Apple has appointed Amar Subramanya as its new leader in the field of artificial intelligence, following the departure of John Giannandrea who is now transitioning into an advisory role. Subramanya comes to Apple with a unique hybrid experience working at the front lines of Microsoft and spending 16 years at Google, most recently as head of engineering for Gemini Assistant, and that gives Apple someone who knows its biggest rivals going back to front.
Subramanya reports to software lead Craig Federighi and takes over Apple’s top AI strategy, machine learning infrastructure, and the long-promised Siri revamp. The mandate couldn’t be clearer or more unforgiving: Rebuild trust after a rocky year for Apple Intelligence, and further an entire road map that will allow it to keep pace with the AI ecosystems developed by Google, Microsoft and OpenAI.
Who is Amar Subramanya, Apple’s new AI leader
Subramanya is known for being a delivery-focused engineering leader who has shipped consumer-scale AI systems across search, assistants and productivity workflows. At Google, he helped direct the engineering around Gemini Assistant, which involved cross-functional collaboration across foundation models, data pipes, safety review and product surfaces — exactly the type of cross-functional muscles Apple needs to flex as AI transitions from demos to daily utility on iPhone, iPad and Mac.
That adds another dimension because of his most recent experience at Microsoft: rolling assistant features into big cloud backends and enterprise-grade protections. That viewpoint may prove crucial in how Apple navigates between its on-device-first approach, and the there-when-needed practicality of generating cloud inference for more complicated tasks.
Why Apple is rewiring its AI brain after recent stumbles
Apple Intelligence — the suite designed to marry generative capabilities into system apps — rolled out with a whimper and quickly fell on its face. A notification summary facility created some false and embarrassing clippings, leading the BBC to lodge a public complaint after it erroneously reported that someone called Luigi Mangione, charged with a high-profile crime, had committed suicide and declared darts starlet Luke Littler the winner before he had even played in the final. The episodes underscored a larger problem: Apple’s guardrails and vetting pipes were not battle-tested.
It was the reboot of Siri that became the bigger flashpoint. A Bloomberg investigation described how internal tests led by senior executives discovered missing features weeks before a scheduled release date, leading to an indefinite postponement and provoking class-action lawsuits from buyers of iPhone 16s who were promised an AI-powered assistant. The same reporting described organizational disconnects — a broken communications channel between AI and marketing, budget discrepancies and an exodus of talent to OpenAI, Google and Meta. Some insiders even took to the biting nickname “AI/MLess.”
In the chaos, Apple reorganised key areas: Mike Rockwell, the exec in charge of Vision Pro, took control of Siri and the company cut back on projects under central AI. The addition of Subramanya is a second phase — one focused more on execution discipline and product polish than splashy reveals.
The Product And Platform Priorities Today
Apple’s AI strategy remains distinct. The company prefers to run models on-device with the Apple Silicon Neural Engine and only escalate to its Private Cloud Compute when necessary. That model maintains privacy and latency, but limits the size and capability of what models can be built versus competitors that rely on massive data center inference. The newest Apple chips prove the point: the A-series Neural Engine on iPhone is estimated at approximately 35 TOPS, and recent silicon for Mac-class systems comes in around 38 TOPS — enough to do summarization, translation and image understanding while still meager compared with multi-hundred-billion-parameter cloud models.
The trade-off is data. Apple’s unwillingness to scrape the data of users at scale means even greater reliance on licensed content, synthetic data and carefully curated corpora. That approach does not break privacy promises but may inhibit model improvement. Apple is reportedly planning to explore a broader use of Google’s Gemini for its next wave of Siri features in 2025, and to do so it’s relying on a vast amount becoming known only through third-party chatter, according to reports. Subramanya’s being so familiar with Gemini’s hot spots and failure modes could make him unusually effective at structuring any partnership, while building Apple’s own models behind the scenes.
Look for the first to be little more than updates on three near-term deliverables: a reliable Siri with crystal clear scope and demonstrable follow-through, system-level AI features that quietly save time without hallucinatory results, developer hooks that let third-party apps tap safely into what we can only presume will be called Apple Intelligence. On the platform side, that likely looks like more rigorous pre-release evals, red-teaming scaled across languages and accents, telemetry engineered for privacy-preserving feedback loops rather than raw data capture.
What success will look like for Apple’s AI reboot
Apple has the pieces: world-class silicon, distribution to a vast installed base and a privacy brand that still carries resonance. It’s also been quietly piling up research scaffolding — like the MLX framework for Apple Silicon and the OpenELM family of compact language models — that can speed on-device innovation. It’s been the repeatable execution that has been missing.
If Subramanya can deliver a Siri that completes multistep tasks in a reliable way, lower the hallucination rates for features like summaries and rewriting, and boost engagement without sacrificing privacy, Apple’s AI narrative can change quickly. Expect to see quantifiable gains in task completion, fewer escalations to the cloud and an acceleration in release cadence for core apps. A clean handoff from Giannandrea during the advisory window will smooth things over, but even that won’t leave Cook much time for course correction in a market where Google, Microsoft and OpenAI iterate at a breakneck pace.
Apple is betting that trust, rock-solid integration and best-in-class devices will eventually trump raw model sizes. With a leader who understands how competition constructs and ships it, it has an opportunity to prove as much.
