Siri AI beta and what we think so far

Siri AI / Apple Intelligence: overall reaction

The hosts discuss their early experiences with Siri AI (developer beta / Apple Intelligence), noting that it is already more capable than older Siri—especially as a personal assistant that can reason over a user’s own data (texts, contacts, calendar, notes, receipts, and more).

They also highlight ongoing LLM limitations and beta-era weirdness, including failures, “thinking,” wrong answers, and access delays.

Overall: impressive capabilities, but still unstable and sometimes unreliable during the beta phase.

Key capability themes (what’s working well)

Personalization via on-device indexing

A central point is that Siri AI relies on Apple’s revamped content indexing. In the beta, indexing can take unusually long during initial enablement (reportedly several days). The indexing builds the “relational map” Siri needs to access user data quickly and thoroughly.

Contextual understanding of your life

The hosts share examples where Siri can infer plans and details from ongoing conversations—even when the user doesn’t provide every field explicitly. For example, it can assemble birthday-related plans from prior message threads.

Task conversion and rewriting

Siri can take unstructured personal notes and reformat them into structured outputs—such as turning notes about bread baking into a recipe format.

Visual intelligence improvements

They describe improved parsing for checks/receipts and images. For example, pointing Siri at a receipt and asking it to remove categories and split totals can work accurately.

Multistep automation potential

They expect Siri AI to become more valuable as third-party integrations mature—enabling searches within apps, gathering context, and taking actions beyond simple Q&A.

Image generation is better (still imperfect)

The image generator is described as dramatically improved compared with earlier attempts, though it may still look somewhat “AI.”

Limitations / “weird” or disappointing behavior

LLM-style factual/logic traps

The discussion covers why Siri (and LLMs generally) can fail on tasks involving tokenization rather than letter-by-letter counting, leading to inconsistent results on puzzle-like questions such as:

• “How many Rs are in strawberry”
• Other letter-based word problems

Confident yet wrong answers

Siri may correct itself when prompted (“Are you sure?”), but it can still behave unreliably.

App/web source ambiguity

In a weather-related test, Siri provided answers that didn’t align with the Weather app, and it wasn’t clear where the data came from—suggesting it may sometimes use a different source.

Beta-stage instability

There were access delays (waitlist-like behavior), prolonged indexing, and occasional issues where Siri fails to respond correctly.

Access timing and beta logistics

• Siri AI access is gated, with waitlist behavior similar to earlier Apple Intelligence rollouts.
• Indexing takes much longer than usual in the beta, leading users to wonder if something is broken (with reports of days-long indexing).
• For watchOS, Siri AI availability is not initially included in the earliest beta stages, even though other devices get it.
  • Apple indicates availability later in the year / later beta.
  • It may arrive in a later watchOS update rather than version 27.0.
• Apple’s messaging suggests Siri AI may require:
  • supported languages, and
  • Apple-Intelligence-capable devices
  • with initial rollout starting in the US and English first.

Third-party app integration: expectations and uncertainty

Apple isn’t providing full “deep indexing” access to all third-party apps immediately. The hosts are unsure whether Siri AI will:

• only perform actions via APIs, or
• also deeply index and reason over app data.

They reference Apple developer building blocks:

• App Intents (actions Siri can perform)
• App Entities (data Siri can access inside an app)

They hope for stronger integrations (e.g., Slack search/history, other note apps, calendar/workflows), but note that app developer updates are likely required.

Privacy stance: important distinction

The hosts argue that privacy likely limits training data collection, but doesn’t necessarily reduce Siri AI’s day-to-day usefulness.

They frame privacy as a reason Apple didn’t use the “train on everything” approach taken by some other AI ecosystems—keeping the system on-device/encrypted and designed to protect user data.

“Apple history” segment: end of Intel Mac support context

During the WWDC keynote on June 22, 2020, Apple announced a two-year transition from Intel to Apple silicon.

The show notes that the transition ultimately completed on schedule. They emphasize that this is the first year where new macOS updates no longer support Intel chips (with macOS 27 described as Apple-silicon-only). However, older Intel Macs should still get security/bugfix updates for at least a year.

They also credit the scale of the architecture shift and the software ecosystem effort required for such a migration.

“Comment Corner” highlights

Listener questions focused on:

• how much Siri AI can use third-party apps and data
• whether privacy constraints will reduce capabilities
• app integration mechanisms (app entities / intents)
• why Apple apps appear to be gaining landscape orientation again (possibly tied to future devices like folding hardware or new screen form factors)

One listener recommended an Apple Watch model for Roman (SE3). Another discussed landscape UI changes possibly connected to upcoming form factors.

Presenters / contributors

• Michael Simon (host)
• Jason Cross (host)
• Roman Lyola (producer / contributor)