3 ms·

There’s lots of new AI features in iOS 26, iPadOS 26, etc. But they aren’t the blockbuster features. They’re more like quality of life issues that users will a

There’s lots of new AI features in iOS 26, iPadOS 26, etc. But they aren’t the blockbuster features. They’re more like quality of life issues that users will appreciate. A now that 3rd parties can access Apple’s LLM models… let me correct that. Shortcuts, a visual automation app, can also access models on device or bigger, more capable models using Apple’s Private Cloud Compute. Apple’s not playing multidimensional chess… but they are playing the long game, where users won’t have to use multiple AI chatbots to get work done, because most of what they want to do is handled by their current apps with new AI capabilities—on device.

bitpush 1y ago
Apple's models are not competitive. Apple has not demonstrated any leadership in fundamental models so far, and I don't expect that to change any time soon. If anything I'd expect Google, OpenAI, Anthropic ... Or even Meta to have a better on-device "lite" model before Apple.
- alwillis 1y ago
  > Apple's models are not competitive. 1. You have no way of objectively assessing this. 2. Apple’s been using machine learning, neural networking and other AI technologies in their operating systems before most of these AI companies even existed. 3. Apple was the first to ship AI-specific hardware (Neural Engine) in the iPhone X in 2017 > Google, OpenAI, Anthropic ... Or even Meta to have a better on-device "lite" model before Apple. Google's "lite" model Gemma 2B has 2 billion parameters and their Gemini Nano has 1.5 billion. The largest lite model Meta tops out at 1.5 billion. The model demonstrated at WWDC has 3 billion parameters [1] with 2-bit and 4-bit quantization. So Apple has already surpassed Meta and Google when it comes to LLMs on smartphones from the jump. Seems pretty competitive. 3rd party developers will have access to it as well. If the task requires a more comprehensive model, it can access Apple's Private Cloud Compute [2] seamlessly. Sounds like a win-win to me. This model is also exposed to automation; you'll be able to create custom workflows that incorporate AI very easily. [1]: From Apple's press release: A key component of Apple Intelligence is a new on-device Foundation Model. This model, with approximately 3 billion parameters, has been engineered for efficiency and is optimized to run directly on Apple's latest silicon. To achieve this, Apple has employed advanced quantization techniques, including 2-bit and 4-bit quantization for the model's weights and an 8-bit KV cache, which significantly reduce the model's memory and computational footprint without compromising performance [2]: "Private Cloud Compute: A new frontier for AI privacy in the cloud" https://security.apple.com/blog/private-cloud-compute/ https://security.apple.com/blog/private-cloud-compute/

bitpush 1y ago

Apple's models are not competitive. Apple has not demonstrated any leadership in fundamental models so far, and I don't expect that to change any time soon. If anything I'd expect Google, OpenAI, Anthropic ... Or even Meta to have a better on-device "lite" model before Apple.

alwillis 1y ago
> Apple's models are not competitive. 1. You have no way of objectively assessing this. 2. Apple’s been using machine learning, neural networking and other AI technologies in their operating systems before most of these AI companies even existed. 3. Apple was the first to ship AI-specific hardware (Neural Engine) in the iPhone X in 2017 > Google, OpenAI, Anthropic ... Or even Meta to have a better on-device "lite" model before Apple. Google's "lite" model Gemma 2B has 2 billion parameters and their Gemini Nano has 1.5 billion. The largest lite model Meta tops out at 1.5 billion. The model demonstrated at WWDC has 3 billion parameters [1] with 2-bit and 4-bit quantization. So Apple has already surpassed Meta and Google when it comes to LLMs on smartphones from the jump. Seems pretty competitive. 3rd party developers will have access to it as well. If the task requires a more comprehensive model, it can access Apple's Private Cloud Compute [2] seamlessly. Sounds like a win-win to me. This model is also exposed to automation; you'll be able to create custom workflows that incorporate AI very easily. [1]: From Apple's press release: A key component of Apple Intelligence is a new on-device Foundation Model. This model, with approximately 3 billion parameters, has been engineered for efficiency and is optimized to run directly on Apple's latest silicon. To achieve this, Apple has employed advanced quantization techniques, including 2-bit and 4-bit quantization for the model's weights and an 8-bit KV cache, which significantly reduce the model's memory and computational footprint without compromising performance [2]: "Private Cloud Compute: A new frontier for AI privacy in the cloud" https://security.apple.com/blog/private-cloud-compute/ https://security.apple.com/blog/private-cloud-compute/