A look at Apple's technical approach to AI including core model performance etc.
Apple’s newly announced “Apple Intelligence” strategy is prompting debate over whether polished, tightly integrated but slightly behind state-of-the-art models can beat more powerful, generic services like ChatGPT. Commenters highlight Apple’s focus on on-device inference, privacy-preserving “Private Cloud Compute,” and deep OS-level integration as potential long‑term advantages, while questioning constraints like limited RAM and the reliance on OpenAI for some queries. Others remain skeptical that these AI features are truly novel or will materially boost iPhone sales, contrasting Apple’s consumer-centric approach with Nvidia’s explosive growth selling training hardware to enterprises.
State of the Art vs Product Fit
- Several argue Apple is intentionally choosing “good enough” models with polished UX over chasing top benchmark scores.
- Others counter that today’s SOTA (e.g., leading LLMs) is also where most bugs are ironed out, so lagging on SOTA risks shipping an inferior assistant again.
- Some note that many valuable features don’t need SOTA; integration, context, and UX matter more.
Local Models, Privacy, and On-Device Focus
- Strong support for Apple’s emphasis on local inference for privacy and latency, even if models are smaller/weaker.
- Local processing is seen as more aligned with Apple’s brand and user expectations, particularly for highly personal data on phones.
- There is interest in adapter-based approaches (LoRA-like “skills”) and multi-agent / tool-use integrations at the OS level.
Role of OpenAI / ChatGPT
- Opinions split: some view integrating ChatGPT as a minor, almost unnecessary fallback for “party trick” use cases.
- Others note OpenAI’s strength is largely B2B via APIs rather than a consumer app.
- Some see Apple’s keynote treatment of ChatGPT as a symbolic downgrade of pure SOTA chatbots.
Hardware, Nvidia, and Apple Silicon
- Debate on whether Apple’s edge-centric AI undermines the Nvidia “GPU gold rush”; most agree Nvidia’s surge is from selling training GPUs, which Apple doesn’t.
- Some speculate Apple’s server-side Apple Silicon (used for Private Cloud Compute inference) has Nvidia-like efficiency per watt and could hint at future server hardware, others think Apple will never sell such hardware broadly.
- It’s noted Apple reportedly trained models on non-Nvidia hardware (e.g., TPUs), reinforcing that Nvidia isn’t strictly required.
- Concerns raised about Apple’s RAM pricing and low default RAM undermining on-device AI potential.
Impact on Users, Platforms, and Upgrades
- Some think integrated, context-rich assistance (calendar, mail, photos, home automation) could become the best consumer AI experience and drive deeper ecosystem lock‑in.
- Others doubt AI features will materially change iPhone upgrade behavior; camera, screen, and obvious performance still dominate for most users.
- Mixed views on whether this will attract Android switchers; some interest reported, but many expect Android to match features quickly.
Novelty, Hype, and Skepticism
- Several commenters see little that’s conceptually new; features resemble existing capabilities on Android, Google Photos, Samsung, WhatsApp stickers, etc.
- Others argue the novelty is in breadth and depth of OS‑level integration, not any single feature like emoji or image generation.
- Some view the article and keynote as overly positive or “fan”‑like and question how much is real vs. marketing.
- Past disappointments with Siri fuel skepticism; many adopt a “wait until shipping” stance.
Private Cloud Compute & Privacy Guarantees
- Apple’s Private Cloud Compute is discussed as a way to offload heavy tasks while keeping data ephemeral and non-attributable, using Apple Silicon with secure boot/enclave.
- Exact details of what context is sent (full images vs. extracted features, single vs. multiple photos/texts) remain unclear.
- Some argue it’s better to invest in strong infrastructure and auditing rather than prematurely freezing strict constraints on context.