Report: Apple fully relies on local AI models at WWDC
With Google Gemini help, Apple wants to strengthen its on-device AI. At the upcoming developer conference, Apple Silicon's capabilities will be key.
Stylized Apple CPU: Apparently doesn't break a sweat with AI.
(Image: Bild erstellt mit KI in Bing Designer durch heise online / dmk)
So far, the capabilities of Apple Intelligence have been considered rather meager – Apple's combination of local language models and Private Cloud Compute (PCC) consistently lags what more modern systems like ChatGPT or Claude deliver. Nevertheless, Apple will continue this course with its new Google Gemini partnership. This is reported by the IT news service The Information from Silicon Valley. Apple will once again prioritize on-device AI models and relegate the cloud aspect to the background, even though the company plans to utilize Google's server infrastructure.
Advertising for local models based on Apple Silicon
Apple apparently continues to see local models as a privacy argument. According to The Information, the iPhone manufacturer wants to highlight its experience in building its own Apple Silicon chips, which have existed for 16 years (originally started with the iPhone, now present in all major devices). Apple Silicon thus gives the company the advantage of using good on-device models on smartphones, Macs, and even smartwatches.
Videos by heise
Apple wants to move away from the poor reputation of Apple Intelligence among users with Google's help: According to the contract with the internet giant, the company is allowed to use its large Gemini models to train local variants – through distillation. Whether this will really lead to significant quality improvement remains to be seen. Apple is also looking for smaller AI companies specializing in the topic, according to The Information, citing informed sources. This includes, for example, the company Liquid AI from Cambridge, which specializes in model scaling.
Apple spends comparatively little
Apple has been promoting the strategy of on-device models for a long time – and often repeats this point to the media and politicians. For users, this is relatively transparent, as they should be able to check whether PCC is being used or not. But PCC was also implemented in such a way that Apple does not see the data. Observers assume that this will also be the case when using Google's servers for the large Gemini models.
Local models are intended, among other things, to serve the long-overdue renovation of the Siri voice assistant. However, “parts” of it will also run on Google servers. Apple is said to have internally allowed the use of Nvidia's Confidential Computing, among other things. Apple held back significantly on the major AI expansion with the creation of data centers, for which Google, Meta & Co., alongside AI specialists like Anthropic and OpenAI, have spent many, many billions. The rate of ”Capex,” i.e., capital expenditures, was only $12.72 billion at Apple last year, while Meta invested $72 billion – a sum that has since increased significantly. Analyst Richard Kramer of Arete Research estimates that the total number of Apple devices in the market, if used for local AI, would represent around $50 billion in computing capacity, according to The Information.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(bsc)
