645: More Frosting

June 26, 2025

102 min episode · 2 min read

Episode

102 min

Read time

2 min

AI-Generated Summary

Published Dec 25, 2025

Key Takeaways

✓macOS Tahoe Beta Evolution: Beta 2 reverses Finder icon colors with blue now on left side, adds menu bar background toggle in settings (off by default with tinted not opaque background), and increases frosting opacity in iOS toolbars making previously unreadable Safari and Music controls more legible though still problematic.
✓Icon Customization Tools: Pictogram and Replacion applications enable persistent custom icon replacement on Mac that survives app updates and auto-updates, automatically reapplying custom icons fast enough that relaunched apps display them immediately without manual intervention, solving the recurring problem of lost customizations after software updates.
✓AI Training Data Consent: Apple's Feedback Assistant now requires agreement that submitted bug reports including system diagnostics may train Apple Intelligence models with no opt-out option, forcing developers to choose between filing bugs or preventing their proprietary code and comments from appearing in future autocomplete suggestions to other developers.
✓Meta Copyright Defense Strategy: Meta argues in court that individual books have zero economic value as training data despite their AI requiring millions of books collectively worth billions, claiming paying authors would be impossible while simultaneously asserting fair use because single books change model performance by only 0.06 percent on benchmarks.
✓BMW CarKey Reliability: Apple CarKey on BMW iX works reliably approximately 97 percent of the time over one year of daily use, with occasional 2-3 second Bluetooth delays and rare CarPlay connection issues in first 30 seconds of driving, performing significantly better than earlier Rivian Bluetooth implementations and eliminating oversized key fobs.

What It Covers

Accidental Tech Podcast episode 645 examines macOS Tahoe beta updates including menu bar transparency changes, Safari toolbar improvements, iOS 18 liquid glass design iterations, and debates around AI training data ethics involving Meta, New York Times licensing deals, and copyright lawsuits.

Key Questions Answered

•macOS Tahoe Beta Evolution: Beta 2 reverses Finder icon colors with blue now on left side, adds menu bar background toggle in settings (off by default with tinted not opaque background), and increases frosting opacity in iOS toolbars making previously unreadable Safari and Music controls more legible though still problematic.
•Icon Customization Tools: Pictogram and Replacion applications enable persistent custom icon replacement on Mac that survives app updates and auto-updates, automatically reapplying custom icons fast enough that relaunched apps display them immediately without manual intervention, solving the recurring problem of lost customizations after software updates.
•AI Training Data Consent: Apple's Feedback Assistant now requires agreement that submitted bug reports including system diagnostics may train Apple Intelligence models with no opt-out option, forcing developers to choose between filing bugs or preventing their proprietary code and comments from appearing in future autocomplete suggestions to other developers.
•Meta Copyright Defense Strategy: Meta argues in court that individual books have zero economic value as training data despite their AI requiring millions of books collectively worth billions, claiming paying authors would be impossible while simultaneously asserting fair use because single books change model performance by only 0.06 percent on benchmarks.
•BMW CarKey Reliability: Apple CarKey on BMW iX works reliably approximately 97 percent of the time over one year of daily use, with occasional 2-3 second Bluetooth delays and rare CarPlay connection issues in first 30 seconds of driving, performing significantly better than earlier Rivian Bluetooth implementations and eliminating oversized key fobs.

Notable Moment

A federal judge ruled that Anthropic can legally train AI models on physically purchased and scanned books under fair use, but explicitly did not address whether charging users to access those trained models constitutes copyright infringement, leaving the commercial exploitation question unresolved and setting up future litigation battles.

Know someone who'd find this useful?