Could Apple’s biggest AI problem be the very strategy it’s built on? The sudden upheaval at the top replacing John Giannandrea with Amar Subramanya is a change not so much of personnel as it is an admission that Apple Intelligence, the generative AI flagship suite it released in October 2024, just hasn’t gone as planned, revealing deeper engineering and organizational cracks.
Subramanya brings impeccable qualifications to the table: 16 years at Google, during which time he was the engineering lead for the Gemini Assistant, then a period at Microsoft as corporate vice president of AI, working on foundation models that power Microsoft Copilot. At Apple, his role encompasses Apple Foundation Models, ML research, and AI Safety and Evaluation, and he reports directly to Craig Federighi. That is to say, he’s the one shaping the technical architecture that enables features powered by AI in the iPhone, Mac, iPad, and all Apple services.
The need for it is obvious. There were a number of high-profile failures that marred the launch of Apple Intelligence, with perhaps its most notorious failure being the notification summary feature that condensed alerts into brief headlines. In several instances, it fabricated and defamatory statements e.g., stating that BBC‑reported let Luigi Mangione shot himself when he was the one who was shot, or that darts player Luke Littler had won the championship before the final even started. Formal complaints by the BBC, New York Times, and other sources saw Apple halt the feature at beta releases and revise it with italicized text and error indications. These episodes were not just PR fiascos; they signaled the existence of weak summarization pipelines and a lack of support for AI‑generated content.
The overhaul of Siri didn’t work out, either. It was Federighi’s own testing that, weeks before the scheduled release in April 2025, showed that announced features did not work at all. The delay caused a class‑action lawsuit from iPhone 16 buyers, claiming that Apple advertised features falsely. Internally, a restructuring occurred with Siri being taken out of Giannandrea’s domain and given to the creator of Vision Pro Mike Rockwell. Bloomberg’s articles mentioned the causes of the dysfunction to be issues with communication between teams, problems with the budget, and employees leaving for AI competitors, while at one point workers joked that their group should be called “AI/MLess.”
Apple’s approach involves privacy-first on-device AI, which is at the heart of the battle. Most of the work is done locally on Apple Silicon’s Neural Engine, using methods such as LoRA adapters and heavy quantization-down to 3.7 bits per weight to accommodate multi-billion-parameter models within mobile resource limits. This is done to prevent sending sensitive information to the cloud, assured by Private Cloud Compute servers that encrypt and delete user content immediately after processing. But there is a significant trade-off herein: the on-device models are smaller and less powerful than the large, cloud-hosted LLMs from Google, Microsoft, or OpenAI.
The training data strategy further widens the gap. Foundation models of Apple are built from the archives of licensed publishers, curated public datasets, and content taken from the web through Applebot all filtered for quality and safety but without private user data. To compensate for the gaps, Apple relies heavily on synthetic data generated by auxiliary models for math, code, multilingual tasks, and safety alignment. Although synthetic data can solve issues of scarcity and compliance, industry research has benchmarked its limitations: it is hard to control outputs, while accuracy and representation of underrepresented groups may be poor, and validation is done primarily via manual “spot‑checking” as opposed to rigorous benchmarks. This may result in weak generalization and slight bias in downstream systems.
In order to get more from limited hardware, Apple’s engineering teams have worked on architectural changes. The on‑device model’s 3B parameters are divided into two blocks; thus, memory use and first‑token latency are reduced by 37.5% without any quality loss. Regarding cloud inference, Apple’s Parallel‑Track Mixture‑of‑Experts design can do this by having only the relevant subnetworks activated for a given prompt; this way, bottlenecks are lessened, and efficiency is increased. Significant extension of support for different languages has been done-the amount of training data in non‑English languages increased from 8% to 30%, and the tokenizer vocabulary has grown from 100K to 150K tokens-thus measurable improvements have been made in both accuracy and fluency for the tested areas.
However, by embracing these optimizations, Apple’s conservative step-by-step approach still sharply contrasts with quick cloud‑based iteration cycles of its competitors. For instance, Google’s Gemini and Microsoft’s Copilot are the results of training on huge, real‑world datasets and are rapidly deployed in data centers; thus, they have richer reasoning capabilities with broader contextual understanding. Since Apple refuses to harvest user data, it needs to come up with workarounds for inherent scale disadvantages-a decision that guarantees privacy at the risk of being less capable.
Subramanya has to figure out a way to harmonize these limitations with what the market expects. His role at Google and Microsoft, where he was the link between research and product, might be of help to Apple in that it facilitates tighter integration between Siri, Apple Intelligence, and the ecosystem in general. However, with the lawsuits hanging over his head, the features being delayed until 2026, and the rivals accelerating, the room for making mistakes is very narrow – and the engineering trade‑offs that are part of Apple’s AI philosophy will be put to test as never before.
