687: You Can Bend This Line

What It Covers

Accidental Tech Podcast episode 687 covers Apple's expansion of advertisements into Maps, the legal ambiguity surrounding AI-generated code and copyright, a controversial open-source relicensing attempt using LLM-generated code, Apple's broader trend of degrading user experience for incremental services revenue, and fitness watch measurement discrepancies between competing wearable platforms.

Key Questions Answered

• •Apple Maps Ads Revenue Reality: Apple's ad sales infrastructure is technically primitive compared to Google's, meaning Maps ads will likely generate minimal revenue relative to the user experience cost. The App Store search ads product, which launched earlier, already demonstrates this pattern — it primarily forces developers to bid on their own brand keywords just to reach existing customers, creating a double-dip alongside Apple's existing 30% transaction cut.
• •Open-Source LLM Relicensing Risk: A maintainer of the Python library Chardet released version 7.0.0 as a complete LLM-generated rewrite, switching the license from LGPL to MIT without legal authority to do so. The original author, Mark Pilgrim, returned from internet retirement to contest this. Developers relying on LGPL-licensed libraries should audit whether maintainers have attempted similar relicensing, as this tactic may spread and creates unresolved legal exposure for downstream users.
• •AI Code Copyright Remains Legally Unresolved: Despite widespread claims online that AI-generated code cannot be copyrighted based on Supreme Court rulings about AI art, no court has specifically adjudicated this for code. Art, code, and patents have historically received distinct legal treatment. Developers building products with AI coding agents should not assume copyright ownership is settled law — the question remains open and will likely be forced into courts by large commercial interests soon.
• •Clean Room Implementation Standard Challenged by LLMs: The legal precedent for clean room reimplementation — established when Compaq cloned the IBM PC by isolating engineers from original source material — requires that implementers have zero exposure to the original code. LLM-based rewrites fail this standard because the model itself was trained on the original code, and the prompting developer had direct prior exposure. This distinction matters for any team attempting to use AI to reimplement GPL-licensed software under a permissive license.
• •Platform Degradation Trend Line Matters More Than Individual Incidents: Each individual Apple ad placement or upsell prompt in system settings appears minor in isolation, but the directional trend is what signals long-term brand erosion. Apple's leadership approves these incremental revenue moves because individual ROI is measurable — AppleCare upsells in Settings provably increase sales — while brand damage is not quantifiable in the same meeting. Developers and users should evaluate Apple's platform trajectory by trend direction, not snapshot quality.
• •Chip Binning Strategy Creates Supply Constraints: The MacBook Neo's A18 Pro chip uses a five-of-six GPU core configuration sourced from binned chips — units where one GPU core failed quality testing. Demand has outpaced Apple's initial projection of five to six million units, exhausting the binned supply. Apple's options include paying a premium to restart TSMC's N3E production lines at maximum capacity, reallocating chips from other products, or quietly shipping units with all six GPU cores enabled — a roughly 16% GPU performance difference.
• •Fitness Watch Calorie and Distance Measurements Diverge Across Platforms: Wearing two fitness watches simultaneously — in this case a Suunto and a Garmin — produces measurable differences in heart rate, calorie estimates, and GPS distance tracking. Calorie counts in particular can shift significantly when switching watch platforms, similar to the variation seen when changing between Apple Watch models. Users switching fitness tracking platforms should expect a recalibration period and should not treat historical calorie data as directly comparable to new platform readings.