Report: Apple wanted to make basic AI models open source
Meta's Llama and Google's Gemma are already open source. Apple was apparently planning to do the same, but the software boss put his foot down.
Apple Intelligence: Based on Apple's basic models.
(Image: T. Schneider/Shutterstock.com)
In the meantime, Apple is said to have planned to shift up a gear with the help of external developers at Apple Intelligence: The company was planning to release parts of its basic models as open source this year. According to a report in the Silicon Valley trade journal The Information, the company also wanted to do this for image reasons – to show the world that the iPhone manufacturer has made progress with regard to Large Language Models (LLMs).
However, this would apparently also have meant admitting to performance losses that occur as soon as the models no longer run on Apple's private cloud compute servers (PCC), but locally on the device. For these and other reasons, software boss Craig Federighi is said to have put a stop to the idea.
Models are considered inferior to the competition
Apple's AI models are considered inferior to the competition. The company admitted this in its latest "Tech Report" on Apple Intelligence with some benchmark results. In some categories, similar values GPT-4o were achieved, but then the results plummeted again. Local models generally perform worse due to the lower available computing power – Apple relies on this approach more than other manufacturers because it has data protection advantages. Apple also recently lost key employees from its AI model department to Meta's "Superintelligence" team, which offers huge salaries (or share awards).
Videos by heise
The open sourcing of important models was seen as a possible lifeline here. But Federighi, Apple's most powerful software man, thought this was problematic. In an email to Ruoming Pang, who now works at Meta, he wrote that there were already "many open source models from other companies" that already allowed research.
In addition to the fact that Apple's foundation models perform worse than those of Alibaba or Google, Federighi is said to have feared above all that there would be criticism of the local models. Open-source foundation models might have shown that Apple was "making too many compromises" here, it was said.
Federighi must keep the team together
These have apparently been controversial in Apple's AI team for some time. The idea of "on-device first" is holding back development at the company, according to individual team members. Even Federighi seems to have doubts as to whether this approach will work in the long term.
At the moment, however, he has completely different problems: he has to keep the Apple Foundation Model team together after the aforementioned departures. Apple therefore wants to pay AI researchers more money, according to –, but the company cannot match the huge sums that Meta, for example, wants to spend. Meanwhile, users are still waiting for Siri to finally be improved
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)
