Alex Heath: Siri still lacks persistent memory, Apple-OpenAI partnership quietly dying, Google is the new default AI partner

Apple's AI Moment — and Its Limits

Alex Heath attended Apple's post-WWDC tech talk in Cupertino and came away with a clear verdict: the new Siri is a meaningful step forward, but it's roughly what Apple should have shipped two years ago. The demo reel Apple ran in that earlier cycle — showing Siri doing agentic tasks on device — never delivered. The rebuilt stack does now, at least in early testing.

The most striking gap is persistent memory. Siri still forgets everything between queries. For anyone running Claude or ChatGPT daily, that alone will feel like a generation behind.

OpenAI relationship quietly fading

The ChatGPT integration has been significantly downgraded. Users now have to manually prompt each time to hand off to ChatGPT, and the software UI element has been removed entirely. Heath draws a direct parallel to the Facebook-iOS relationship under Steve Jobs — close collaboration, then ideological drift, then a quiet parting. Craig Federighi's on-stage dig about not running ads was a tell. The OpenAI partnership, Heath argues, is heading in the same direction.

“I think they finally have a Siri that works. It's arguably what they should have shipped two years ago... The other the other big thing about this new Siri is they have really nerfed the ChatGPT partnership... I really think there's this OpenAI is kind of like the new Facebook to Apple where they work together closely... and then they just kind of drifted apart.”

Google steps in — but not as a product

Google is now Apple's primary AI partner, with Gemini supplying the underlying model. But Apple is doing extensive post-training and reinforcement learning to make responses distinctly "Apple" rather than standard Gemini outputs. Federighi ran what Heath describes as an "AI 101" whiteboard session with press specifically to make that point — Apple controls the stack, Google supplies the base.

The strategic framing Heath finds most credible comes from Dave Moran (quoted on the record): Siri is the harness, not the product. Apple has built flexible infrastructure underneath that's designed to accept future models. Google is the best available partner today, but the architecture is meant to outlast any single partnership. Heath expects Apple to eventually want to own more of its AI destiny, citing Meta's lab reboot as a comparable multi-year effort.

Inference costs are the real constraint

The new Siri launches in English only, with no EU or China availability, on an invite waitlist. During the demo at Apple Park on launch day, photo retouching tasks were taking around 25 seconds and servers were visibly strained. Apple has been telling a story about being an AI winner without major CapEx, but cloud inference through Google Cloud Platform will show up on the balance sheet. Heath expects a slow, deliberate ramp rather than a broad rollout.

Developer opt-in is the next bottleneck

Siri can see on-screen text and images by default, but understanding an app's UI requires the developer to explicitly opt in — which means handing Apple access to that data. Whether major platforms like Meta or Anthropic will open their apps to Siri is unresolved. Heath flags this as the next thing to watch: user demand may pressure developers to participate, but the data tradeoff is real.

Acquisition unlikely to reset the strategy

Heath had expected Apple to have made a significant AI acquisition by now. They looked at Perplexity and Thinking Machines, but appear to have moved past both. Apple has since built its own web knowledge graph, which was likely what a Perplexity deal would have unlocked. Heath doesn't see a strategic gap that a large acquisition now fills — the moment for a foundational Oculus-style bet has likely passed. Future deals, if any, will be talent-driven.

The question Apple can't fully answer yet is how open the harness will become. Heath thinks there will be some degree of model switching, but doubts it will be as frictionless as swapping a search default. Models are too integrated with the OS layer, and the product logic is pulling in the opposite direction from plug-and-play.