This article doesn't mention TPUs anywhere. I don't think it's obvious for people outside of google's ecosystem just how extraordinarily good the JAX + TPU ecosystem is. Google several structural advantages over other major players, but the largest one is that they roll their own compute solution which is actually very mature and competitive. TPUs are extremely good at both training and inference[1] especially at scale. Google's ability to tailor their mature hardware to exactly what they need gives them a massive leg up on competition. AI companies fundamentally have to answer the question "what can you do that no one else can?". Google's hardware advantage provides an actual answer to that question which can't be erased the next time someone drops a new model onto huggingface.
> 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
Assuming that DeepSeek continues to open-source, then we can assume that in the future there won't be any "secret sauce" in model architecture. Only data and training/serving infrastructure, and Google is in a good position with regard to both.
Google is also in a great position wrt distribution - to get users at scale, and attach to pre-existing revenue streams. Via Android, Gmail, Docs, Search - they have a lot of reach. YouTube as well, though fit there is maybe less obvious.
Combined with the two factors you mention, and the size of their warchest - they are really excellently positioned.
Making your own hardware would seem to yield freedoms in model architectures as well since performance is closely related to how the model architecture fits the hardware.
... except that it still pretty much requires Nvidia hardware. Maybe not for edge inference, but even inference at scale (ie. say at companies, or governments) will still require it.
TPUs aren't necessarily a pro. They go back 15 years and don't seem to have yielded any kind of durable advantage. Developing them is expensive but their architecture was often over-fit to yesterday's algorithms which is why they've been through so many redesigns. Their competitors have routinely moved much faster using CUDA.
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.
Not sure how familiar you are with the internal situation... But from my experience think it's safe to say that TPU basically multiplies Google's computation capability by 10x, if not 20x. Also they don't need to compete with others to secure expensive nvidia chips. If this is not an advantage, I don't see there's anything considered to be an advantage. The entire point of vertical integration is to secure full control of your stack so your capability won't be limited by potential competitors, and TPU is one of the key component of its strategy.
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.
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.
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.
The one thing they lack that OpenAI has is… product focus. There’s some kind of management issue that makes Google all over the shop, cancelling products for no reason. Whereas Sam Altmans team is right on the money.
I've used Jax quite a bit and it's so much better than tf/pytorch.
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?
There's basically a difference in philosophy. GPU chips have a bunch of cores, each of which is semi-capable, whereas TPU chips have (effectively) one enormous core.
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.
Way back when, most of a GPU was for graphics. Google decided to design a completely new chip, which focused on the operations for neural networks (mainly vectorized matmul). This is the TPU.
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"?
TPUs (short for Tensor Processing Units) are Google’s custom AI accelerator hardware which are completely separate from GPUs. I remember that introduced them in 2015ish but I imagine that they’re really starting to pay off with Gemini.
Amazon also invests in own hardware and silicon -- the Inferentia and Trainium chips for example.
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.
I think Amazon, Meta have been trying on inference hardware, they throw their hands up on training; but TPUs can actually be used in training, based on what I saw in Google’s colab.
The problem is always their company never the product. They had countless great products. You cant depend on a product if the company is reliably unreliable enough. If they don't simply delete it for being expensive and "unprofitable" they might initially win, eventually, like search and youtube, it will be so watered down you cant taste the wine.
I am the author of the article. It was there since the beginning, just behind the paywall, which I removed due to the amount of interest the topic was receiving.
And yet google's main structural disadvantage is being google.
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.
> 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 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.
> Google is too risk averse to do anything, but completely doomed if they don't cannibalize their cash cow product.
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.
This would be like claiming in 2010 that because Page Rank is out there, search is a solved problem and there’s no secret sauce, and the following decade proved that false.
Do we have insights on whether they knew that their business model was at risk? My understanding is that OpenAI’s credibility lies in seeing the potential of scaling up a transformer-based model and that Google was caught off guard.
I think what may save Google from an Innovator's Dilemma extinction is that none of the AI would-be Google killers (OpenAI etc.) have figured out how to achieve any degree of lock-in. We're in a phase right now where everybody gets excited by the latest model and the switching cost is next to zero. This is very different from the dynamics of, say, Intel missing the boat on mobile CPUs.
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.
Google is winning on every front except... marketing (Google has a chatbot?), trust (who knew the founding fathers were so diverse?), safety (where's the 2.5 Pro model card?), market share (fully one in ten internet users on the planet are weekly ChatGPT users), and, well, vibes (who's rooting for big G, exactly?).
But I will admit, Gemini Pro 2.5 is a legit good model. So, hats off for that.
My experience with their software has been horrible. A friend was messing around with Gemini on my phone and said my name is John, and it automatically saved that to my saved info list and always called me John from then on. But when I ask it to forget this, it says it can't do that automatically and links me to the Saved Info page, which is a menu they didn't implement in the app so it opens a URL in my browser and asks me to sign into my Google account again. Then a little toast says "Something went wrong" and the saved info list is empty and broken. I tried reporting this issue half a year ago and it's still unresolved. Actually the only way I was ever able to get it to stop calling me John is to say "remember to forget my name is John" in some way that it adds that to the list instead of linking me to that broken page
"Hello John, I notice your username on HN has not been updated. I will make that change for you in 2 hours, from a2128 to john2128. If you want to keep your current username, please follow steps in our help me discord channel"
Google is also terribly paranoid of the LLM saying anything controversial. If you want a summary of some hot topic article you might not have the time to read, Gemini will straight up refuse to answer. ChatGPT and Grok don't mind at all.
I noticed the same in Gemini. It would refuse to answer mundane questions that none but the most 'enlightened' could find an offensive twist to.
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.
I think that's the "trust" bit. In AI, trust generally means "let's not offend anyone and water it down to useless." Google is paranoid of being sued/getting attention if Gemini says something about Palestine or drawns images like Studio Ghibli. Meanwhile users love to these topics and memes are free marketing.
Not a fan of Google, but if you use Gemini through AI studio with a custom prompt and filters disabled it's by far the least censored commercial model in my experience.
I have the same experience in the web UI. Asking for that famous Obama chili recipe gets a refusal. But when I use the API, I can dial back the safety settings to the point where things work much more smoothly.
>Google is also terribly paranoid of the LLM saying anything controversial.
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.
Seems like a feature. Last thing we need is a bunch of people willing to take AI at it's word making up shit about controversial topics. I'd say redirecting to good or prestigious source is probably the best you can do
I wouldn't even say Gemini Pro 2.5 is the best model. Certainly not when you do multimodal or function calling, which is what actually matters in industry applications. Plain chatbots are nice, but I don't think they will decide who wins the race. Google is also no longer in the mindset to really innovate. You'll hear surprisingly similar POVs from ex-Googlers and ex-OpenAI guys. I'd actually say OpenAI still has an edge in terms of culture, even through it fell deep.
I did some experiments with Gemini Pro 2.5 vs Sonnet 3.7 for coding assistants, and, at least as far as code quality and ability to understand complexities in existing codebase goes, Gemini is noticeably stronger.
I have found function calling and ‘roll my own agents’ work much better now with Gemini than they did late last year, but I also do a lot of function calling experiments with small open models using Ollama - much more difficult to work with to get a robust system.
You really hit the nail on the head with trust. Knowing the power of these AIs and how absolutely little I trust Google, I’d never tell trust Gemini with the things I’ll say to ChatGPT.
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?
My hesitancy to adopt Gemini, despite being a heavy GCP and workspace user, is I kinda lost trust when trying to use their earlier models (I don't even remember those models' names). I just remember the models were just so consistently bad and obviously hallucinated more than 50% of the time.
Maybe Gemini is finally better, but I'm not exactly excited to give it a try.
Well, Google is also very well placed to integrate with other products that have big market share.
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.
Microsoft has been integrated Copilot in their Office products. In fact, they don't even call it Office any more. Guess what? If you ever had first hand experience with them, they are absolutely a dumpster fire. (Well, maybe except transcription in Teams meeting, but that's about it.) I used it for 5 minutes and never touch it again. I'll be very impressed if that's not the case with Google.
Exactly. Google may have a lead in their model, but saying they are "winning on every front" is a very questionable claim, from the perspective of everyday users, not influencers, devoted fans or anyone else who has a stake in hyping it.
I look more to Google for efficient and inexpensive LLM APIs, and in a similar way to Groq Cloud for inexpensive and fast inferencing for open models.
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.
I had to stop using Gemini 2.5 because the UI peaks my MPB cpu at max and I can’t type my prompt at more than a character every 2 seconds. I can’t even delete my chats lol. Anyone else?
On deleting chats, I accidentally discovered that AI Studio creates a 'Google AI Studio' folder on your Google Drive with all the links to chats. If you delete the 'link' from there, it will disappear in AI Studio...interesting UX :-)
It might be worth throwing in an analogy to windows PCs vs Mac vs Linux. G appeals to a subset of the market at the end of the day, being “best” does not mean everyone will use it.
I would like to think they just let other companies have the first mover advantage on chatbots because it only disrupts Google in their search business, which was already pretty far gone and on the way out. Where is AI actually going to change the world? Protein folding, robotics, stuff that the public doesn’t hype about. And they looked at the gold rush and decided “let’s design shovels”. Maybe I’m giving them too much credit but very bullish on Google.
They "lost from a position of strength" in that they had they had the most public-cloud like data centers and started thinking about selling that later than they should have. Bard/Gemini started later than chatgpt , but there's not really a moat for this LLM stuff, and Google started moving a lot earlier relative to GCP vs Amazon.
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.
Trust is important, and Google has a big rep for killing its projects. As well as making the most moronic braindead decisions in handling what they don't kill off.
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?
I’m scared they’re going to kill it off. Every good idea they’ve had in the last 20 years has been killed off. Even Fuchsia/Zircon, which should have supplanted Android a full decade ago.
Winning =/= won. The point is that they are improving on many fronts. If they were already recognized as THE leader there would be no point in making a HN post about it.
As an Ex-OpenAI employee I agree with this. Most of the top ML talent at OpenAI already have left to either do their own thing or join other startups. A few are still there but I doubt if they'll be around in a year. The main successful product from OpenAI is the ChatGPT app, but there's a limit on how much you can charge people for subscription fees. I think soon people expect this service to be provided for free and ads would become the main option to make money out of chatbots. The whole time that I was at OpenAI until now GOOG has been the only individual stock that I've been holding. Despite the threat to their search business I think they'll bounce back because they have a lot of cards to play. OpenAI is an annoyance for Google, because they are willing to burn money to get users. Google can't as easily burn money, since they already have billions of users, but also they are a public company and have to answer to investors. But I doubt if OpenAI investors would sign up to give more money to be burned in a year. Google just needs to ease off on the red tape and make their innovations available to users as fast as they can. (And don't let me get started with Sam Altman.)
> there's a limit on how much you can charge people for subscription fees. I think soon people expect this service to be provided for free and ads would become the main option to make money out of 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.
> No guarantees that Google discover a meaningful business model here...
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.
I don’t think “AI” as a market is “winner-takes-anything”. Seriously. AI is not a product, it’s a tool for building other products. The winners will be other businesses that use AI tooling to make better products. Does OpenAI really make sense as a chatbot company?
"A surprising number of people pay for the ChatGPT app and/or competitors."
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.
Google aren’t interested in <1bn USD businesses, so it’s hard for them to build anything new as it’s pretty guaranteed to be smaller than that at first. The business equivalent of the danger of a comfortable salaried job.
'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
>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.
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.
Absolutely agree Microsoft is better there - maybe that's why Google hired someone from Microsoft for their AI stuff. A few people I think.
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.
> 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
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.
"The actual business models, at this point, aren't even known."
"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.
The business model question applies to all of these companies, not just Google.
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.
Contextual advertising is a known ad business model that commands higher rates and is an ideal fit for LLMs. Plus ChatGPT has a lot of volume. If there’s anyone who should be worried about pulling that off it’s Perplexity and every other small to mid-sized player.
> I think soon people expect this service to be provided for free and ads would become the main option to make money out of chatbots.
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.
I agree with you. There is also a move toward "agents", where the AI can make decisions and take actions for you. It is very early days for that, but it looks ike it might come sooner than I had though.
That opens up even more potential for influence on financial decisions (which is what adtech wants) - it could choose which things to buy for a given "need".
You're assuming ads would be subtly worked into the answers. There's no reason it has to be done that way. You can also have a classic text ads system that's matching on the contents of the discussions, or which triggers only for clearly commercial queries "chatgpt I want to eat out tonight, recommend me somewhere", and which emits visually distinct ads. Most advertisers wouldn't want LLMs to make fake recommendations anyway, they want to control the way their ad appears and what ad copy is used.
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.
The ads angle is an interesting one since that's what motivates most things that Google and Meta do. Their LLMs' context window size has been growing, and while this might the natural general progression with LLMs, for those 2 ads businesses there's pretty straight paths to using their LLMs for even more targeted ads. For example, with the recent Llama "herd" releases, the LLMs have surprisingly large context window and one can imagine why Meta might want that: For stuffing in it as much of the personal content that they already have of their users. Then their LLMs can generate ads in the tone and style of the users and emotionally manipulate them to click on the link. Google's LLMs also have large context windows and such capability might be too tempting to ignore. Thinking this, there were moments that made me think that I was being to cynical, but I don't think they'll leave that kind of money on the table, an opportunity to reduce human ad writers headcount while improving click stats for higher profit.
EDIT: Some typo fixes, tho many remain, I'm sure :)
When LLMs are essentially trying to sell me something, the shit is over.
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.
The real threat to Google, Meta is that LLMs become so cheap that its trivial for a company like Apple to make them available for free and include all the latest links to good products. No more search required if each M chip powered device can give you up-to-date recommendations for any product/service query.
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.
Google offering Gemini 2.5 Pro for free, enough to ditch OpenAI, reminds me of an old tactic.
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.
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 pay for ChatGPT, Anthropic and Copilot. After using Gemini 2.5 Pro via AI Studio, I plan on canceling all other paid AI services. There is no point in keeping them.
I believe it. This is what typically happens. I would go to AWS re:invent and just watch people in the audience either cheer or break down as they announced new offerings wash away their business. It's very difficult to compete in a war of attrition with the likes of Google, Microsoft, and Amazon.
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.
Based on his interview with Joe Rogan, he has absolutely no imagination about what it means if humans actually manage to build general AI. Rogan basically ends up introducting him to some basic ideas about transhumanism.
To me, he is a finance bro grifter who lucked into his current position. Without Ilya he would still be peddling WorldCoin.
What cards has google played over the past three years such that you are willing to trust them play the "cards at hand" that you alleged that they have? I could think of several things they did right, but I'm curious to hear which one of them are more significant than others from someone I think has better judgement than I do.
I get your perspective, but what we're seeing looks more like complex systems theory, emergent behavior, optimization, new winners. If models become commoditized, the real value shifts to last-mile delivery: mobile, desktop, and server integration across regions like China, Korea, the U.S., and Europe.
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.
‘think soon people expect this service to be provided for free’
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.
It's interesting to hear your perspective as a former OpenAI employee. The point about the sustainability of subscription fees for chatbots is definitely something worth considering. Many developers mention the challenge of balancing user expectations for free services with the costs of maintaining sophisticated AI models. I think the ad-supported model might become more prevalent, but it also comes with its own set of challenges regarding user privacy and experience. And I agree that Google's situation is complex – they have the resources, but also the expectations that come with being a public company.
> "[Google is] a public company and have to answer to investors"
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.
Microsoft CoPilot (which I equate with OpenAI ChatGPT, because MS basically owns OpenAI) already shows ads in it's chat mode. It's just a matter of time. Netflix, music streamers, individual podcasters, YouTubers, TV manufacturers – they all converge on an ad-based business model.
Google has so many channels for ad delivery. ChatGPT is only competing against Google Search, which is arguably the biggest. But dont forget, Google has YouTube, Google Maps, Google Play, Google TV and this is before you start to consider Google's Ad Network (the thing where publishers embed something to get ads from Google network).
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.
> 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.”
Open AI don't always have the best models (especially for programming) but they've consistently had the best product/user experience. And even in the model front, other companies seem to play catchup more than anything most of the time.
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.
AI startups were easy cash grabs until very recently. But I think the wave is settling down - doing real AI startup turned out to be VERY hard, and the rest of the "startups" are mostly just wrappers for OpenAI/Anthropic APIs.
I don't know what you did there, but clearly being ex OpenAI isn't the intellectual or product flex it is: I and every other smart person I know still use ChatGPT (paid) because even now it's the best at what it does and we keep trying Google and Claude and keep coming back.
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 use a service where users can choose any frontier model, and OpenAI models haven't been the most used model for over half a year - it was sonnet until gemini 2.5 pro came out, recently.
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.
Don't think it's a flex, I think it's useful context for the rest of their comment.
> 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.
It is sort of funny to me how the sentiment about whoever seems to be leading in ML changes so frequently (in particular here on HN.) A couple months ago it felt like people were sure that Google completely fucked it up for themselves (especially due to the fact that they invented the transformer but didn't productize it themselves at first.)
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.
Personally, I don't really think there's a team at Google, nor at OpenAI, paying for "astroturfing" on sites like HN.
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?
Yeah... I wish there were laws that would require disclosure of such behavior. Might be tricky to implement though, and probably contradicts the interests of politicians.
AI is changing fast! And to be fair to the model companies, they have been releasing products of (mostly) increasing quality.
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.
The sentiment changes this fast because SOTA changes this fast. E.g. Google models were objectively crappy compared to OpenAI, but Gemini 2.5 really turned the tables (and I'm not talking about synthetic benchmarks here but real world coding).
The state of affairs with local models is similarly very much in flux, by the way.
Several people have suggested that LLMs might end up ad-supported. I'll point out that "ad supported" might be incredibly subtle/insidious when applied to LLMs:
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
No one would ever need to/be able to know if the output:
"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.
LLMs should be legally required to act in the interest of their users (not their creators).
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.
Would be illegal in Germany ('Schleichwerbung') and perhaps the EU?
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.
For me ads on web are acceptable as long as they are clearly distinguished from the content. As soon as ads gets merged into content, I'll be unhappy. If LLM would advertise something in a separate block, that's fine. if LLM augments its output to subtly nudge me to a specific brand which paid for placement, that's no-no.
Free might as well be a curse-word to me, and I'm not alone. I'm old enough to have experience in pre-internet era magazines, and the downgrade in quality from paid publications to free ones has been quite substatial.
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.
I'm not convinced this is any worse than searching for results or reviews and being directed to content that is affiliate supported (or astroturfed by companies). Humans already do this sort of subtle nudging and lots of people position themselves as unbiased. So many blogs are annoying "buried lede" advertising where the article seems vaguely useful until you realise that it's just a veiled attempt to sell you something. Virtually every reviewer on YouTube seems obliged to open with "my thoughts are my own, the company doesn't get to edit my review, etc."
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.
To put on my techno-optimist hat, some specific searches I make already thinking please, please sell me something and google's results are horribly corrupted by SEO.
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.
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.
I was a loyal Claude user until I decided to try Gemini 2.5. "After all", I thought, "I already use a Pixel phone, so it's integrated with Android. And with Google Drive. And I can get it through my Google One subscription."
And now that I'm on it, I don't think I'm going back. Google did it again.
2.5 Pro Experimental and Deep Research showed up in the Gemini app for me today days after it was available on web so it seems to be different roll outs for different platforms.
Out of interest: Using Gemini on your phone, integrated and all, obviously reduces friction, but would you say convenience is the only reason for you not going back or do you feel Gemini is a real improvement as well?
The improvement in Gemini 2.5 is real, but I wouldn't say it's miles away from Claude 3.7. The fact that web browsing still isn't in Claude in Europe bothered me. It's many little things.
> Gemini 2.5 Pro in Deep Research mode is twice as good as OpenAI’s Deep Research
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.
Matches also my experience that openai fell behind with their deep search product. And that deep search is basically the top tier benchmark for what professionals are willing to pay. So why should i shell out 200 dollar for an openai subscription when google gives me a better top-tier product for 1/10th of the price openai or anthropic are asking. Although i assume google is just more willing to burn cash in order to not let openai take more market share which would get them later on soo more expensive (e.g. iphone market share, also classic microsoft strategy).
It may actually be affordable for Google to charge $20 vs OAI's $200. Google already has an extensive datacenter operation and infrastructure that they're amortizing across many products and services. AI requires significant additions to it, of course, but their economy of scale may make a low monthly sub price viable.
Great stuff. My prompts are falling behind after seeing what you are doing here.
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.)
> "produce a comprehensive analytical report exploring the conceptual and methodological intersections between Surrealist art techniques, Freudian dream analysis, and the practice of prompt engineering for AI image generation models (such as DALL-E, Midjourney, Stable Diffusion)."
yes they also have very decent free tier which is great, but keep in mind prompts "used to improve our products". We will see if free tier will stick for long or is temporary - they removed details how big free tier is but still great for testing.
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.
Readit News
[1]https://blog.google/products/google-cloud/ironwood-tpu-age-o...
> 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
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.
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.
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.
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.
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?
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.
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"?
https://en.wikipedia.org/wiki/Tensor_Processing_Unit
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.
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.
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.
I doubt this. Embedding models are no panacea even with a lot simpler retrieval tasks like RAG.
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.
Deleted Comment
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.
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.
Deleted Comment
Who is? (Genuine question, it's hard to keep up given how quickly the field moves.)
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?
Are you pretty sure that Google won't eventually buy OpenAI and thus learn everything you've said to ChatGPT?
Maybe Gemini is finally better, but I'm not exactly excited to give it a try.
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!
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.
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?
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.
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.
But not profitable yet.
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.
'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
¹ https://en.wikipedia.org/wiki/Business
² https://en.wikipedia.org/wiki/Predatory_pricing
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.
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.
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.
"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.
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.
Especially when post-tarrifs consumption is going to take a huge nosedive
How so? Amazon were the first with S3 and EC2 including API driven control.
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.
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 hope local models remain viable. I don't think ever expanding the size is the way forward anyway.
>... ads would become the main option to make money out of chatbots.
What if people were the chatbots?
https://youtu.be/1iqra1ojEvM?si=xN3rc_vxyolTMVqO
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.
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.
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.
https://ai.google.dev/gemini-api/terms#data-use-unpaid
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.
Please do.
Key Facts from "The Secrets and Misdirection Behind Sam Altman's Firing from OpenAI": https://www.lesswrong.com/posts/25EgRNWcY6PM3fWZh/openai-12-...
Deleted Comment
To me, he is a finance bro grifter who lucked into his current position. Without Ilya he would still be peddling WorldCoin.
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'
https://news.ycombinator.com/item?id=41034829
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.
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.
Deleted Comment
Dead Comment
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.
Another instantiation: people like cheap goods more than they dislike buying foreign made goods
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.
> 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.”
Why not? That's one of the reasons I visit HN instead of some random forum after all.
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
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.
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.
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 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 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.
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?
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.
The state of affairs with local models is similarly very much in flux, by the way.
Hey it's me!
An LLM-based "adsense" could:
No one would ever need to/be able to know if the output:"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.
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.
lofty ideal... I don't see this ever happening; not anymore than I see humanity flat out abandoning the very concept of "money"
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.
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.
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.
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.
https://nlp.elvissaravia.com/i/159010545/auditing-llms-for-h...
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.....
Using Gemini via Google Workspace.
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
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...
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...
Haha, what a perfect project for AI.
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.