IdeaWell

Project ideas from Hacker News discussions.

Apple picks Gemini to power Siri

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

6 Most Prevalent Themes

  1. Siri's Longstanding Poor Performance
    Users decry Siri as outdated and unreliable after 15 years.
    "Siri has been bad for what, 15 years now? It does seem like a bt of an outlier." – tibbar
    "Siri is utterly awful and that’s an understatement, for at least half a decade." – ProofHouse

  2. Pragmatic Outsourcing to Avoid AI Training Costs
    Apple wisely skips expensive frontier model training, treating LLMs as commodities.
    "Apple admitted that the cost of training SOTA models is a capex heavy-lift they don't want to own. Seems like they are pivoting to becoming the premium 'last mile' delivery network for someone else's intelligence." – Fiveplus
    "Models are becoming commodities, and their economy doesn't justify the billions required to train a SOTA model. Apple just recognized that." – elzbardico

  3. Lock-in Risks from Explicit Gemini Branding
    Public acknowledgment of Gemini hinders future switches, echoing Apple Maps failure.
    "The biggest NEW thing here is that this isn't white-labeled. Apple is officially acknowledging Google as the model that will be powering Siri. That explicit acknowledgment is a pretty big deal. It will make it harder for Apple to switch to its own models later on." – johnthuss
    "When Apple switched to their own maps it was a huge failure resulting in a rare public apology from the company." – johnthuss

  4. Limited Everyday Use of Siri
    Many users rarely or never use Siri, viewing it as niche or unnecessary.
    "You guys use Siri?" – hu3
    "I don’t think I used it once in my life and it’s one of the first thing I turn off every time I have a new device." – manuelmoreale

  5. Privacy Mitigated by Private Cloud Compute
    Data stays with Apple via on-device/PCC hosting, not Google servers.
    "This story says the custom model will run on-device and in Apple's Private Cloud Compute. The implication is that Google will not see the data." – WorldMaker
    "Apple Intelligence will continue to run on Apple devices and Private Cloud Compute, while maintaining Apple's industry-leading privacy standards." – Google (quoted)

  6. Apple's Strategic Timing and Strengths
    Apple waits for AI maturity, excels at integration/edge inference over training.
    "Apple is competent at timing when to step into a market and I would guess they are waiting for AI to evolve beyond being considered untrustworthy slop." – LexGray
    "Apple has the best edge inference silicon in the world (neural engine), but they have effectively zero presence in a training datacenter." – Fiveplus

🚀 Project Ideas

Siri LLM Intent De-anonymizer

Summary

  • A local, on-device utility that analyzes outgoing Siri requests before they are sent to Apple/Gemini servers.
  • It identifies and flags requests that contain personally identifiable information (PII) or sensitive data, explaining why the request might not be fully private and what data might be shared.
  • Core Value Proposition: Restores user agency and transparency in the "Private Cloud Compute" era by making the privacy boundaries of Siri explicitly visible.

Details

Key Value
Target Audience Privacy-conscious iOS/Mac users concerned about the new Gemini integration.
Core Feature Local network analysis (or screen overlay/extension) that flags requests containing PII (names, addresses, medical terms) before transmission.
Tech Stack Swift (for iOS/macOS Extension), Core ML (for on-device NLP/PII detection).
Difficulty Medium
Monetization Hobby

Notes

  • HN users are deeply skeptical of "privacy theater" (greentea23: The source code is closed so Apple can claim X but do Y). This tool provides the verification they crave.
  • Discussion value: High. It addresses the core tension in the thread: trusting Apple vs. trusting Google vs. trusting the government.
  • Practical utility: Helps users decide when to use the "opt-in" vs. "default" Siri features.

Private Cloud Compute (PCC) Simulator

Summary

  • An open-source simulation environment that mimics Apple’s Private Cloud Compute node architecture.
  • It allows researchers to audit the theoretical security claims of PCC by simulating how a request could be intercepted or logged if the "secure enclave" guarantees were hypothetically bypassed.
  • Core Value Proposition: A "Red Team" tool for the HN community to stress-test Apple's privacy claims without needing physical access to Apple's data centers.

Details

Key Value
Target Audience Security researchers, iOS developers, privacy advocates.
Core Feature Containerized simulation of PCC cryptography and attestation flow, allowing users to inject "malicious" logging code to test detection.
Tech Stack Docker, Go, Rust (for crypto primitives), Swift (for client-side attestation logic).
Difficulty High
Monetization Hobby

Notes

  • Addresses the debate around the CLOUD Act and "privacy theater" (tempodox: The gov’t can force them to reveal any user’s data and slap them with a gag order).
  • Allows the community to discuss the actual security mechanisms rather than just speculating on government subpoenas.
  • Fits the HN ethos of "verify, don't trust."

Siri Abstraction Layer (SAL)

Summary

  • A developer-facing framework that standardizes the interface between iOS apps and the underlying LLM (whether it's Apple's small model or the off-device Gemini/ChatGPT).
  • It detects which model is currently active and automatically adjusts the prompt engineering and tool-calling format to ensure consistent app behavior.
  • Core Value Proposition: Prevents "vendor lock-in" scenarios for developers where Siri's capabilities degrade or change drastically if Apple swaps LLM providers again.

Details

Key Value
Target Audience iOS App Developers, Shortcut power users.
Core Feature A Swift API that normalizes LLM responses and tool calls, hiding the complexity of which AI is generating the response.
Tech Stack Swift, App Intents Framework, Core ML.
Difficulty Medium
Monetization Revenue-ready: Enterprise licensing for large app developers.

Notes

  • Responds to Angostura’s point about needing an abstraction layer to swap models without losing quality.
  • Solves a real pain point for developers who don't want to rewrite their Siri integrations every time Apple changes backend providers (e.g., OpenAI to Gemini).

App Context Bridge

Summary

  • A local background service that aggregates "App Intents" and "Core Spotlight" data into a structured JSON dump accessible via a local API endpoint.
  • This allows the off-device LLM (Gemini) to query specific app data locally without the app needing to be specifically "LLM-ready" by the developer.
  • Core Value Proposition: Makes Siri smarter now by giving the off-device LLM immediate access to the full context of the user's installed apps, bypassing the slow adoption of App Intents by third-party developers.

Details

Key Value
Target Audience Power users, HomeKit automation enthusiasts.
Core Feature Local API server exposing app data (contacts, notes, third-party app data) to the LLM with strict permission boundaries.
Tech Stack Swift, SQLite, JSON, HTTP Server (SwiftNIO).
Difficulty Medium
Monetization Hobby

Notes

  • Addresses the frustration voiced by al_borland: The fact that Apple still prompts me to download Yelp to view images of a business is insane to me.
  • Also addresses the "dumb terminal" critique (Fiveplus): If the iPhone is a dumb terminal, this tool makes the terminal smarter by feeding better context to the brain.
  • Practical utility: Greatly enhances Siri's usefulness for home automation without waiting for developers to update their apps.

Gemini Data Scrubber

Summary

  • A pre-processing tool for the "Private Cloud Compute" that sanitizes user requests sent to the external Gemini model.
  • It acts as a middleware filter to remove specific data patterns (e.g., email addresses, phone numbers, health data) before the request leaves the PCC, ensuring that even Google does not see raw PII.
  • Core Value Proposition: True data privacy for Apple users by ensuring the external AI provider (Google) receives anonymized or pseudonymized inputs.

Details

Key Value
Target Audience Enterprise users, paranoid individuals.
Core Feature Pattern matching and redaction of PII from text/audio transcripts prior to LLM inference.
Tech Stack Swift (for integration), Python (for regex/NER models).
Difficulty Low
Monetization Hobby

Notes

  • Directly addresses the "Google seeing the data" concern (layer8: An abstraction layer doesn’t prevent Google from seeing the data).
  • Users like natch expressed distrust in Apple's shifting privacy language; this tool restores trust by applying an external verification layer.
  • Discussion: Sparks debate on whether this level of filtering degrades LLM performance (hallucinations).

Siri Utility Switchboard

Summary

  • A configuration tool that allows users to granularly route specific Siri queries to specific AI providers (e.g., Timers/Reminders to Apple's on-device model, Web Search to Gemini, Coding questions to ChatGPT).
  • It mimics the functionality of iOS 18.2's ChatGPT opt-in but for the new Gemini integration and adds custom logic for routing.
  • Core Value Proposition: Gives users the "opt-in" granularity that HN users demand, preventing the "forced" integration of Google services that johnthuss warned about.

Details

Key Value
Target Audience iOS Power users, privacy advocates.
Core Feature Rule-based routing engine for Siri requests based on keywords, intent, or app context.
Tech Stack Swift, Shortcuts API, iOS Intents.
Difficulty Low
Monetization Hobby

Notes

  • Addresses the concern about Siri becoming a "dumb terminal" for Google. Users can still keep local tasks local.
  • Responds to the user frustration about Siri "kicking me out to Google search result pages to show results" (al_borland).
  • High discussion potential regarding the UI/UX of managing these complex rules within Apple's walled garden.

Offline Siri Bridge (PCC Local)

Summary

  • A tool that attempts to run the "Small" Apple Intelligence model locally on older hardware (or Macs) that lack the Neural Engine required for official support, acting as a bridge to handle basic commands without reaching out to the cloud.
  • It intercepts "off-device" calls and routes them to local open-source models (like Phi-3 or Qwen) if the user prefers not to use Gemini for low-stakes queries.
  • Core Value Proposition: Extends the life of older Apple hardware and provides a truly offline fallback for basic commands, mitigating reliance on the cloud.

Details

Key Value
Target Audience Owners of older iPhones/Macs, users with poor internet connectivity.
Core Feature Local model hosting and request interception for basic Siri intents (Timers, Alarms, basic math).
Tech Stack Python (for model hosting via Ollama), Swift (for request interception via network proxy).
Difficulty High
Monetization Hobby

Notes

  • Addresses the "competence" issue (IOT_Apprentice: It appears to be lack of competence given they lied about the initial features of Apple Intelligence).
  • Provides a solution for users like rootusrootus who only use Siri for simple tasks (opening/closing the garage door, setting timers) but don't want cloud dependencies.
  • Discussion: Validates the local-first approach while acknowledging the hardware limitations Apple imposes.

Read Later