But that was unintentional. While Grok is purposely being manipulated to push an agenda. You must surely understand there is a massive distinction.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple in Talks With Google to Power Next-Gen Siri With Gemini AI
- Thread starter MacRumors
- Start date
- Sort by reaction score
I don’t care where the servers are, I don’t want Google’s AI on my iPhone
Gemini would be an interesting choice. In some testing I asked the same questions to Claude, Gemini and ChatGPT. I then discussed with each LLM the differences in their replies. Claude gave a clear overview of major themes before asking if I wanted to drill down on individual topics. Meanwhile ChatGPT and Gemini both got straight into the weeds with a lot of detail up front but no broad overviews. When asked about this Gemini explained that Claude was very well developed to explain things at a broad conceptual level while it was trained on massive amounts of data to give as much detail as possible. So, it maybe depends on what Apple think its customers want - high-level overviews or lots of detail.
I just want to see options. Same with storage - I’d love to ditch iCloud, but you’re forced into it.
Not sure where the supposed drama is in Apple not having its own LLM. Not every tech company needs to have every “thing” of their own. Facebook isn’t dying because they don’t have a search engine. Google isn’t dying because they have their own social network. Apple’s business is selling (often great) hardware at premium prices. Search isn’t their business. Social networking isn’t their business. I don’t think AI is their business either.
A friend of mine texted me last night about how he used chatgpt, Google AI studio and Claude to build some stuff for him. Now, I’ll be phasing out a Dynamics CRM solution for a customer soon, which means all their accounts and contacts will need to be exported into something else. I’ll be using Python for that. The Dynamics CRM api uses the typical refresh_token/access_token setup where the access token can expire but you can get a new one through the refresh token.
So I asked an AI to write the Python code to refresh the access token. Of all the choices I had I picked Google AI studio. No real reason other than it was mentioned in the text that I got earlier. The code was written. I’ve used the Dynamics CRM api in other languages and as far as I can tell, the Python code generated by Google looks correct.
Makes me wonder though … does it even matter which AI I got this code from? I’m sure Claude would have spit it out as well. Come to think of it, I bet chatgpt would have too.
So, in the end, does it really matter which AI wrote that code? It certainly doesn’t for me, because it’s irrelevant. The part that matters is that I got the code that I wanted. My customer doesn’t care. Just like they don’t care if I use(d) Stack Overflow for other things.
So, let’s say I’m in Xcode doing some Swift UI stuff. And some AI in Xcode helps me to achieve that. Does it matter if the code was generated by Google, Claude or Chatgpt? No, it doesn’t.
In fact, I’d rather have great code generated by a non-Apple LLM than dodgy code generated by Apple themselves.
A friend of mine texted me last night about how he used chatgpt, Google AI studio and Claude to build some stuff for him. Now, I’ll be phasing out a Dynamics CRM solution for a customer soon, which means all their accounts and contacts will need to be exported into something else. I’ll be using Python for that. The Dynamics CRM api uses the typical refresh_token/access_token setup where the access token can expire but you can get a new one through the refresh token.
So I asked an AI to write the Python code to refresh the access token. Of all the choices I had I picked Google AI studio. No real reason other than it was mentioned in the text that I got earlier. The code was written. I’ve used the Dynamics CRM api in other languages and as far as I can tell, the Python code generated by Google looks correct.
Makes me wonder though … does it even matter which AI I got this code from? I’m sure Claude would have spit it out as well. Come to think of it, I bet chatgpt would have too.
So, in the end, does it really matter which AI wrote that code? It certainly doesn’t for me, because it’s irrelevant. The part that matters is that I got the code that I wanted. My customer doesn’t care. Just like they don’t care if I use(d) Stack Overflow for other things.
So, let’s say I’m in Xcode doing some Swift UI stuff. And some AI in Xcode helps me to achieve that. Does it matter if the code was generated by Google, Claude or Chatgpt? No, it doesn’t.
In fact, I’d rather have great code generated by a non-Apple LLM than dodgy code generated by Apple themselves.
Noooo - I love Anthropic. I don’t want Tim Cook to ruin it.Here's an idea for Apple: If Anthropic's fees are too high, JUST EFFING BUY THEM ALREADY AND GET IT OVER WITH.
Although, I must say, this actually makes sense. This way, when Google is compelled by the courts to stop paying Apple 20 billion a year for default Google search results, Google can pay Apple the same amount for default Gemini search results instead.It's pretty obvious by now that Google really likes paying Apple 20 billion dollars a year, and Apple really likes accepting that 20 billion, so why mess with a winning formula? Leave these poor kids alone for gosh sake.
And the wheels keep spinning, round and round.
I'm mostly interested in Apple Intelligence for what it can do for my smart speakers. I started with Google devices using Google Assistant and moved on to Amazon Alexa because Alexa was much faster at the time, had more integrations, and because I hated having to say "Hey" or "OK" before the Google wake word. I've also played with Siri a bit on my iPhone and Watch enough for me to decide it's not something I would want to use in a smart speaker right now, a step down from Alexa let alone Google Assistant (which in my opinion is still the smartest smart speaker assistant right now).
One thing I've discovered is that for my use of smart assistants ground truth is really important in terms of distances, directions, opening hours etc. At least outside of the USA Google is streets ahead (no pun intended!) of even Apple let alone Amazon in terms of what it knows about local geography, public transport, shops and restaurants etc.
If Apple were to go with Gemini I wonder whether that would also give it access to any Google ground truth. From what I've experienced so far with playing with various public LLMs they're pretty hopeless at effectively interfacing with dynamic sources of information. Asking what time the next train is into London and being told to go and look it up on the local train company's timetable isn't the quality of answer I'm looking for from a smart assistant.
The above is a real life example from one of the major ChatGPT/Claude/Gemini LLMs but I can't remember which one it was. It's a test I try now and then on all of the major LLMs and from the last time I tried it just after ChatGPT5 came out one told me point blank to go and look it up on the train company web site, another told me the train company name and web site but said it couldn't navigate certain interactive elements of that company's web site so I'd have to do it myself, and the third one (that was ChatGPT 5) told me the train times of the next 5 trains, and even departure platforms, but the first three "next" train departure times that it suggested were actually in the past. Google Assistant was answering that same question perfectly every time even way back in 2018 when I got my first smart speaker.
LLMs can do some clever party tricks but from what I see there is still work to be done to get them to give good answers to basic questions like "when is the next train to <somewhere>?" or "When does <some-shop> close today?".
One thing I've discovered is that for my use of smart assistants ground truth is really important in terms of distances, directions, opening hours etc. At least outside of the USA Google is streets ahead (no pun intended!) of even Apple let alone Amazon in terms of what it knows about local geography, public transport, shops and restaurants etc.
If Apple were to go with Gemini I wonder whether that would also give it access to any Google ground truth. From what I've experienced so far with playing with various public LLMs they're pretty hopeless at effectively interfacing with dynamic sources of information. Asking what time the next train is into London and being told to go and look it up on the local train company's timetable isn't the quality of answer I'm looking for from a smart assistant.
The above is a real life example from one of the major ChatGPT/Claude/Gemini LLMs but I can't remember which one it was. It's a test I try now and then on all of the major LLMs and from the last time I tried it just after ChatGPT5 came out one told me point blank to go and look it up on the train company web site, another told me the train company name and web site but said it couldn't navigate certain interactive elements of that company's web site so I'd have to do it myself, and the third one (that was ChatGPT 5) told me the train times of the next 5 trains, and even departure platforms, but the first three "next" train departure times that it suggested were actually in the past. Google Assistant was answering that same question perfectly every time even way back in 2018 when I got my first smart speaker.
LLMs can do some clever party tricks but from what I see there is still work to be done to get them to give good answers to basic questions like "when is the next train to <somewhere>?" or "When does <some-shop> close today?".
I'll never get tired of saying it: I can't believe Apple was first with Siri and just sat on that watching everyone surpass them. They were years ahead and did nothing. Huge missed opportunity.
You make some good points but you are overselling things that lots of models can do. I'd add that the reason Apple is considering google is because google is also heavily invested in on device models so they understand what Apple needs, and are in a more advanced state than anything Apple has been able to produce.
Like for instance google released Gemma 3n which is a preview of where gemini nano is going. It is a 5 - 8B parameter model that effectively runs as a 2 - 4B parameter model.
Gemma 3n is
1. multimodal, can take in text, images and audio
2. multilingual, handling over 140 languages
3. has video understanding
4. is agentic and can use tool calls to operate your phone and apps (the uber eats thing you are talking about)
Not true. Literally everyone with a platform is working on this that's what MCP and tool calling is. Its an "API" that agentic models can use to call certain functions that require external service.
Last edited:
That's like saying everyone else has the same thing as SwiftUI because they have XAML.
Details matter...
The very fact that 1/4 of your readers will consider Elon the tech devil, while another quarter will consider that to be Zuch, and the quarter that consider themselves cool and leading edge have already transitioned to hating Sam Altman as the tech devil really answers your own question.
You're not going to get anywhere in either acquiring customers or making deals by treating the AI world in terms of "who is the REAL tech devil"...
And yet Apple created something very good. Whereas Intel did not under the same circumstances.
Management competence is more important than people like you think.
Wow this is pathetic if the deal is finalized! Apple used to be all about privacy.
If you followed the news you would know Musk claims he backtracked his blatant interference with the model. But the fact he did it once already says enough. What stopping him from tampering with Grok in more subtle ways? We know he also messed with the X algorithm as well.
You could argue that Intel weren’t giving modems the required focus. Apple brought two sets of experts together as Cook recognised he needed something he didn’t have.
- The Engineer’s Plan:
The engineer has an idea, makes a plan.
And the Plan was good. - The Supervisor looked at the Plan:
The supervisor said, “This is good.”
And the Plan was good. - The Manager looked at the Plan:
The manager said, “Change this, change that.”
And the Plan became complicated. - The Director looked at the Plan:
The director said, “Add this, add that.”
And the Plan became expensive. - The Vice President looked at the Plan:
The VP said, “We need this immediately.”
And the Plan became late. - The President looked at the Plan:
The president said, “Make it cheap.”
And the Plan was impossible.
Out maybe this
- Starts simple, makes sense,
- Every level of management adds nonsense,
- Ends up unworkable or doomed.
No, you're mistaking a developer framework for a revolutionary AI capability. The details of the API matter for developers, but the core function is the same.
Code for tool calling in Apple swift
In Google Kotlin using genkit
It's called tool calling. It's an industry standard feature. Google, OpenAI, Anthropic, opensource models can all do it. It's literally just giving a model a set of tools (app functions) and having it figure out how to use them to get something done. The protocol is just a structured way for a Swift developer to define a function that the AI can call. It has a name, a description for the model to understand what it does, and a call function to execute it. This is the same concept that exists in every other major AI framework. Its a basic feature that allows all these AI models to access external services.
Your Uber Eats example is a prime use case for it. It's not new. You want to see it in action? Watch the Gemma 3n video I linked;
- Looks at a photo of a receipt.
- Opens a completely different expense app.
- Fills in the expense name and the exact amount from the photo.
- Saves the entry.
Last edited:
Why?! Gemini is arguably the worst relevant AI out there! Go back to the days where Siri would just reference wolfram-alpha when it didn’t know the answer.
What metrics are you using? For day to day use and Live chat I prefer Gemini. For other stuff it’s a mix of perplexity or Claude. At work I’m limited to copilot which is functional I suppose