Erik Hornberger

SwiftAI Frameworks Engineer at Apple

[Shun Takeishi @ShunTakeishi]

9/11だけではありますが、Foundation Model フレームワークのエンジニアも参加してくれることになりました。もし直接エンジニアと話したいって方は9/11にぜひご応募ください。また、9/9,10の枠も、もちろんお待ちしております。

[Erik Hornberger @mopperson]

If you’re passionate about the intersection of AI and API design, there is arguably no better team than this one. Opportunities to be part of something this big are few and far between.

[Richard Wei @rxwei]

The team that built the Foundation Models framework is hiring! Join us to create the best LLM APIs, both for app developers and for Apple teams.
https://t.co/OWZuGOUjaO

[Erik Hornberger @mopperson]

The update also contains optimizations to the inference stack that improve token throughput. I'm curious to see how noticeable the difference is in your use cases. (6/6)

[Erik Hornberger @mopperson]

iOS 26 Beta 5 dropped today and contains some exciting additions to the Foundation Models framework. (1/6)

[Erik Hornberger @mopperson]

GeneratedContent is no longer opaque. It now has a kind property that you can inspect. This makes it possible to create views that display arbitrary generated content, which can be useful when working with user configurable generation schemas. (5/6)
https://t.co/lsmkZddfof

[Erik Hornberger @mopperson]

@bdkjones @_julianschiavo Yes, it put it in a Task and cancel the task!

[Erik Hornberger @mopperson]

Very clever idea, I love to see stuff like this! You can also add a slider to your preview view to scrub back and forth across different degrees of partially generated states.

https://t.co/UXnpDGguvE by @iosartem

[Erik Hornberger @mopperson]

@winstondu Thanks for the report!

[Erik Hornberger @mopperson]

@rudrank Thanks for filing feedback!

[Erik Hornberger @mopperson]

@bilqisium @NatashaTheRobot Thank you!

[Erik Hornberger @mopperson]

@bilqisium @NatashaTheRobot Please file feedback for that. If you have specific prompts that you can include for us to repro, that would be very helpful.

Seed 3 should have a new session.dumpFeedbackAttachment() that you can use to capture the transcript in a json file that you can attach!

[Erik Hornberger @mopperson]

@NatashaTheRobot Great write up! First I’ve seen the recommendation to put reasoning before classifications and use structured concurrency. 👍👍

A small optimization is to use an enum for the classification. After the first token, the label is deterministic, so we can skip computing the suffix!

[Erik Hornberger @mopperson]

@irangareddy @azamsharp @generable There is a type named DynamicGenerationSchema that exists specifically for expressing schemas that are constructed at runtime instead of compile time.

It’s a little more typing, but can be very powerful.

[Erik Hornberger @mopperson]

@gyin94 @normarayr Also, the entires Transcript type is Codable, which is an even easier way to persist and restore sessions.

[Erik Hornberger @mopperson]

@gyin94 @normarayr It’s a known bug — we listed it in the release notes.