Hacker News Reader

Comment by minimaxir | Hacker News Reader

I wonder how much of Siri AI is Apple-developed and how much of it is Google-developed as a result of Gemini. The a) search demos and b) image generation demos seem unlikely to have been done by Apple alone, the demos being closer to Google Search and Nano Banana respectively.

View on HN

It looks almost entirely like gemini. The images they showed are obviously nano banana, and the text responses are almost obviously Gemini (I say as a somewhat frequent Gemini user).

I'm sure they customized some of it, but this looks basically like Gemini integrated with iCloud instead of Google Workspace.

I'm curious what the obvious tells were for you. I never use any of these tools so I do wonder what sets them apart for those in frequent contact.

The images were the biggest tell, generating using a reference photo of a person, at least Gemini and ChatGPT have two distinct styles. ChatGPT is a little less uncanny valley than Gemini which tries to be too realistic looking, in a bad way because it tries to preserve the original person in the photo, but still can't seem to help altering facial features.

The text responses had Gemini's verbosity. Asking ChatGPT to show me iconic dishes from both Brazil and Morocco (Apple's example), is much cleaner, less verbose. Quick list of dishes and links to the recipe. Gemini just spews a wall of text and bullet points and goes on and on with fluff. Tons of "What this dish is" "Why it works" Same with its frequent use of tables, which I see less of with ChatGPT.

Each Siri demo they did in the keynote had that hallmark verbosity I typically get with Gemini without prompting it to not do that.

Yeah I noticed the same things! I've been using Gemini on my Pixel a lot like this and this feels like Siri skin for Gemini

There was not a single thing they launched that I have not seen Gemini already showcase capability or existing feature wise

The only thing you saw was phoning home, that’s what’s gonna be interesting when Apple releases their version. Is it phoning home 100% of the time or can you turn off the Internet and have it perform in the same way, there will be plenty of YouTubers that will give it the test a test I might add that they haven’t done up until this point for anything that Google has put out?

Is Siri any more or less than “just” an agentic harness such as OpenClaw? How much of what that harness does is up to the LLM or the harness itself?

In my mind the Gemini LLM defines the bounds of capability and capacity, but any actual functionality or usefulness (or lack of) comes from Apple’s Siri harness.

I was wondering the same. I have to imagine it's mostly Gemini, unless Apple has a big, secret, SotA foundation model no one has heard of? But if it is Gemini, how does that work with their Private Cloud thing? Are they able to load the Gemini weights into it?

IIRC Apple cut a deal to have their own version of Gemini that is hosted just for them.

It’s almost all Gemini and the “Apple local models” part seems to just be image embeddings/descriptions powering new spotlight and the like which is also likely someone else’s model.

Reply on news.ycombinator.com