Apple reveals new AI architecture built around Google Gemini models

Original Article ↗ Hacker News Discussion ↗

Apple’s new “Apple Intelligence” stack blends its own on-device foundation models with a high-end Gemini-based cloud model running on NVIDIA GPUs in Google Cloud, raising questions about how much of Siri’s new capability is really Apple’s versus Google’s. Commenters focus on privacy claims around Apple’s Private Cloud Compute, debating whether confidential-computing hardware and external audits meaningfully protect user data when third-party infrastructure and legal backdoors are involved. The move also intensifies scrutiny of Apple’s decision to delay the features in the EU under the Digital Markets Act, its reliance on a direct mobile competitor instead of OpenAI or Anthropic, and whether tightly integrated but closed assistants should allow users to choose alternative AI providers.

Architecture & Model Choices

  • Apple described five “Apple Foundation Models” (AFMs): two on‑device (Core, Core Advanced) and three cloud (Cloud, Cloud Image, Cloud Pro).
  • Cloud Pro is said to be “Gemini frontier‑level” and runs on NVIDIA GPUs in Google Cloud under Apple’s Private Cloud Compute (PCC); others run on Apple Silicon.
  • Everything except Cloud Pro is described as custom Apple models “refined” using Gemini; commenters speculate this means some form of distillation or fine‑tuning, but details are unclear.
  • Some links claim earlier on‑device AFM was ~3B parameters; people note the new stack is “more complicated” than that.

Gemini Quality & Hallucinations

  • Several posters complain that public Gemini (especially search “AI mode”) is inaccurate and hallucinatory; some prefer Claude or ChatGPT.
  • Others distinguish Google Search’s AI mode from the Gemini app/API, saying the latter is much better, especially paid tiers (e.g., Ultra).
  • Local Gemma‑based models are praised on phones, but some report current iOS local Gemini is slow, hot, and still hallucinates.

Privacy, Private Cloud Compute, and Trust

  • Apple claims: on‑device first, PCC for offload, data only used per request, and not accessible to Apple or third parties; security researchers can verify via published PCC design.
  • PCC uses confidential computing (Apple Silicon, and now also Intel/NVIDIA on Google Cloud) plus OHTTP‑style relays; keys are held so operators allegedly can’t inspect data.
  • Supporters call this the best available privacy architecture for off‑device inference.
  • Skeptics point out: users must still trust Apple (and underlying hardware vendors, and jurisdiction); nation‑state backdoors, zero‑days, or legal compulsion remain possible; some call it “security theater” unless independently and continuously audited.

EU DMA, Regulation, and Feature Delay

  • Siri AI / Apple Intelligence is delayed in the EU. Apple blames the DMA’s requirement for parity of access to device data and actions for third‑party assistants.
  • One side: giving “any AI app” Siri‑level permissions (read all personal data, control apps) is too risky; Apple is right to resist.
  • Other side: this is about preserving lock‑in and avoiding competition; users should be allowed to choose other assistants with explicit permissions, as with contacts/photos today.
  • Debate over paternalism vs user autonomy; some argue safeguards and system prompts would suffice, others think average users will be tricked by Meta‑style dark patterns.

Why Google (and Not Others)?

  • Reasons discussed:
    • Google’s strength in small/edge models (Gemma, on‑device Gemini) and prior edge‑AI work.
    • Massive compute capacity (TPUs, GPUs, data centers) and willingness to host Apple’s own AFMs under PCC.
    • Existing multibillion‑dollar search partnership and perceived corporate stability vs newer labs.
  • Many think models are becoming commodities; the real differentiation will be Apple’s integration, orchestration, and UX.

User Choice, Lock‑in, and Third‑Party Assistants

  • Some want a system‑level way to plug in alternative models (Claude, Mistral, DeepSeek, self‑hosted) behind the same Apple Intelligence APIs.
  • Others argue Apple won’t (and isn’t required to) operate everyone’s models in PCC, and that open routing to arbitrary clouds would weaken privacy guarantees.
  • DMA debate recurs here: whether “same access as Siri” should apply only to on‑device APIs, or also to cloud backends.

Siri, UX, and App Integration

  • Many say legacy Siri was “terrible” and are skeptical that reusing the Siri brand will change perception.
  • Technical talks about App Intents and Shortcuts suggest a deep agentic layer: AI can coordinate across apps, change passwords, book travel, etc.
  • Some are excited about OS‑level integration that third‑party chatbots can’t match; others fear brittle automation (e.g., AI mis‑changing passwords) and complex failure modes.

Views on Apple’s Overall AI Strategy

  • Critique: Apple is “weirdly behind,” relying on others for core AI, showing loss of innovation leadership and over‑focus on operations.
  • Counter‑view: Apple wisely avoided burning tens of billions on training frontier models; now it can rent or co‑develop models once the landscape is clearer.
  • Several note Apple has long positioned itself as a hardware+integration company, not a search/AI lab; treating the model as a swappable implementation detail fits that philosophy.