Behind Apple AI's Repeated Delays: The 14-Year Internal Struggle of Siri Revealed

From Grand Promises to Broken Deadlines: Apple Intelligence's Embarrassing Reality

At WWDC 2024, Apple unveiled Apple Intelligence with unprecedented fanfare. The "new Siri that truly understands you" had countless consumers buzzing with excitement. Yet a year later, Apple quietly informed media through a spokesperson that the key new Siri features showcased at last year's keynote might not actually ship until 2026.

The core capabilities demonstrated at the Apple Intelligence launch included: cross-app contextual understanding (such as reading email content and automatically linking it to calendar events), personalized semantic search (building a personal knowledge graph from user data), and multi-step task planning (like automatically planning a pickup route based on flight information). These features fundamentally require a powerful on-device + cloud collaborative large language model architecture that not only possesses general reasoning capabilities but also needs to access users' private data within a strict privacy framework. Apple designed the Private Cloud Compute architecture for this purpose, promising that user data would be processed on dedicated Apple Silicon servers and immediately destroyed afterward, inaccessible to Apple or any third party. The implementation difficulty of this technical promise far exceeds that of traditional cloud-based AI services.

Even more baffling, according to former Apple employees, the features demonstrated at WWDC—reading emails, planning routes, reasoning about schedules—may never have actually worked. Perhaps the only real thing in the entire demo was the colorful ribbon animation with its purple glow that appeared when Apple Intelligence was activated.

When the news broke, the capital markets responded first: Apple's stock plummeted 4.85% overnight. Consumers were even more furious—many had purchased the iPhone 16 specifically for Apple Intelligence, only to be told after more than six months that the core features would require another year of waiting. Class-action lawsuits followed.

Seeds of Trouble: Siri's Predicament After Jobs' Departure

The Fall of a Pioneer

Apple once held a commanding lead in the voice assistant space. As early as 1984, when the original Macintosh was unveiled, Jobs had the computer introduce itself using speech. In 2010, he personally greenlit the acquisition of the fledgling Siri company and integrated it into the iPhone 4S—the "S" stood for Siri.

Siri originally emerged from the CALO project at SRI International, a cognitive assistant project funded by DARPA (Defense Advanced Research Projects Agency), with the full name "Cognitive Assistant that Learns and Organizes." Early Siri used a rule-based and statistical model natural language understanding (NLU) system, parsing user commands through Intent Classification and Slot Filling. While this architecture performed precisely in specific scenarios, it scaled terribly—every new feature required manually writing extensive rules and training data. In contrast, Google's later end-to-end neural network approach could automatically learn semantic understanding capabilities from massive datasets, with scaling costs far lower than Siri's traditional architecture.

But fate had other plans. Apple acquired Siri in 2010, and Jobs passed away in 2011. Combined with the server crashes caused by users flooding in when Siri launched, the Siri team found itself in an increasingly awkward position within the company.

The Battle Over Direction and Endless Infighting

After Jobs' departure, the fundamental question of whether Siri should be a "personal assistant" or a "universal search engine" remained unresolved for years. Williamson, who took over as Siri's lead, insisted on aligning Siri's update cadence with iOS—once a year. But Siri is fundamentally an online AI service that could be continuously optimized. The annual update rhythm caused it to quickly fall behind competitors.

Tying an AI service to the iOS annual update cycle meant that Siri's model improvements, new feature launches, and bug fixes all had to wait for the major iOS release each fall. Competitors like Google Assistant and Amazon Alexa adopted a Continuous Deployment model where backend models could be updated weekly or even daily, with users receiving improvements without upgrading their systems. This difference was especially fatal in an era of rapid AI iteration—when the evolution from GPT-3 to GPT-4 took just over a year, Siri was trapped in an annual release cadence unable to respond quickly to technological change.

Even more dramatic, in a 2018 interview, Williamson publicly complained that "Siri was a disaster at launch," shifting blame to the original team. Siri's original CEO Dag Kittlaus fired back passionately on Twitter, pointing out that Siri's launch issues were merely server overload, while Apple Maps under Williamson's watch was "the real product disaster in Apple's history."

Apple executives publicly feuding online—this was extremely rare in Apple's history. Before leaving, Dag delivered one devastating parting shot: "If Steve were still alive, I'd probably still be working at Apple."

Just one year after launch, Samsung's S-Voice was already on par with Siri; by 2014, Google Now had surpassed it. Amid infighting and playing catch-up, Siri faded into mediocrity.

The Blind Leading the Blind: Apple's AI Team's Seven Years Adrift

Good Guy Old John's Laid-Back Management

In 2018, Apple poached AI chief John Giannandrea from Google while simultaneously building a unified AI platform department, attempting to end Siri's chaos. The strategy seemed brilliant, but Old John was no Steve Jobs.

Compared to other Apple executives, Old John seemed out of place. Among the Senior Vice Presidents listed on Apple's website, almost everyone except him and a legal counsel had spent decades at Apple, with work styles leaning toward iron-fisted, high-demand, and fast-paced. Old John, by contrast, was described by colleagues as "quiet, easygoing, conflict-averse, and a good listener."

His technical direction was equally laid-back—he believed that as long as they steadfastly pursued machine learning research, Siri would "gradually improve like climbing a mountain." When ChatGPT burst onto the scene, he even reassured employees that "chatbots won't deliver much value to users," showing zero sense of urgency or decisiveness.

After ChatGPT's release in November 2022, the entire tech industry's AI competitive landscape was fundamentally reshaped. ChatGPT's Transformer-based large language model demonstrated "Emergent Abilities"—when model parameter scale breaks through certain thresholds, it suddenly acquires capabilities not explicitly included in training objectives, such as logical reasoning, code generation, and multilingual translation. This paradigm shift meant that the traditional approach of "training specific models for specific tasks" (precisely the path the Siri team had long pursued) might already be obsolete. A sufficiently large general-purpose model could accomplish virtually any language task through Few-shot Prompting, directly challenging Apple's years of incremental improvement strategy in machine learning. Old John's misjudgment of this paradigm shift became one of the most fatal errors in Apple's AI strategy.

Wavering Direction and Team Disillusionment

According to former employees, the AI team initially planned to build two models for Apple: a small model running locally on iPhone for simple tasks, and a large model in the cloud for complex tasks. This approach preserved both privacy and capability—very "Apple."

From a technical perspective, this "edge-cloud collaborative inference" architecture was designed as follows: a small model (approximately 3 billion parameters) deployed on iPhone's Neural Engine, handling simple queries, text summarization, and quick responses with low latency; a large model (hundreds of billions of parameters) deployed in the cloud, handling complex reasoning and multi-step planning requiring substantial computing power. The core advantage: simple tasks completed entirely on-device with user data never leaving the device, protecting privacy while reducing latency; only complex tasks would go to the cloud. Apple's eventually released Apple Intelligence actually returned to this concept, using an approximately 3-billion-parameter on-device model paired with a cloud-based large model via Private Cloud Compute—but the directional wavering in between wasted precious development time.

Before long, the plan was scrapped in favor of using a single model to solve everything—contradicting the team's previously held privacy policies and directly causing some disillusioned employees to leave.

Old John's deputy, Robby Walker, was equally exasperating. His greatest achievement was spending two and a half years reducing Siri's wake response time by 33% while maintaining system safety and stability—in plain terms, removing the "Hey" from "Hey Siri."

Under this pair's leadership, the entire AI and Machine Learning team (AIML) earned a sarcastic nickname within Apple: "Aimless."

Departmental Warfare: The Open Conflict Between AI and Software Engineering

Old John's nice-guy management style created another problem: his employees got faster promotions and longer vacations, with the sole exception being work output keeping pace. This "inequality breeds resentment" situation caused relations between the AI team and Craig Federighi's Software Engineering team to grow increasingly tense.

The AI team thought the Software team was "too stubborn and demanding," while the Software team felt they were "cleaning up after the AI team." Earlier this year, the AI team even required engineers to leave written documentation on collaborative projects "in case blame gets shifted"; the Software team simply spun up their own "Intelligent Systems" team, bypassing the AI department to work on AI independently.

This internal war even affected Vision Pro. In 2022, Vision Pro lead Rockwell wanted both teams to collaborate on voice interaction features. The AI team developed slowly and failed to deliver a single planned feature, with meetings essentially turning into struggle sessions against the AI team.

Better Late Than Never: Apple's Path to Self-Rescue

In March 2025, Tim Cook finally made the decisive call. The Siri engineering department was separated from the AI team and placed entirely under Vision Pro lead Rockwell, reporting directly to Software team chief Craig. Old John was stripped of authority entirely—even his secret robotics division was transferred to someone else.

On the hardware front, Apple is also accelerating its efforts. Since the A11 Bionic chip, Apple has included a Neural Engine specifically designed to accelerate machine learning inference tasks. By the A17 Pro and M4 chips, the Neural Engine has reached 35 trillion operations per second (35 TOPS). However, the computing power required to train large language models is on an entirely different scale from inference—OpenAI's training of GPT-4 reportedly used approximately 25,000 NVIDIA A100 GPUs over several months. Apple had originally planned to co-design custom AI chips with Broadcom (expected to enter mass production in 2026), primarily targeting data center training and inference scenarios to reduce dependence on NVIDIA. But since chips typically require 2-3 years from design to mass production, facing urgent market pressures, Apple has pivoted to directly purchasing NVIDIA's H100/B200 and other off-the-shelf GPU clusters to build AI infrastructure and accelerate its catch-up.

Conclusion: Too Many Nice Guys, Too Few Steve Jobs

Apple's stumble in AI is fundamentally a case study in organizational management failure. When a company faces a technological revolution, a "nice guy culture" can be more dangerous than a "bad product"—everyone's first instinct is to play it safe and avoid mistakes, rather than seizing opportunities and breaking conventions like Jobs would have.

Changing the world isn't won by executing every step perfectly—it demands determination, boldness, and the courage to act against the current. Apple has now woken up, but its window of opportunity is narrowing rapidly—consumers have waited a year for nothing, and competitors are taking turns poaching Apple's users. Whether Apple can reclaim the legacy Jobs left behind in this AI race remains an open question.