Skip to content

ai-summaries.md

Latest commit

 

History

History
68 lines (40 loc) · 3.03 KB

ai-summaries.md

File metadata and controls

68 lines (40 loc) · 3.03 KB
Raw

AI summaries

One-sentence summaries of talk descriptions generated on-device by Apple Intelligence. Optional, gated behind a Settings toggle.

Requirements

  • iOS 26 or later (the FoundationModels framework first ships in iOS 26).
  • A device that supports Apple Intelligence (iPhone 15 Pro / Pro Max, iPhone 16 family, M-series iPad Pro / Air / mini, all later devices).
  • The user has Apple Intelligence enabled in iOS Settings → Apple Intelligence & Siri.

On devices that don't meet the requirements, the Settings toggle is hidden entirely.

Turning it on

Settings → AI Summaries → toggle on.

AI Summaries toggle in Settings

The toggle has a footer explaining the feature:

Show one-sentence summaries of talk descriptions, generated on-device by Apple Intelligence. Summaries are cached and only generated for descriptions longer than 100 characters.

What you see

When enabled, each talk row gets a new line below the speaker names:

✨  AI-generated one-sentence summary, lineLimit(2).

The summary appears in .foregroundStyle(.secondary) so it reads as ambient metadata rather than primary content.

Row with AI summary line

Long-press peek

Sometimes the summary skips a detail you care about. Long-press a row with a visible summary → opens a sheet showing the original full description with selectable text. Tap Done to dismiss.

The peek sheet only triggers on long-press for rows that have a visible summary. Long-pressing other rows behaves normally (no-op for the navigation tap).

How it works

Generation is triggered opportunistically when a row materializes in the LazyVStack — i.e., when you scroll close enough that the row is about to be visible.

Each row's generation runs in a background Task, deduplicated per row id. The cache is a global TalkSummaryCache actor that:

  • Stores summaries in UserDefaults as JSON, keyed by content id with a SHA256 of the source description.
  • Returns cached summaries instantly on subsequent visits.
  • Re-generates when the source description's hash changes (the talk got edited server-side).
  • Caps at 1,000 entries; evicts oldest first.
  • Throttles to 4 concurrent generations to keep battery cost reasonable.

The prompt is biased toward "what will the audience learn" instead of marketing copy, with a cleanup pass that strips prefatory text ("Summary:", surrounding quotes) some model outputs still emit.

What stays the same

  • The original talk description is never modified — the summary lives alongside it.
  • AI summaries are never sent to a server — generation is entirely on-device via Apple's bundled model.
  • The 100-character minimum on source descriptions means short blurbs (e.g. "Hands-on workshop") don't waste a model call.

Disabling

Settings → AI Summaries off — rows hide the sparkle line. Cached summaries persist (no point throwing them away) but new ones won't generate.

See also