Apple may be looking to build on-device AI for iPhones, but this does not mean it is forgoing cloud-based AI. The Cupertino giant is still in talks with Google and OpenAI for licensing generative AI, as per a new report from Bloomberg.
Bloombeg’s Mark Gurman cites internal sources that claim that the iPhone maker has restarted negotiations with OpenAI to potentially incorporate the startup’s technology into the forthcoming iOS 18. Discussions are underway regarding the terms of a potential deal and how OpenAI’s features might be integrated.
Furthermore, Apple is engaging with Google to explore licensing the Gemini chatbot. A final partnership decision remains pending, and it’s conceivable that Apple could finalize deals with both companies or another entity. Meanwhile, older reports have claimed that Apple is also discussing a partnership with Baidu to run AI on its devices in China.
By collaborating with external partners for generative AI capabilities, Apple aims to advance its entry into the chatbot market and mitigate certain risks.
More details regarding the matter should be confirmed in June as Apple is reportedly planning to announce generative AI for iPhones through iOS 18 that month. The company is expected to unveil new AI software at this developer conference which will be integrated into its devices for better privacy and protection.
Apple On-Device AI
Apple has recently unveiled a new family of compact AI models designed to run natively on devices thanks to their efficiency. However, unlike Apple’s usual secrecy and lock-in practices, these new AI models are fully open-source.
Known as OpenELM, Apple’s open-source AI models consist of eight models with four different parameter sizes (270M, 450M, 1.1B, and 3B), all trained on public datasets.
OpenELM comes as a direct rival to Microsoft’s recently announced Phi-3, which are also open-source compact models designed to run directly on devices. But despite being small in size, Phi-3 is able to beat OpenAI’s free GPT 3.5 in benchmarks and also Meta’s Llama 3, in not one, but several tests. Meta’s Llama 3 is open-source as well, but not compact. It already powers Meta AI, which is found in its social media apps including WhatsApp and Instagram.
