Apple Intelligence Foundation Language Models Tech Report 2025

Original Article ↗ Hacker News Discussion ↗

Contributor List & Anti‑Poaching Speculation

  • Some suspect the long, randomly ordered contributor list is meant to blunt poaching.
  • Others note large projects often use flat or alphabetical ordering, and with many non‑Latin names, English alphabet order is arbitrary anyway.
  • People point out it’s easy to find core researchers via references or LinkedIn, so any anti‑poaching effect is weak at best.

On‑Device 3B Model & Hardware

  • The ~3B-parameter model already runs on current iPhones, iPads, Macs and visionOS in the 26 betas, accessible via the Foundation Models framework and even Shortcuts.
  • Users report decent latency (few seconds) and that it can act as an offline chatbot despite being tuned for summarization, extraction, and writing tools.
  • Discussion of ANE vs GPU: GPU tends to be faster but less efficient; ANE is optimized for low power. Fitting 7B+ models in 8 GB RAM is seen as technically impressive but impractical for real use.

Siri & Apple Intelligence User Experience

  • Many comments complain that Siri still fails simple or multi-part requests, especially compared with ChatGPT or Gemini, and call Siri “a joke.”
  • Some note current Apple Intelligence models already power features like notification summaries and writing tools, but users often find these underwhelming or intrusive.
  • Multiple insiders-style comments describe attempts to bolt an LLM onto legacy Siri as an integration nightmare (multi‑turn flows, “smart” UI, privacy, backwards compatibility), arguing a full reset is needed.

Apple’s AI Strategy, Privacy, and Data Centers

  • Strong split in perception:
    • One side sees Apple as badly behind frontier models, over‑promising with “Apple Intelligence,” and now partly retreating to OpenAI/Anthropic APIs and legal/PR positioning (e.g., “responsibly sourced” training data, Private Cloud Compute).
    • The other side argues Apple should not chase frontier models; their differentiator is privacy, on‑device inference, and tight OS integration, leaving generic chatbots to apps.
  • Private Cloud Compute is praised as a technically strong privacy design, but skeptics doubt scalability and note that truly powerful, tool‑using agents will likely require large cloud models with lots of tokens and user context.

Training Data, Applebot, and Robots.txt

  • Apple claims not to use user private data, to filter PII/profanity, and to honor robots.txt for Applebot.
  • Critics argue Apple scraped before clearly documenting AI training use, making later “you can opt out” messaging feel disingenuous.
  • Others respond that robots.txt has long been the standard opt‑out mechanism and that expectations for advance crawler disclosure are new and mostly born of anti‑LLM sentiment.
  • Some propose adding crawler “categories” (e.g., search vs. LLM‑training) to robots.txt to give publishers finer control.

Alt Text, Accessibility & Training

  • The paper’s use of image–alt‑text pairs sparks debate: alt text is heavily advocated for accessibility, yet is now prime supervised data for vision‑language models.
  • Some see this as “free annotation labor” for AI; others argue it’s still morally consistent to write alt text for disabled users even if it’s scraped.
  • A few note they already use LLMs to draft alt text but still review edits manually.

Developer Experience & Structured Output

  • iOS developers are enthusiastic about the Foundation Models framework: typed Swift structs, guided/structured output, streaming partial fields, and a clean bridge to external models.
  • Commenters compare this to “structured output” / grammar‑based sampling already available elsewhere, noting that forcing strict structure can reduce model quality and sometimes needs a two‑pass “think then structure” approach.

Model Updates & LoRA Adapters

  • People wonder how often on‑device models will change; the base weights are gigabytes, so frequent silent updates seem unlikely.
  • Apple appears to rely on LoRA‑style adapters for specialization; these must be retrained per base‑model version, suggesting model changes will likely track OS point releases, not constant churn.

Is Apple Behind or on Its Own Track?

  • Critics frame Apple’s current AI as a “slow‑motion train wreck”: small models, weak Siri, lack of headline features vs. ChatGPT/Gemini, and RAM‑constrained hardware.
  • Defenders counter that:
    • Apple has always moved slowly and then integrated deeply;
    • They don’t need to win the model race, only own the device, UX, and privacy story;
    • Users who want frontier chatbots can easily install apps or use the OS‑level ChatGPT integration.
  • There’s broader pessimism about Apple’s product coherence post‑Jobs, contrasted with arguments that partnering (e.g., with OpenAI) fits a long history of Apple leveraging outside tech while keeping tight control of the platform.