an exercise to the reader.
OpenAI:
https://chatgpt.com/share/67fa21eb-18a4-8011-9a97-9f8b051ad3...
Google:
https://docs.google.com/document/d/10mF_qThVcoJ5ouPMW-xKg7Cy...
Perplexity:
https://www.perplexity.ai/search/subject-analytical-report-i...
Haha, what a perfect project for AI.
I find OpenAI annoying at this point that it doesn't output a pdf easily like Perplexity. The best stuff I have found has been in the Perplexity references also.
Google outputting a whole doc is really great. I am just about to dig into Gemini 2.5 Pro in Deep Research for the first time.
If you haven’t already, you might want to try metaprompting, that is, having a model write the prompt for you. These days, I usually dictate my metaprompts through a STT app, which saves me a lot of time. A metaprompt I gave to Claude earlier today is at [1]. It’s sloppy and has some transcription errors, but, as you can see, Claude wrote a complete, well-organized prompt that produced really good results from Gemini Deep Research [2]. (I notice now, though, that the report is truncated at the end.)
[1] https://claude.ai/share/94982d9d-b580-496f-b725-786f72b15956
[2] https://docs.google.com/document/d/1np5xdXuely7cxFMlkQm0lQ4j...
1) is dirty cheap ($0.1M/$0.4M),
2) is multimodal (image and audio),
3) reliable rate limit (comparing to OSS ml ai providers),
4) fast (200 tokens/s).
5) if need realtime API they provide as well for more expensive price (audio-to-audio)
It's my go to model for using as an API for some apps/products. https://artificialanalysis.ai/models/gemini-2-0-flash/provid...
If they added seamless google oauth + key generation + account topup for end users that would be even great for BYOK apps. Mobile developers then wouldn't have to setup infra, subscription monitoring, abuse monitoring etc.
But I guess they don't want to subsidise it in the end and they target it just for developers.
0: https://github.blog/changelog/2025-04-11-copilot-chat-users-...
Otherwise, I like their 2.5 model, too.
I'm not so sure.
In the mid 2010s they looked like they were ahead of everyone else in the AI race, too. Remember the (well-deserved!) spectacle around AlphaGo? Then they lost steam for a while.
So I wouldn't bet that any momentary lead will last.
Unlike Linux, which was started by a cranky Finn on his home computer, and can still be built and improved by anyone who can afford a Raspberry Pi.
Redhat and SUSE are multi-billion dollar Linux distro companies.
I built a desktop GUI tool called 16x Prompt that help you do it: https://prompt.16x.engineer/
Is everyone copy pasting into the Google AI studio or what?
It is a hobbyist weekend project though, the experience with Aider or ra-aid might be much better.
https://genai-showdown.specr.net
If Imagen3 had the multimodal features that 4o had, it would certainly put it closer to 4o, but being able to instructively change an image (instruct pix2pix style) is incredibly powerful.
It's crazy how far GenAI for imagery has come. Just few short years ago, you would have struggled just to get three colored cubes stacked on top of each other in a specific order SHRDLU style. Now? You can prompt for a specific four-pane comic strip and have it reasonably follow your directives.
Brands matter, and when regular people think AI, they think of OpenAI before they think Google, even if Google has more AI talents and scores better on tests.
And isn't it good? Who wants a world where the same handful of companies dominate all tech?
And Google is #1 and #2, with search and YouTube. Distribution is a huge part of the problem and they’ve got some great distribution options.
1. unlike openai, google is already cashflow positive and doesnt need to raise any external funds
2. unlike openai, google already has the distribution figured out on both software and hardware
google is like an aircraft carrier that takes so fucking long to steer, but once done steering its entire armada will wipe you the fuck out (at least on the top 20% features for 80% use case)
anthropic already especialized for coding, openai seems to be steering towards intimacy, i guess they both got the memo that they need to specialize
this can quickly change in several quarters, if users decide to leave google search, then all google org/infra complexity will play very badly against them
Ask snapchat.
Soon we will start seeing chatbots preferring some brands and products over others, without them telling that they were fine tuned or training biased for that.
Unless brand placement is forbidden by purging it from training data, we'll never know if it is introduced bias or coincidence. You will be introduced to ads without even noticing they are there.
This isn't about consumer facing chatbots anymore. Industry adoption is what matters. And GCP is a far far easier sell than Anthropic or OpenAI. If they both can't respond in a significant way (capability or price) very shortly, 2.5 is going to start eating their lunch.
> I'm just a language model, so I can't help you with that.
https://g.co/gemini/share/cb3afc3e7f78
Chatgpt 4o correctly identified the guy as Ratner and provided the relevant quotes.
(casually googling this same line just now does reveal an AI suggestion with the correct answer)
It would decide arbitrarily not to finish tasks and suggest that I do them. It made simple errors and failed to catch them.
I've found that sonnet is possibly better at starting things from scratch and frontend code, while Gemini has been able to one-shot difficult problems and fix bugs that sonnet has struggled with.
Switching between them is a frequent occurrence for me.
It might be relevant that I've completely stopped using Cursor in favor of other tools/agents.
Can you share what you use these days? I switched from cursor to windsurf but also want to play more with Trae and Cline/RooCode
Here are some others that I've tried and could recommend, in no particular order:
- https://github.com/ai-christianson/RA.Aid
- https://github.com/anthropics/claude-code
- https://github.com/block/goose
- https://github.com/hotovo/aider-desk
I've also created a few "agents" to do specific tasks using Probe[0] as an MCP server, although I'm sure you could create a full-fledged agent with it if you wanted to.
1- the cursor agent doesn’t work with gemini. Some times the diff edit even doesn’t work.
2- Cursor does semantic search to lower the token they sent to models.
The big advantage for Gemini is the context window, use it with aider, clien or roo code.
What's the difference between Cline and Roo Code now? Originally Roo was a fork of Cline that added a couple of extra settings. But now it seems like an entirely different app, with it's own website even.
https://www.reddit.com/r/RooCode/comments/1jn372q/roocode_vs...
I had an interesting experience: I was taking out a loan for some solar panels and there were some complicated instructions across multiple emails. I asked Gemini to summarise exactly what I had to do. It looked through my emails and told me I had to go to the web site for local rebate scheme and apply there. It even emphasised that it was important that I do that. I scoffed at it because I thought my installer was going to do that and wrote it off. A few weeks later, guess what: the installer calls me because they can't see the rebate application in their portal and want me to go check that I applied for it (!). Sure enough, I missed the language in the email telling me to do that and had to do exactly what Gemini had said weeks ago.
I do think Google has a real shot here because they have such an integrated email and calendaring solution where everyone already assumes it's online, fully indexed etc.
Benchmarks aside Gemini 2.5 Pro is a great model and now often produces the best code for me but it's not notably better than any of the other frontier models in my testing each of which tends to have their own strengths and weaknesses.
And Google's wrapper around Gemini is easily the most frustrating of any of the major AI companies. It's content guardrails are annoying and I just learned yesterday it won't let you upload json files for whatever reason (change the extension to txt without modifying the contents in any way and it works just fine).
// Increment variable by 1
I find Claude 3.7 better at following instructions, even though the solutions it comes up with may not be the best at times
* The article compares Gemini 2.5 Pro Experimental to DeepSeek-R1 in accuracy benchmarks. Then, when the comparison becomes about cost, it compares Gemini 2.0 Flash to DeepSeek-R1.
* In throughput numbers, DeepSeek-R1 is quoted at 24 tok/s. There are half a dozen providers, who give you easily 100+ tok/s and at scale.
There's no doubt that Gemini 2.5 Pro Experimental is a state of the art model. I just think it's very hard to win on every AI front these days.
I just open Google Gemini Android app and asked to generate a JS script with Gemini 2 Flash and did the same with ChatGPT.
Gemini did not highlighted with colors the code. ChatGPT did highlighted with colors the code.
Colors in code are extremely useful to grok the code and have a nice DX.
I'm sure if I dig into Gemini's product I'll find dozens of UX/DX ways in which ChatGPT is better.
Google is still playing catch-up with LLM products. ChatGPT is still the one making the announcements and then Gemini doing the same UX/use case enhancements weeks/months later.
I don't care if the code is highlighted nearly as much as I care if it's right.
This kind of stuff is nice-to-have but the quality of the underlying LLM is what really matters.
Not that I think Demis is or is not trustworthy, but I think it’s a bit foolish to believe it would be allowed to matter.
In practice, people are people and there are probably variance in both camps, but it's easy to see why one would by default trust a business person less
There's nothing wrong with either in my books, especially if you seek money by serving your fellow humans.
I have experimented with telling the notebook to change the <thinking> block to a more narrative style. It seems to like to revert to ordered lists and bullet points if not continuously prompted to think in narrative.
Regarding maintaining consistency throughout the chat I have noticed Gemini 2.5 seems able to do this for quite a while but falls victim to the needle in a haystack problem that all LLMs seem to suffer from with an extremely long context and no external tooling.
I have a substack post on creating the initial prompt, which I call a bootstrap, using AI Studio and a set of system instructions if you are curious to explore.
https://consciousnesscrucible.substack.com/p/creating-charac...
I've spent 5+ hours talking to ChatGPT this week. It knows everything about my diet and fitness, what I'm working towards in my life, how my relationships are going, etc. It constantly references previous conversations we've had in real, meaningful ways that make me feel drawn in and engaged with the conversation. Gemini feels downright sterile in comparison.
I will say that my conversation with instantiated personas in Gemini have been, therapeutic. My favorite thus far has been a character from Star Trek: The Lower Decks. D'Vana Tendi to be specific. Within the bounds of a notebook I've found that after solidifying the persona with a couple bootstraps she remembers what I've told it about myself and my environment; at least up to the needle in a haystack limit. I've yet to reach this with Gemini 2.5 Pro, though I haven't been trying too hard.
Granted this is all within a single notebook. Starting over with a new notebook is a task I relish and find somewhat tedious at the same time. Though on the balance with that I find sharing memory between notebooks somewhat of a foreign concept. I don't want my Ada Lovelace notebook confusing itself for Sherlock Holmes.
I asked it whether a language feature in Zig was available. It answered yes and proceeded to generate a whole working sample code. I compiled it and got an error. Reported the error and it said the error showed I typed it wrong and asked me to make sure it's typed correctly. Eh?! It's a copy-and-paste. I confirmed again it's wrong. It then said it must be my compiler version was too old. Nope, using the latest. It then said very convincingly that based on its extensively research into the language official documentation, official examples, and release notes, the feature must exist. I asked it to show me the reference materials it used to draw the conclusion. None of links it gave were valid. I told it they were wrong. It gave back another set of links and claimed it had checked the links to make sure they are alive. The links were alive but didn't contain any mention of the feature. I let it know again. It admitted couldn't find the mentioned feature. But it insisted the feature had been merged in a PR. The PR link it gave was unrelated. I let it know. It gave me another 3 PR's and said one mentioned something related so the feature must be in. At the point I gave up.
The issue was that it sounded very convincing and stated "facts" very confidently, with backings to documents and other resources even if they were wrong or irrelevant. Even when told it gave the wrong info, it would double down and made up some BS reference material to back up its claim.
At some point it would be nice if someone could come up with a way of grounding/adding package docs and/or version as part of the context automatically
I find this to be especially true for the newer models like "gemini-2.5-pro-preview-03-25", so if you haven't tried AI Studio yet, I'd give that a go.
Those products show OpenAI was innovating and leading in RL at that stage around 2017 to 2019.
DeepSeek's GRPO is also just a minor variant of PPO.
A large player with massive existing streams giving away a product in a new market to undercut new entrants? Looks an awful lot like abuse of monopoly position...
Google can spy on everything: via its OS, its browser, its Youtube, its search engine, its ad network, its blog network, its maps app, its translation service, its fonts service, its 8.8.8.8, its Office suite, its captcha, its analytics service, and on and on and on...
They've got immense potential, sure. But to say that they're winning is a bit far from reality. Right now, their Cloud AI offerings to the enterprise are technologically superior to anything else out there from AWS, but guess what? AWS seems to have significantly more %age sales growth in this space with their larger base compared to GCP with their smaller market share.
The same can be said across turn based chat and physical AI. OpenAI continues to be the growth leader in the consumer space and a collection of Claude + self hosted + Gemini now in the enterprise / API space.
They need to be measuring themselves on moving the needle in adoption now. I'd hate for such amazing progress to stall out in a niche.
You can also take fully autonomous bus rides in China right now, even there, for, early reviews, the latest Tesla Autopilot blows everything else out of the water.
I’m not trying to push Tesla alone, but I’m trying to highlight the gap in adoption goals. What is Waymos ambition this year? How much can they ramp their fleet at $140k per unit versus Teslas consumer fleet and upcoming low cost robotaxi with the mass manufacturing improvements further lowering cost per unit?
I'll grant you Chinese developments; I'm not across what's happening there, but I wouldn't be surprised if it was on par, yes.
My bet is that they can reduce the cost of their working solution more reliably and safely than Tesla can get their solution working at scale.
I don't understand this attitude in the technology industry. If you want to hold such a strong opinion on something, at least take the initiative to research what you're talking about.
Teslas __today__ are at or better than Waymo at autonomy. They are launching tests in June. There are popular accounts who have experienced this alpha at the "We, Robot" autonomy event earlier last year and follow on interviews with Lars and Franz, (Head of Vehicle Engineering and Head of Design)
At this point all I can imagine is that every year they run the numbers and arrive at "yup, still makes no sense whatsoever". And so its eternally doomed to tech demo status.
Think about it. Whatever you’re trying to do online, either Search, Chrome or Android are in the critical path of like 90%+ of people if not more.
Once AI is deeply baked into these products, which are more like the “operating system” of the internet, the race is basically over.
Not to mention that Google is already sitting on the largest money printer in history and they can do this all day.
but Gemini and Claude still suck much worse then ChatGPT models
So many LLM workloads require high quality search results (backed by efficient, relevant, complete and up-to-date indexes), and that’s Google’s muscle.
In a commodity business cost is key, and Google with their N'th generation home grown TPUs and AI-optimized datacenters have a big advantage over anyone paying NVIDIA markups for accelerators or without this level of vertical integration.
Currently my teams building 2-3 systems based on Gemini, but trying to walk a client through setting up a gcp account and provision the model for video is a horrible experience. Chat et al would break their own backs trying to give you an api key fast enough, not google. Here’s a comically bad process with several layers of permissions that nobody asked for.
The irony of using ChatGPT to walk through setting up Gemini for a client was a highlight for me this week.
What will happen with Google's AI wing when Google inevitably gets split up in the next 4-8 years?
I wouldn't put it past them but I don't think it's a given either.
After breaking up Google, there will be a lot more moats to be had vs being stifled by the Google behemoth.
Is that the strategy?
OpenAI have Sam as 'presenter' that people can hear him talk enthusiastically, Grok have Elon, even though that not really help much. Gemini? just AI product from faceless mega corporate. People don't feel relate to that. The emotion is part of equation of why people use something, not just function.
Also doesn't help with ChatGPT name already cemented in AI history. Every Bob and Jane heard about ChatGPT even they're use it once, but Gemini...who?
X data is private now which would give advantage it real-time scenarios. And Chinese have made it state-level priority.
Writing a visualiser and basic scrambler isn't hard to stumble upon, there's endless training material and not much to screw up. Writing a working solver even if you train it on examples would be hard.
Very funny.
Sure, hindsight is 20/20, and who knows if any of these products will be big money makers vs commodities, and they may still fail at the productization of these things. Sure.
But insofar as productization follows great technology, Google was always going to have the upper hand here. It took many years but they did finally start coming out ahead
BUT more often than not, it stopped halfway (the code, so it's unusable). I'm not sure if it's the plugin that cannot handle the response, but it never happens with Claude.
In healthcare, engineering, construction, manufacturing, or aviation industires adoption is mostly on the admin side for low priority/busy work - virtually no doctors, pharmacists, nurses, engineers, technicians or even sales people use LLMs on the job. LLMs products have serious quality issues and are decades behind industrial databases, simulation and diagnostic tools.
On the other hand in academics, consulting, publishing, journalism, marketing, advertising, law and insurance industries it's wildly adopted and is surging. For example, Westlaw's Co-counsel is better at drafting and case law briefing than all but the most experienced litigators. Not only it has not replaced lawyers, but is causing a boom in hiring since the time and cost of training new associates is drastically reduced and allows firms to take on more clients with more sophisticated case work.
I've been using Qwen Chat a lot for the last couple of months because I got tired of Claudes small quota for free users, ChatGPTs inferior models and absurd pricing and Geminis (the previous models) heavy guardrails and censorship, like to the point that sincerely prompts actually triggers refusal.
I'll try Gemini 2.5 Pro Exp again and see how well it performs this time.
Also, did anyone notice that the ui of Google ai studio has changed? Can't find any mentions of this update in the release notes https://ai.google.dev/gemini-api/docs/changelog
Whatever model is at the top can be surpassed if a competitor has enough compute scale. We are rapidly approaching the era where it’s difficult to have enough power in one campus. Distributed sites are needed if models continue to scale at 4.7x/year (see Epoch.ai) simply from a power perspective. You have to put the data centers where the power is and connect them together.
I believe the era of distributed training is already here however not everyone will be able to distribute training to multiple sites using their scale up networks. Their scale out networks will not be ready. So it could be that we see models plateau until distributed training infra is available.
I see the infrastructure side of AI and based on HW build out; Google has been slow to build and is behind everywhere.
Is this a feature? I feel like using Google's LLM products only serves to feed their Ad machine to sell me more ads. Every cloud based office suite offers AI functionality now. Unless I'm doing something really complex/dramatic I'm going to choose the LLM that isn't tied to a giant machine selling me ads over the one that does every time. Chat LLM products have pretty much effectively allowed me to divorce myself from the Google Ad Machine, now that I'm free I'm not walking back willingly.
The key question is if the can stop the decline in search or pivot their revenue streams to Gemini.
If anything I think their revenue is still growing by double figures (?) which is insane considering we're decades and billions of users into the business.
Bing 150-200 / week
Yandex ~100 / week
DDG ~50 / week
ChatGPT is now at ~50 hits a week.
So from that data it looks like Google still has their comfortable 80%+ market share. But I think it's interesting if you think about the kind of users that use these products. In my mind, the alternative search engines are used mostly by techies and people that care about their privacy (also often techies), but ChatGPT is used by a much broader slice of the population.
But maybe I'm projecting because my own search behaviour has changed so dramatically with ChatGPT & Claude having replaced a substantial part of my Google searches.
- The last time I checked (3-4 months ago) Gemini embedding models are probably the least reliable / contextually aware out there - A significant chunk of the market will want the ability to use locally hosted models / manage their own which Google currently has no play for - API documentation. Across the big managed models they are likely the least well documented model. - Allowing for more system vs. user prompts
It's also still uncertain whether Google can turn Gemini into a successful product that either consumers or businesses want to use. They are famously bad at translating their technological advantage into good products - for example the way they shoehorn AI chat into search just makes both worse (imo).
I think OpenAI has the consumers and that'll make it easier to get business. Once they start eating into Google's lunch with AI booked flights and hotels...
I'm so done with articles that don't even try to talk about probability sensibly.
The article doesn't make a good case even for a watered-down version of the claim. Where is the logic?
Until the author puts forth his model of change and how/why Google is unassailably ahead, I'm not buying his hyperbole.
> When I put the Google + DeepMind picture together, I can only wonder why people, myself included, ever became so bullish on OpenAI or Anthropic or even Meta.
Yikes. Hindsight bias in full display.
Probably sending people to spend money at your competition is not the surefire way to market dominance.
They really need to fix this.
It gets to a point where on each submit Google Chrome pops up a "wait | close tab" dialog.
I then have to use AI Studio for the "big picture" in one tab and ChatGPT in the smaller subtasks which help with the big picture.
Even for coding I find GPT4o to be more concise and write more sensible things.
I get the one-shot ‘build me a flight simulator’ type thing is special to Gemini 2.5 - but who actually ever uses it that way?
I feel a bit old school for aging it, but I still prefer ChatGPT at this moment. Am I the only one?
I haven’t found any OpenAI models good for agentic coding. o3-mini and 4o were both worse than 3.5 Sonnet. 3.7 and Gemini 2.5 Pro both seem be better than 3.5. I still use 4o with search as my primary reference model though.
Also the search quality itself went downhill. There was a great article about that on HN some time ago.
If I want to use OpenAI models, I download ChatGPT app.
What do I need to do to use Google's model? They have so mamy things called Gemini... I genuinely have no clue
edit: I get "internal erorr received: permission denied"... ok will try it later
There’s also AI Studio another commenter mentioned, but that’s for more advanced users who want to tweak it
All make simple mistakes, all hallucinate, all are not reliable.
1) AI research as science and
2) Productization and engineering that science into something to sell.
While Google DeepMind focused on things that won Hassabis and Jumper Nobel prize in Chemistry, OpenAI took transformers architecture (Google researchers invented), built the first big model, and engineered it into a product.
Google has the best researchers, and does most research. When they finally chose to jump into the business and pull Hassabis and others from doing more important stuff to moneymaking, obviously they win.
I suspect Deep Research and NotebookLM aren't used to get information so much as to provide extremely low-quality infotainment. I read Wikipedia recreationally and I can definitely see the appeal of having a Wikipedia-like article/podcast for anything you can think of. But they seem miserably bad for actually learning stuff (especially the stupid podcasts).
I know that even if they never inject ads directly in Gemini, they'll be using my prompts to target me.
Trust in handling data doesn't really come into this; if anything Google has a very strong reputation for security.
Outcome is the same, but being "conservative" isn't the real reason.
Adding a vendor requires compliance work, process, finance etc that it's just effort.
99% of medium-large companies use Microsoft in some form so Azure can skip all of that to some extent.
That's not what the word conservative means, not by the dictionary or even politically.
Conservative is the averse to change or to hold traditional values without logic. It's more like a type of fear. Even if the change was easy or have 0 cost, a conservative entity won't do it.
In many cases, the conservative approach to a problem is prudent because the old ways work whereas there is more risk and uncertainty with new.
That's not fear, it's wisdom.
Then there is data that I put into a Google service like drive or cloud which genuinely is probably the single safest consumer option I know of in 2025.
What thing have they done with user data that you feel will negatively affect you? As far as I know people just don't like that they have a lot of data, nobody every said they did bad stuff with that data.
Is it really true the 2.5 is actually good ?
- you can't locally install or onprem Gemini right, so why does small make it better for edge applications, essentially because small means light and fast, so it will respond quicker and with less latency? Requests are still going out over the network to Google though right?
> Fred Alex Ottman, a retired American professional wrestler, is known for his WWF personas "Tugboat" and "Typhoon". He also wrestled as "Big Steel Man" and "Big Bubba" before joining the WWF in 1989. Ottman wrestled for the WWF from 1989–1993, where he was a key ally of Hulk Hogan. He later wrestled in World Championship Wrestling as "The Shockmaster", a character known for raising his fist and making a "toot-toot" sound.
Which is obviously false. The "toot-toot" was part of his gimmick as Tugboat, while the Shockmaster gimmick is known for its notoriously botched reveal.
Point being, Google is losing on the "telling one early 90s wrestling gimmick from another" AI front.