TL;DR
Apple is replacing Siri’s underlying architecture with a new large language model system, code-named "Project Greymatter," aiming to close the gap with ChatGPT and Google Assistant. This matters right now because Siri has fallen so far behind competitors that Apple’s entire ecosystem strategy—from iPhone to HomePod to Vision Pro—depends on a credible voice assistant by the 2026 holiday season.
What Happened
Apple is overhauling Siri from the ground up, replacing its decade-old neural network with a new large language model system internally called "Project Greymatter" — a move that The Wall Street Journal reports is the most significant AI investment in the company’s history. The overhaul, expected to be previewed at Apple’s Worldwide Developers Conference on June 22, 2026, aims to give Siri the ability to hold multi-turn conversations, execute complex app commands, and generate text responses rather than simply fetching web results.
Key Facts
- Apple has allocated over $5 billion to Project Greymatter, including acquiring at least three AI startups in 2025 and 2026 focused on conversational AI and on-device inference.
- The new Siri will run a 70-billion-parameter language model on-device for iPhone 17 and later models, with cloud fallback to a 500-billion-parameter model for complex queries.
- Anthropic is facing internal turmoil over its "Fable" project—a consumer chatbot product—after 40% of its safety team resigned over disagreements about deployment speed versus safety testing.
- Meta is offering $1.2 million annual retention bonuses to top AI researchers working on its Llama 4 large language model, up from $800,000 in 2025, according to internal compensation documents.
- Deepfake porn creation has surged 550% since 2023, with 96% of all deepfakes being non-consensual explicit content, according to the 2026 Deepfake Pornography Report from the AI Safety Institute.
- Google has already integrated its Gemini model into Assistant, achieving 40% higher user satisfaction scores than Siri in internal benchmarks leaked in March 2026.
- The European Union’s AI Act, fully enforceable starting August 1, 2026, will require all voice assistants to disclose when users are interacting with AI and provide opt-out mechanisms for data training.
Breaking It Down
The Siri overhaul represents a belated but necessary admission from Apple that its original approach—a rules-based assistant with limited machine learning—has failed to keep pace with the generative AI revolution. Since its 2011 debut, Siri’s underlying architecture has been patched with upgrades to speech recognition and limited question-answering, but it never received the foundational rewrite that competitors like Google and OpenAI have invested in. Project Greymatter is that rewrite: a full migration from Siri’s legacy system to a transformer-based architecture that can understand context, remember previous questions, and execute chains of commands like "Find the Thai restaurant we went to last month and text Sarah the address."
Apple has lost an estimated 18 percentage points of smart speaker market share since 2022, falling from 35% to 17% as Amazon Alexa and Google Assistant integrated generative AI features that Siri could not match.
The internal pressure is immense. Apple’s services revenue—which hit $95 billion in fiscal 2025—is increasingly tied to ecosystem lock-in, and a weak voice assistant undermines the value proposition of Apple’s hardware. The $3,499 Vision Pro, in particular, relies on voice and gesture input; without a competent Siri, the device’s primary interface is crippled. Apple is betting that on-device processing—using its own A18 and M4 chips with dedicated neural engines—will give it a privacy advantage over cloud-dependent rivals, but that advantage only matters if the assistant actually works.
Anthropic’s Fable woes illustrate the tension ripping through the AI industry. The startup, once celebrated for its safety-first approach, is now dealing with a 40% resignation rate from its safety team after pushing Fable to beta in March 2026 without completing red-teaming on speech generation and memory features. The resignations have delayed Anthropic’s planned $3 billion funding round, with investors demanding proof that the safety exodus won’t lead to regulatory liability. This is the same dilemma facing every major AI lab: move fast and risk safety blowback, or move slow and lose market share to less cautious competitors.
Meta’s $1.2 million retention bonuses for Llama 4 researchers are a direct response to the talent war. OpenAI is reportedly offering $2 million total compensation packages for senior researchers, while Google DeepMind has created a "superstar" track that bypasses normal salary bands. The bonuses are not just about keeping people—they are about keeping secrets. Meta’s Llama 4 is expected to be released under a more restrictive license than Llama 3, as the company seeks to monetize its open-source strategy without giving away its best models for free.
What Comes Next
-
WWDC 2026 (June 22–26): Apple will preview Project Greymatter’s capabilities, likely demonstrating multi-app workflows and on-device privacy features. The actual public rollout is expected with iOS 20 in September 2026, but only for devices with A17 or newer chips—leaving hundreds of millions of older iPhones with the legacy Siri.
-
Anthropic board meeting (July 15, 2026): The board will vote on whether to slow Fable’s rollout or accelerate it to capture market share from OpenAI’s GPT-5. The outcome will determine whether Anthropic remains a safety leader or pivots to a growth-at-all-costs strategy.
-
EU AI Act enforcement (August 1, 2026): Apple, Google, Amazon, and Anthropic must all have compliance systems in place. For Apple, this means Siri must disclose its AI nature—potentially undermining the "natural" assistant experience it has marketed for years.
-
Deepfake porn legislation vote (September 2026): The U.S. Senate is expected to vote on the "No AI Fakes Act" which would criminalize non-consensual deepfake porn and require platforms to remove such content within 24 hours of notification. The bill has bipartisan support but faces opposition from free-speech groups.
The Bigger Picture
This story sits at the intersection of three major technology trends. First, the Voice Assistant Arms Race has entered a new phase where raw intelligence—not just accuracy—is the differentiator. Siri’s overhaul is Apple’s attempt to catch up, but the company is starting from a position of weakness: Google Assistant and Alexa already have generative AI integrations, and OpenAI’s ChatGPT Voice mode has set a new bar for natural conversation. Second, the AI Safety Backlash is real and accelerating. Anthropic’s internal crisis and the deepfake porn surge are two sides of the same coin: the technology is advancing faster than society’s ability to manage its harms. The 550% increase in deepfake porn since 2023 has made this a mainstream policy issue, with both the EU and U.S. moving toward regulation that will reshape how AI models are trained and deployed.
Finally, the Talent War for AI Researchers is distorting the entire tech labor market. Meta’s $1.2 million bonuses are not outliers—they are becoming the baseline. This concentration of talent at a handful of companies (OpenAI, Google, Meta, Anthropic) risks creating an AI oligopoly where only the wealthiest players can compete. Smaller startups and academic institutions are being starved of the very people who could develop alternative approaches to AI safety and alignment.
Key Takeaways
- [Siri’s Last Stand]: Apple’s Project Greymatter is a make-or-break effort to salvage its voice assistant business, with a $5 billion investment and a 70-billion-parameter on-device model that must ship in September 2026.
- [Safety vs. Speed]: Anthropic’s 40% safety team resignation rate over Fable deployment shows the AI industry’s fundamental conflict between moving fast and preventing harm—a conflict that will only intensify as regulation tightens.
- [Talent Escalation]: Meta’s $1.2 million retention bonuses for Llama 4 researchers signal that AI talent costs are spiraling, with top researchers now commanding $2 million+ total compensation packages.
- [Deepfake Crisis]: The 550% surge in deepfake porn since 2023 has made non-consensual AI-generated content a top-tier policy issue, with the No AI Fakes Act and EU AI Act set to impose real consequences on platforms and developers.
