What AI models does Apple Intelligence rely on? Apple explains, warning us about possible hallucinations

Among the new features announced in WWDC 24Apple has devoted enough space to it Apple Intelligencethe new set of AI features that will power Apple’s future operating systems starting with iOS 18.

But what models is it based on exactly? Apple Intelligence, it is not clear. Or rather, Apple has stated that certain features, such as Writing tools and some selected questions a SYRIA will be managed by ChatGPT and users will be able to link their subscription as well ChatGPT Plus to access GPT-4o functions. And the others?

A new blog post dedicated to research on machine learning i Apple does little clarity, it is the same TIM Cook warns that these tools are not 100% reliable.

Apple Intelligence It is therefore a set of AI functions that partially relies on ChatGPT, but most of the work is done by self-developed large language models (LLMs).

The models built into Apple Intelligence are said to be refined for user experiences such as write AND perfect the text, prioritize and summarize notifications, create imaging to chat with family and friends and take in-app actions to simplify cross-app interactions.

pattern in the device i Apple has a size of approx three billion parametersand it’s optimized with techniques like quantization for use on devices like the iPhone 15 Pro and 15 Pro Max with 8GB of RAM.

The cloud-based model, however, is bigger and more powerful. Apple has not specified it explicitly dimensions, but is designed to run entirely in data centers powered by Apple siliconeand therefore directly on its servers and not relying on those of third parties.

But how do these models behave? According to standards, which should obviously be taken with a grain of salt as they could have been performed on specific training data to highlight the goodness of the product, very well.

In the graph below, you can see that human raters prefer answers from Apple’s cloud model about 50% of the time over those provided by GPT-3.5while GPT-3.5 achieved a complete victory in only 24.7% of cases.

Compared to many more powerful GPT-4 TurboHowever, Apple saw its lead shrink to just 28.5%, with a tie 29.8% of the time.

The model too in the device Apple performs quite well, beating or keeping up with it Mistral-7B AND Gemma-2B in most responses tested.

These standards may be why the company took a hybrid approach Apple Intelligence and Siri in particular, as she was unsure of the actual security of some responses, and therefore relied on an external tool, adding a disclaimer to warn users.

On the subject of security, Apple says its base models are much safer of competition.

The model based on NEW of the company returned “replies with harmful content, sensitive topics and factual facts“At just 6.6% of the time, much lower than GPT-3.5 Turbo’s 15.5% and GPT-4 Turbo’s 20.1%.

In an interview with Washington PostTim Cook said Apple didn’t want it to compromise your values ​​to present AI functionsand that there are limitations to protection users.

According to Cook, AI will help users who save timebut when asked if UA e Apple Intelligence suffers from hallucinations, that is, if she can make up facts, Apple’s boss had to admit that she is not 100% accurate.

About the partnership with OpenAI, and if you trust the content created by ChatGPT especially after the latest charges of several company employees (and Google) about its security, Cook did not elaborate, saying only that they were looking for the best possible solution.

In general, Cook tried to bring attention to Apple’s commitment to a responsible AI and safe, and in the initiatives taken to guarantee privacy of users.

Back to Apple’s blog on it Foundation modelsthe company Apple states for example that they do not use private personal data of users or their interactions during training.

Overall, Apple’s AI seems to have done remarkable things PROGRESS (especially from the absolute nothing of a year ago), but these technologies are evolving rapidly. Apple compared its models with GPT-4 Turbobut OpenAI has already released GPT-4o, and who knows what will happen in four months.