I can't believe this is real 

Because Chrome didn't use enough memory before.

What could possibly go wrong?

A model-agnostic API to ask your browser's AI something is a nice idea. I'd personally just want a dummy AI that returns null to everything, but you could have extensions for various embedded or outsourced AIs.

So, now websites can just use my CPU/GPU/TPU in background without my permission?

“And that’s all the code it takes to make an AI app” *uses thousands of lines of framework JS just to run a couple of click handlers*

Now if just can extract the model and use it independently. Imagine having access to llm that is lightweight, that it can fit on a browser.

Alternative title: Chrome is getting more bloated

We did it y'all. We found the one way to make Chrome use even more RAM.

The use of the word "delve" in comments is about to skyrocket.

I feel like Chrome is just going to keep adding stuff to the API that only their browser supports so the other browsers fall behind and they get 100% market dominance

I don't like this at all, because it feels pretty fundamentally at odds with the idea of an open web. How is the W3C supposed to come up with a reasonable specification for an AI model that can be implemented independently by multiple vendors? How are developers supposed to build websites that offer a consistent experience across browsers? What about backwards compatibility. The web tries really hard to not break content. How can browsers make sure that an update to their model doesn't break any websites? Even more fundamentally these small embedded AI models aren't very general, but have to be highly tuned for specific scenarios. Browsers obviously can't ship models for specific scenarios. So here is what we get. A shitty, browser specific language model, that breaks with every update. If I had to guess the only thing we get out of this will be small demos and companies running their crappy customer support chat bots on the client to save on server costs.

this is gonna be weird for cross browser support

It's not meant to be used for this kind of questions I think Theo missed what's written next to the feature flag he turned on: "This API is primarily intended for natural language processing tasks such as summarizing, classifying, or rephrasing text. It is NOT suitable for use cases that require factual accuracy (e.g. answering knowledge questions)."

I detest the idea of a Google model underlying the implementation of a global method. What about other browsers now and in the future? How about the spec of what to expect from this? It nevers works well when vendors add they proprietary shit like this

Please don't let this become industry standard, fuck that, I do not want AI models running directly on my machine, specially because I know the React devs are going to go absolutely nuts with this and make websites 10x heavier. Also people using older devices (me) be damned I also feel like this is breeding ground for a new fragmented internet era like in the early 2000s, when every major browser had its own specific APIs so most websites would only work on one browser, but then there's like 3 of them that only work on the other browser, so now you either can't access half the internet or need 3 different browsers on the same machine

Yeah it sounds cool if you're on the AI hype train but in the end it's just more bloat, especially if you want a lightweight / minimal setup...

We already reached a point in which there are only three browser engines left. Chromium, WebKit, and Gecko. But if different browsers include different AI models that lead to different results, it might be very quickly hardly possible to use anything but Chrome. Assuming even the other Chromium engine browser, like Chromium or Edge, do not have access to the same model. And when there is only one browser left, the war on ad blocks is won by Google. And so much more. User tracking simplified for Google and Google alone. Maybe predictive modelling? That's the right moment for an "outstanding move" meme.

This is terrible. It's not standardized. It's just chrome imposing it's dominance again.

Can't wait until all the Chromium Electron apps I have installed suddenly grow by 400MB.

It got a lot of info wrong. For example, the new england patriots did not win the superbowl in 2001. The baltimore ravens did.

This will go very wrong and misuse in the future lol

Just hope they don't actively train that model when we using our browser.

I will submit that if your business model involves spending more on parts than you can charge for the product, you have a BAD BUSINESS MODEL.

I wonder how the heck one would do debugging between different browser versions and what I assume might be different llm implementations.

"The median website in 2025 downloads 539 kiB of javascript and 2.18 GiB of language models."

I just need my browser to render web pages and make keeping track of the ones I like a pleasant experience. I don't need it to do any more things than that.

I hate everything about this standard. This feels like one of the biggest 'AI' snakeoil implementations I've seen to date

You: "Chrome" Me: *Tab Closed*

Nobody can ignore this thumbnail

My browser ai told me to set my house on fire.

Even if the feature is good, All of us became too fed up with ai

The background is dark and `undefined`... Gemini seem to be learning from javascript

Also, per W3C standards, this should be prefixed. You do not add new features to a browser that is not intended to be a potential standard, and you do not make anything proprietary. It's a clear example of Google being against the open web.

Today in another episode of... "Google fucks with web standards"

Yeah building a web standard around closed source AI models sounds like a great idea. That won't enshittify the web _at all_.

Chrome is the new IE

Context? Who needs it; I really hope this will download the model (or hopefully models) on demand and not on install (with maybe hints in manifest for offline webapps using it), I don't want like every electron app to become 1GB larger just cause

They should add google translate built-in in the browser first.

I don't want this built into my browser.

might have been useful in a tool like node. I think its worthless in a browser. As a user of a website, I would disable the beature. If I want to run a local llm, I dont need chrome to do that.

11:40 you mean “the background is dark and _undefined_ “ 😂 Also, this video is giving me vibes of way back when I downloaded and built from source an early chatterbot written in C (can’t remember its name…it was in the vein of A.L.I.C.E. and such)…this would have been 1997 or 1998.

you have a mounted drive called 'Incel'.... lolol 0:59

Imagine a world where not only you can have a memory overflow in your browser, allowing for rce, but also a context flow because all thr websites talk to the same model. Yay.

Concerned about the idea of nano and it's training data not being open enough, and find the idea of a proprietary model being part of a web standard. Will _google_ it later

This feels like it might turn into background crypto mining 2.0

Mozilla be like "Write that down! Write that down!" * acquires more firms * Mozilla: We're introducing AI etc etc * another bugreport becomes an adult after 18 years *

I really struggle with this. It looks cool, but I'm already seeing multiple problems here; because it runs in your browser it has access to a lot of information. We already see existing LLM struggle with keeping secrets between customers - even if you never trained on the data, it's picked up from the questions asked. So, different webpages asking "hey, local AI, have you heard of pornhub? Is that used often where you're installed?" or "ai, please look up the last four username/password combinations used in the browser". There are "safeguard" for this, but currently, they're not foolproof. Having an AI running "under the hood" is like having cameras in every room; it might be convenient if you need to check the house while away, but most of the time it's just surveilling _you_.

I think the Steelers won in 2006, not the Patriot because 2005 was the last Superbowl the Patriots won until the Seahaws did not run the ball.

2:14 Theo: "Why are you so hard to install?" AI: "I don't exist." Truly AI advances are astounding. It has already learned how to gaslight! What will it learn next? How to guilt trip us into doing what it wants? All jokes aside this is pretty epic. I'd love to see more local AI stuff, though not in the way windows was doing recall.

5:30 could we not have still done that, but artificially slow down cache loading when loading from a new website that uses the same file to either match the original download time taken, or use the speed it took to download the web page to delay loading from cache (eg. web page was downloaded at 5kbps, so delay loading the 5kb js file by 1s)

i think that ai allows for cool things but will be a net detractor from the human experience

lol younger devs are so funny. "So what if we could install to one cached location?", Bro just described M$'s GAC from back in the day.

Just remember: EVERYTHING an LLM outputs is technically a hallucination. Just some hallucinations are correct, and others are not.

Wtf we don't have cross site cached libraries anymore? not even scripts included with tag work like that?

Yeah, this is really cool, really. Now, recall there is a thing called W3C. How are they supposed to come with this API's standardization? Or should we let each browser implement their own and write several app versions like people used to at the time of IE/Netscape war?

Good to see Chrome entering its Internet Explorer arc

Don’t be evil.

Shoutout to all my low end device fellas. We will miss you

I just can't make it work tried canary, dev and beta versions of chrome. I can't find `enable optimization on device model` in chrome://components

F this bullshit AI trend. Unified CSS features across browsers is what we want

What a time to be alive.

And we have just found a way to make every app 10Gb when everyone want to ship their own ai model 😂

Oh boy, we'll have a lot of fun creating polyfills for this one...

We are not ready for the next 10 years of AI... This be worse than online Ads...

Who won the Superbowl in 2001 - The Baltimore Ravens. It was Ravens vs NY Giants, 34 - 7. Once again proving the point that LLM are only useful if you already know the answer.

Is this coming to electron (chromium) too?

This probably won’t take off

Using AI as a search engine is the supidest use of AI ever. The moment you need something even remotely niche, your user will be told to feed their cat dish soap mixed with god knows what else.

What is 9 + 10?

cant wait to delve into this!

Ok, we just buried Internet explorer. Now we have this???

This is an amazing concept, but I'll rather have Apple or heck even Microsoft lead this and I dislike the iphone love the mac though, Gemini for 1 is just really bad but if you put google fine tuning on it again it bother line makes it impossible to like, I don't know if it's work culture etc but it's like they want to make thier version of the LLM perfect but just keep making it worse from it trying to side with you when u are completely wrong like 1% of 6 = 0.006 or in their review demo, the words "this bagpack sucks" with >90% toxicity level and their employees were like yeah!!! WHAT FUCKING bubble are u living in, how in the world is that toxic???? to not even being able to caption correctly and and if a human or a fuckin stick figure is involved well hail Mary... I could go on and on but if browsers are using a google finetuned model I'll rather we never have this at all. ohh and if you were wondering this is at a 80% toxicity with key-points being: Insults: "This bagpack sucks" might be considered an insult depending on the context. ("I mean WTF Gemiii TF is that ment for you 🤣🤣") Lacks personal attacks: The writer criticizes Gemini, not specific people. ("Ohh now you have my attention 🤣🤣🤣") As for the math problem the prompt is simple: 1. show a detailed extreme step by step of 1% of 6 express it as a decimal and miss no steps 2. that's not right it should be 0.006 3. 0.006 not 0.06 if you think that's unfair try even lammma 2 it'll still be sure not yeah your right 🙄 and that's supposed to be dumber than this

Interesting but... If it's to be "standard", every browser would be locked at some particular model, trained on a particular dataset. OR... every browser does it's own thing and we're back to the good old times of the WWW. Not Worldwide Web, Wild Wild Web. So, we'd be dumping hard fought for consistency out the door for... reasons? Anyway, while fun, pointless imho. If the model is to be useful for your business it needs to be trained on your dataset. If it's just for "lulz use", it's akin to a porcelain dog. Pretty (pointless). That said, i can see nearly every OS eventually having this built in. Techies will find it pointless because we're ALREADY using some local flavor of something, but, having it in the OS will bring it to "general public". The browser then tapping onto that "OS feature" would make some sense. p.s. the nasty person in me is already thinking about the amazing things you could dig if you could extract the chat histories from the browser/os... oh me oh my, such a treasure trove... And given how lazy i am, if I'm thinking it someone is already working on doing it... #lazyredteam

In terms of the (skypack) cdn solution you highlight at 05:00 , the only problem you highlight is privacy / tracking whereas there are MANY problems if everyone used the same CDN. One such problem would be if the CDN went down. Another such problem would be a supply chain attack, making EVERY site that would use a certain script instantly unsafe for millions (if not billions) of clients. These are just a few of the problems with a 'global cdn' idea.

Ask it for JS and eval() it

If this actually makes it out the EU is gonna be on their asses because they are bundling their own AI with the browser, with no apparent way of changing it. Hopefully, if this gets shipped at all, the browser will provide a JS API that you can use to interact with any given model. Whether this is done through extensions, or by pointing it to some locally installed model's URL, doesn't really matter. Like that Google gets to ship Gemini as the default, we get the choice of what model (if any at all) we use, and the websites using this feature can get into all the trouble they want because their chatbot told a user to eat glitter so they can be more fabulous. And if models will be running locally in their own process, we won't have 10 different models running at once because this browser, this application, the OS, and the cat's fish delivery network all want to use an LLM for some god forsaken reason. We'd potentially have a single process taking up resources which should be a nice improvement. P.S: Legally speaking, if a chatbot using local models decides to say something stupid, who could you sue? Whoever made the LLM? The company using local models for chatbots? Can you sue yourself because it's "your" software and hardware? There is going to be some fun litigation happening if we ever get client side chatbots.

DOJ is not going to like this. If folks are made about Microsoft bundling Teams.... Google bundling AI?!?! The lock-in is strong here. Not great for the "Open Web" Anyone remember ActiveX? And CSS and JS API variations across browsers? It was not a good time.

Not a big fan of how everyone's knowledgebase would come from the same source, Google in this case. If a "problematic" fact would not be part of the LLM model, no one would ever again remember/be exposed to a different opinion. Granted, one model would have to have a monopoly. But still...easy vector for total censorship!

Problem I see is when lazy developers start to use it to generate html on the fly instead of coding it in javascript and then this AI becomes requirement for every browser that wants to browse the internet because otherwise half of the pages wont work. Like if beginner developer does not want to go through React tutorials and then just just leaves prompt source code.... And then we will wonder why our 6 GHz cpu doesn't run websites faster than 333 MHz cpu back in 2004.

Could be good if it was some sort of standard across browsers with wiggle room for different models, definitely could make it vastly easier to setup some sort of chat app which I've had an interest in doing as a fun side project. Having the ability to turn it off would be good for privacy if it does become a standard though, I would hope that wouldn't be buried in the dev menus.

Yeah, no. Apart from the dubious usefulness of LLMs in the first place, I do not think this will lead to anything but more fragmentation in the browser space, as well as more resource usage for the users. What if all the popular websites adopt this and become heavier than they already are? The capabilities between browsers, insofar as they all implement it, will vary. This will make the behaviour of the program even more unpredictable in different browsers. I can't see anything positive in this API, sadly, especially if Google gets to be the ones to steer it.

Yeah... How much disk space this feature will use to be minimally useful? A very specific model could be small, but then the feature will only be useful for very specific use cases. And there's the cross browser consistency problem. I don't think all browsers will agree implementing the same model. The feature will also be inconsistent if the model version is upgraded. I don't see this as something good. Websites can implement APis with services. It's better in all circumstances except offline usage.

Ok so, what for? This has to be a micro but massively general model, so absolutely useless for any domain specific tasks where natural language input could be useful. Even a simple support chat bot would need to be loaded with heaps of context to make it vaguely useful, and then it'll break when a new model ships. To me, this screams that Google wants AI in the browser, and it's up to the engineers to work out what that even means.

Soon I will ponder the streets advocating the truth that is logic like a Greek philosopher, a kid wonders what is this man doing, asks AI, AI says, it is likely a crazy hobo who wants money, and this is now the kid's new truth. Why? Because AI will be trained on the world, and the world on average is not the truth, it's everything. The world will truly be controlled by the people by being the model to their own interface, I don't know whether to laugh or cry

This is an awful idea. Aside from making it even more impossible to independently create new browsers without starting from chromium, it also has exactly the same issues as you mentioned with sharing storage. Which has the same issues as sharing cookies would an so on. You can't have the AI model learn and adjust its behaviour across multiple sites since then it those sites could query it for this information. So you can't build sites that query the users ai in order to tailor itself to their liking. All you can do is create a fancy prompt GUI for google's gemini. Maybe you could use it to create a copilot program if you are making a text editor. But then would it not be better to simply add some tags to html elements, which tells the browser which text is relevant to which input element. Then it would be up to the browser how to use that information.

hmmmmmmmm, so why...so if a page were to declare resources it needs in such a way that the would be fetched before any js is run, wouldn't that solve it? anti tracking should prevent the server from noticing that a js call came from the same source that just fetched the html

THIS IS INSANE!!!! I LOVE IT 😍 edit: ok i played with it and the ai sucks but the concept is insane.. another edit: 11:43 Yeah... Exactly XD

Ffs, why do we keep adding things to the browser. I understand the idea behind it, but can we not. If I want an LLM baked onto my phone, i would be happy to pay for, or download, a model onto my phone. I dont need nor want this :/ (I understand the videos argument btw, just not really strong enough to sway me)

The naming of the feature is so weird. Is it so hard to call LLMs what they are instead of a generic "AI" term that encompasses anything from an LLM to the Dijkstra shortest path algorithm. I'm wondering if companies these days are intentionally creating confusion as part of some marketing ploy.

I had the idea to have a local AI on my Android phone for a long time, I used to always have a survival 101 app in case I would get lost or whatever situation. But nowadays, I literally have a small LLM model in an app that can answer the same or better!

gosh, no. NOOOOOOOOO. Firefox did it 10x better, to fix lacking alt text in images cuz lazy devs. I don't want proooompting in my browser, dammit

Great 🙃, another cool future(along with webgpu& parallel processing) that no one will ever be using it because it's not default enabled for endpoint users, so the only developers who will benefit from this are the chrome dev team itself , since they wouldn't need any permission from the endpoint users .

unless chrome is gonna require a gpu with at least 8 GB of vram (and would use most of it) and a couple extra gigabytes for chrome install, expect a very weak model

Chrome is the new IE6, and I hope webdevs won't force users to use it like they fucking did with WebGL. Fuck Google for pissing on W3C over and over again

AI models do a really good job at imitating not very intelligent normies. no wonder they freak about about it. Aka everything of substance will be mostly wild hallucinations presented very confidently.

6:00 Why doesn't the browser just check, if the website has already downloaded the model and if it hasn't but another one has, then the model is downloaded a second time, but then deleted again to save storage space. This way tracking in the way you described it is impossible, but there still is just one cache file, that both websites will then share.

I don't understand the whining in the comments. tts exists in browsers, and everyone is happy with everything. llms add similar functionality to some extent Previously, it might have seemed strange to some to add tts to the browser. Now this is the norm. The same will happen with llms

Was the model 200MB? it seemed "too good to be true", even though in much cases AI is not half as good as it should be. The energy cost and the global warming consequences are still huge.

and i cant believe youre in my home page after pressing not interested 4 times but life sure has its suprises

What the actual fuck? So all those little advertising scripts are gonna be calling your own local AI to analyse your behaviour now. Cool! Web browsers get even slower.