Google limits Meta's use of its Gemini AI models
cnbc.com
cnbc.com
In the current climate limiting someone's use of AI might be expected to be about restricting access or restricting what someone can do with it, but the story here ostensibly seems to be about capacity constraints, not any limitation on what models or capabilities Google is giving Meta access to.
Cloud services like to present the illusion of an infinite amount of compute available at a fixed price per unit, but the reality is if you try to use too much of any service you'll find you have a quota and requests to increase it will fall on deaf ears if the provider doesn't have more of that resource.
Too much of my working life has been spent shoehorning services into less space/compute/ram/spindles or migrations to other data centers to solve such issues.
Having said that, I agree with you. You have to request limit increases often and can't scale even in those instances if you don't plan ahead.
There has to be a name for this deceptive marketing tactic where you say something is unlimited and then it is only unlimited as long as you don't use very much.
It would be one thing if you occasionally got a "no more capacity" error when requesting large amounts of resources but it doesn't work that way. They confine you to a relatively small amount of resources the entire time you have an account. If you want more you have to request it.
The tiny blog sure isn't for the cloud, but also it's not the main client of the cloud.
> it's 20% more than you are currently using and you pay 300% more for that.
I'm assuming you are comparing to self hosting. Then you need to account for things that are difficult to put a price like your time maintaining a physical infrastructure and the lessons you will learn with it.
Sounds like I'm defending the big cloud, but there is a valid use that is disconsidered because it's trendy to hate on the cloud.
> They confine you to a relatively small amount of resources the entire time you have an account. If you want more you have to request it.
It's a form of KYC, nothing wrong with that.
Like literally 10x times more expensive to do so, to run CI jobs...
I dont want to imagine the margin AWS has like generally, cause it can easily be a 90% too
The margin on the list prices are clearly far higher.
I assume you're using your owned server and not a provider like Hetzner? So you did have a substantial delivery time. Although in my city is a recycled that resells used servers, and I could show up there with a truck and get a server within hours if I'm not too picky. Or use some random desktop or laptop off the pile, short-term.
Right now the biggest issue is the vibe coded CI program is not really meant to be a distributed multi-node thing yet, so we're on the biggest machines (there's some newer bigger stuff we could migrate too) and the only issue is on peak hours queue can get a bit slow.. but that was also some other bugs etc making not ideal.
Tbh it works pretty well, we just need now to scale it to more than one node etc (which is not to say that is easy, but still, x10 headroom to work with)
When Anthropic accuse Alibaba of distilling their models, you have run that by a reality check of what is actually possible.
1) You can use another model as "LLM as judge" to rate alternative outputs that your own model has generated. Useful data perhaps, but certainly not distillation.
2) If what you are interested in are the reasoning steps (that are hidden from you) that arrived at an answer, not the answer itself, then you can try to train a model to guess what those steps were (this is a published technique). This may be better than nothing, but hardly distillation if it's your model that is suggesting the reasoning!
3) Depending on the model, you may be able prompt engineer it to reveal it's reasoning, not just show a summary, but this should be very obvious. Anthropic cite this as something they have seen. This would be useful data if you can get it (presumably they've now done a better job of preventing it), but at the end of the day all you'd be getting is some training data cheaper than if you'd had to create it by hand.
It's probably the best multimodal model I've worked with (if somebody knows a better one for audio analysis, please let me know!)
I want to know how impacted Gemini has been by that, because that will reveal a lot about their margins and revenue generating first party demand. Each MSFT earnings report they discuss the balance they’re dealing with between supplying GPUs to Azure customers and first party demand.
My pet theory is that Gemini is “losing” the LLM race because they’re preferentially selling the TPUs to competitors, while keeping just enough for themselves to stay competitive and build their own products.
I HIGHLY doubt that Gemini is overloaded, Google has been bullshitting with their crap models since release. Waste of everyone's time.
I use Gemini daily, on a paid basis. Mostly 3.5 Flash Medium. It’s very good for software development, has a great context window, and pricing is about 1/3rd of Claude Opus. I haven’t experienced availability problems except once or twice, certainly no more often than OpenAI or Anthropic.
Seems like the complaints about it are mostly from people moaning about Google not giving them AI for free. Good luck with that.
And I'd expect Google will want to prioritize capacity for those - they don't want their google pixel phone to error or google search to barf.
OTOH, if they are stressing Google's capacity then it seems it has to be for production use, which would relfect a massive failure on Meta's side given their investment in datacenters and AI. If they can't utilize their own models and datacenters, then maybe they should just rent the excess capacity to Google! :)
And their safety tuning is neither effective nor precise on edge models.
Who says they aren't? Could be using all of them for "research".
Llama Meta 70b is 50th or so down the list of popular models.
It has 24.1b tokens used in 7 days vs the top models that have trillions or hundreds of billions of tokens.
So practically dead!
Image generation and veo models I’d imagine quite effective for creators; new Instagram accounts with AI content that are garnering millions of followers in spans of weeks are quite common now
honestly better than gemini flash
That said, I expect much of the AI bubble to pop. Google Gemini with Antigravity is a good product, as is a Claude Code subscription but I have switched to using DeepSeek v4 Pro with the Claude Code harness and DeepSeek v4 Flash with the OpenCode harness (when I am not using local models with little-coder/pi) and at least for the foreseeable future I don’t think I am going back. Fast APIs at low cost trumps having to spend a little more time to get the same quality of results.
* Repeated instances of incorrect code insertion that the agent cannot clean up. Sure, version control, but this is often happening in new files that aren't even in version control yet.
* Lost chat history when I close and restart the app.
* Not being able to restore a chat from the history (just saw this last week).
* Overly broad searches that waste time and tokens.
* No vertical scroll bar arrows. WTF?? Doesn't the interface look "flat" enough already? This feels arbitrary and stupid.
* The previous chat prompt takes up a large portion of the vertical space of the chat window, even on a high res display.
When it works Antigravity is excellent. When it doesn't work, it's absolutely horrible. If you check the update history, there are usually just a few items and they're super generic things like "Fixed a bug with text entry.".
I don't see it improving at any kind of reasonable pace, even over the last 6 months As a result, I've mostly relegated Antigravity to a planning tool and it does an excellent job. Or I use it to write prompts that I give to Codex. It definitely can do an excellent job writing code sometimes, but sometimes it also does an absolutely horrible job with not breaking the code when it inserts it. It seems to be terrible at understanding C++ braces. How often? Way too often. I always know it's happening because it prompts me to run Git while it's doing something. LOL, that's how I know that it's broken something.
Codex is definitely way, way, way better. It's not even a contest at this point. Codex never breaks my code. It might not always do what I want, but it's just an order of magnitude better than Antigravity. Antigravity really feels like a comedy of errors at this point. ESPECIALLY from a company with Google's resources.
Understanding the actual problems we are trying to solve with code and efficiently coming up with solutions (essentially, pre-LLM development) will always be better than wastefully brute forcing solutions with LLMs.