We’ll see. To date there’s no local runnable generative LLM model that comes close to the gold standard GPT-4. Even coming close to GPT-3.5-turbo counts as impressive.
We only recently got on-device Siri and it still isn’t always on-device if I understand correctly. So the same level of privacy that applies to in-the-cloud Siri could apply here.
Apple has sold computers with local voice input and command processing for more than 20 years, and iPhones have pretty much always had that feature (it was called “Voice Control” before Siri existed, and it was 100% local).
I’d argue that, for Apple, what they’ve started doing recently is processing commands in the cloud. The list of commands that are processed locally vs in the cloud has changed over time… and they did move most of it to the cloud several years ago when they bought a cloud based smart assistant startup and used it as the basis for a new an improved assistant on iPhone. But every year they remove the dependence on that and are going back to how it used to be with local processing. These days even when a command is processed in the cloud it’s often only part of a multi-step process where the majority of the work was done on device. And many everyday commands are done entirely on device.
For example if you ask it what the weather is, it’s entirely an on device command except for actually checking the latest weather report… and you can ask it what the temperature is “inside” which will check a sensor in your house and be entirely offline (if your home has a temperature sensor. There’s one built into Apple smart speakers and also a small but growing number of third party smart home products)
To date there’s no local runnable generative LLM model that comes close to the gold standard GPT-4.
True - but iPhones do run a local language model now as part of their keyboard. It’s definitely not GPT-4 quality but that’s to be expected given it runs on a tiny battery and executes every single time you tap the keyboard. Apple has proven that useful language models can be run locally on the slowest hardware they sell. I don’t know of anyone else who’s done that?
Even coming close to GPT-3.5-turbo counts as impressive.
Llama 2 is GPT-3.5-Turbo quality and it runs well on modern Macs which have a lot of very fast memory. Even their smallest fanless laptop can be configured with 24GB of memory and it’s fast memory too - 800Gbps. That’s not quite enough to run the largest Llama2 model but it’s close to enough memory. Their more expensive laptops have more memory and it’s faster - they can run the 70 billion parameter llama 2 without breaking a sweat.
And on desktops Apple sells Macs with 192GB of memory and it’s way faster at 6.4Tbps. That’s slightly more memory (and for a lot less money) than the most expensive data center GPU NVIDIA sells (the NVIDIA unit is faster at compute operations but LLMs are often limited by available memory not compute speed).
