In a cautiously worded announcement connected Monday, Apple said John Giannandrea, who has been the company’s AI main since 2018, is “stepping down” to, well, not enactment astatine Apple anymore. He’ll instrumentality astir done outpouring arsenic an advisor.
His replacement is Amar Subramanya, a highly regarded Microsoft enforcement who spent 16 years astatine Google, astir precocious starring engineering for the Gemini Assistant. It’s a savvy hire, fixed that Subramanya knows the contention intimately.
The determination is being characterized arsenic a shake-up. It was seemingly inevitable successful retrospect. Apple Intelligence, the company’s reply to the ChatGPT moment, has been stumbling since its October 2024 launch. Reviews person ranged from “underwhelming” to outright alarmed.
Its archetypal months were immoderate of the roughest. A notification summary diagnostic meant to condense aggregate alerts into digestible snippets generated a bid of embarrassing, untrue headlines successful precocious 2024 and aboriginal 2025. Among different missteps, the BBC complained twice aft Apple Intelligence falsely reported that Luigi Mangione, the antheral accused of sidesplitting UnitedHealthcare CEO Brian Thompson, had changeable himself (he hadn’t) and that a darts player, Luke Littler, won a title earlier the last even began.
Then determination was Siri’s promised overhaul, which became a achromatic oculus for Apple.
A Bloomberg investigation published successful May revealed the depths of Apple’s AI struggles. For instance, erstwhile Craig Federighi, Apple’s bundle chief, tested the caller Siri connected his ain telephone conscionable weeks earlier its planned motorboat successful April, helium was dismayed to find that galore of the features the institution had been touting didn’t work. The motorboat was delayed indefinitely, triggering class-action lawsuits from iPhone 16 buyers who’d been promised an AI-powered assistant.
By that point, Giannandrea had already been sidelined, according to Bloomberg. The quality enactment reported that Tim Cook had stripped Siri from Giannandrea’s oversight wholly backmost successful March, handing it to Vision Pro creator Mike Rockwell. Apple removed its secretive robotics part from Giannandrea’s control, too.
Techcrunch event
San Francisco | October 13-15, 2026
Bloomberg’s probe painted a representation of organizational dysfunction, with anemic connection betwixt AI and selling teams, fund misalignments, and a enactment situation terrible capable that immoderate employees had taken to mockingly calling Giannandrea’s radical “AI/MLess.” The study besides documented an exodus of AI researchers to competitors, including OpenAI, Google, and Meta.
Apple is reportedly present leaning connected Google’s Gemini to powerfulness the adjacent mentation of Siri, an astonishing and also, presumably, humbling twist considering the aggravated rivalry betwixt the 2 companies that dates backmost much than 15 years, crossed mobile operating systems, app stores, browsers, maps, unreality services, astute location devices, and present AI.
Giannandrea came to Apple from Google, wherever helium ran Machine Intelligence and Search. At Apple, helium oversaw the AI strategy, instrumentality learning infrastructure, and Siri development.
Now Subramanya inherits those responsibilities, reporting to Federighi with a wide mandate to assistance Apple drawback up successful AI.
It’s an absorbing infinitesimal for the company. While competitors person been pouring billions of dollars into monolithic AI information centers, Apple has focused connected processing AI tasks straight connected users’ devices utilizing its customized Apple Silicon chips, a privacy-first attack that avoids collecting idiosyncratic data. (When much analyzable requests necessitate unreality processing, Apple routes them done Private Cloud Compute, servers that committedness to process information temporarily and delete it immediately.)
Whether that doctrine pays disconnected oregon whether it has permanently near Apple down is an outstanding question. Apple’s attack comes with wide trade-offs. Among them, on-device models are smaller and little susceptible than the monolithic models moving successful competitors’ information centers, and Apple’s reluctance to cod idiosyncratic information has near its researchers grooming models connected licensed and synthetic information alternatively than the elephantine troves of real-world accusation that substance its rivals’ systems.















English (US) ·