Startup-eyed Apple running AI of 27 billion parameters directly on iPhone – no servers

Apple is watching PrismML, a startup that has developed a technology capable of running artificial intelligence models with impressive 27 billion parameters directly on the iPhone – without relying on external servers. According to The Information, reflected by 9to5mac, the giant Cupertino showed interest in the acquisition of the company, which may represent a leap in the capacity of AI on-device.

What does PrismML do differently in AI for iPhone?

PrismML has developed a method of compression and optimization of AI models that enables massive neural networks – such as those that normally require large clusters of GPUs – to be executed within the limited hardware of a smartphone. The secret lies in a proprietary technique that reduces memory and energy consumption without sacrificing precision. While competing models with 7 billion parameters are already considered heavy for mobile devices, PrismML can run 27 billion parameters on chips such as the A17 or A18 of current iPhones.

Why does unserved AI matter to the end user?

The main advantage is privacy and speed. With AI running locally, no data needs to be sent to cloud servers, which reduces leak risks and eliminates connection latency. For the user, this means that personal assistants, translators, photo editors and productivity tools can work instantly and offline, with an intelligence level comparable to what is only possible today with remote services. In addition, Apple could integrate technology into its native ecosystem without relying on partnerships with cloud providers.

How can Apple use this local AI technology?

The acquisition of PrismML would give Apple a huge competitive advantage in the smartphone market. Large language models (LLMs) like GPT-4 require huge server costs and generate privacy concerns. With startup technology, Apple could offer advanced AI features – such as text generation, document summary, smart image editing and voice recognition – all on the device itself. This would also open doors for new types of applications, such as smarter augmented reality or real-time accessibility tools. Interestingly, the approach to running AI locally resembles initiatives from other companies, such as French ZML startup that launched free software to run AI on any chip.

Privacy and performance: how does the AI on your iPhone go?

Apple has always placed privacy as a central value, and the AI on-device is the strongest bet to maintain that speech. The PrismML differential is precisely not dependent on the internet: the model is loaded onto the device and processed locally. Energy efficiency is also a critical point – startup claims that its technology consumes less battery than competing solutions. For the Brazilian market, where connectivity can be unstable and expensive data plans, having a powerful AI that works offline would be a huge attraction.

Frequently Asked Questions about AI on your iPhone

Has Apple confirmed PrismML's purchase yet?

Nope. The information, published by The Information, indicates that Apple has shown interest, but no deal has been officially closed.

Will this technology work on old iPhones?

Probably, as long as they have compatible Neural Engine chips. Models from iPhone 12 can run smaller versions of the model, but maximum performance will require more recent hardware.

When are we gonna see this on a real iPhone?

If the acquisition takes place, integration will take 1 to 2 years. Apple usually tests and adapts technologies before releasing them to the public, and PrismML still needs to be incorporated into the company's software pipeline.