Reddit user tests Neural Reflex Model for robot blinking

A privacy policy in the iOS 27 release candidate lets Apple store Siri and dictation interactions, including audio recordings and transcripts, for training in-house language models. Apple had previously ruled out using personal data for training; the change is opt-in at Siri AI activation, with data not linked to the Apple Account.

A Reddit post claims an unreleased Astra-family model wrote a persona line for itself during RL training; the post carries no snippet, link, or corroborating detail.

Microsoft CEO Satya Nadella, speaking on the All-In Podcast, said AI can fabricate financial records. The episode was published September 16, 2026.

A LocalLLaMA user compared the new Jev model from TypeSafe's System One release — which outputs probabilities given choices — with Qwen reranker models, then implemented the approach using Qwen3.5 4B logits.

Trajectory-Shaped Discrete Flow Matching uses a lightweight energy compass to guide training trajectories, letting a 170M-parameter student hit 32% lower perplexity than its 1,024-step teacher at 8 steps — 128× faster. Shaping is training-only, so inference cost is unchanged.

DACA-GRPO adds Denoising Progress Scores and Stratified Masking Likelihood to any GRPO-style trainer, targeting missing temporal credit assignment and mean-field likelihood bias. Gains reach 36.3pp on constraint satisfaction, 7.4pp on code generation, and 5.6pp on math reasoning across seven benchmarks.