HN Threads 2
Microsoft AI's leader argues that teaching models person-like interests complicates containment; this is a philosophical/training-policy argument, not an experimental proof about consciousness.
Enclave reports 11/11 vulnerable targets compromised and four fixed controls intact, at $4.65 for accepted runs ($5.14 including retries); five successes used unintended benchmark routes, since repaired.
Tools & Demos 2
OpenSpec tracks proposals, specs, design and tasks through explore/propose/apply/verify/archive stages, keeping coding-agent implementation tied to evolving requirements.
A new merge-gate tool proposes scoring structural decay in AI-assisted changes; commenters question metric scope and whether the project itself meets its quality claims.
Podcasts 1
AIUC cofounder Rune Kvist argues enterprise adoption needs independent testing, quarterly standards and insurance—not merely stronger models; the Sep 16 interview also announces a $40M round.
TLDR: Rune Kvist argues that trust and risk can constrain adoption even when models are capable. Quarterly adversarial tests and insurer-backed promises may make risk legible, but certification cannot cover every daily software change. The practical takeaway is to request concrete audit evidence and understand limits rather than rely on badges. Chapters: 00:07:58 trust versus capability; 00:18:52 controls and adversarial audits; 00:41:45 coverage, limits and insurer relationships; 01:19:10 watchdog incentives and the roadmap. Quotes: Rune Kvist, “standards have to precede insurance” [00:49:12]; “you must get tested every quarter” [00:15:08]. These are a vendor's proposals, not legal/insurance advice.
HF Trending Papers 5
A benchmark mines 1,975 replay-verified behaviours across 26 apps and builds 4,063 tasks; the authors report 49.2% vs 28.8% full-app workflow success for Astra and Opus 5 (11 HF upvotes).
GAI distinguishes who performs improvement and whether evaluation is externally anchored, separating bounded policy improvement from goal drift and self-referential loops (19 HF upvotes).
ModularRSI evolves five restricted harness modules using 2,000 tasks disjoint from downstream evaluation, reporting transfer to unseen tasks and models (8 HF upvotes).
StepAudio 3 Realtime runs private reasoning alongside spoken output and asynchronous tools; authors report 56% τ-Voice task success and 98.9 on Full-Duplex Bench (94 HF upvotes).
Small continuous register states carry progress after earlier generated chunks are cleared; authors report gains up to 8.5 points on math and 19.5 on code versus text carry (3 HF upvotes).
arxiv: Agent Systems 1
Controlled recommender experiments find that assigning an agent to a booking platform rather than a traveler changes how it treats sponsored listings; this is role-prompt sensitivity, not an audit of OpenAI's new ad product.