What changed this week
AI platform signals
Short reads on Siri AI, App Intents, local models, cloud agents, memory systems, and privacy regulation.
Siri AI at WWDC26: generic replies or your trained voice?
Apple unveiled Siri AI at WWDC26: a rebuilt conversational Siri with reply suggestions in Messages, shipping with iOS 27. What it means for your own voice.
Read this firstGuides in this hub
01 Small Models Caught GPT-4o — and a $20B Market Followed Microsoft's 3.8B Phi-4 matched GPT-4o on structured tasks in 8GB RAM; the small-model market is forecast to reach $20.7B by 2030. 2 min 02 Phones Hit 45 TOPS: On-Device AI Crosses the Hardware Line Snapdragon's NPU hit 45 TOPS and edge benchmarks show local models beating cloud latency. The hardware excuse for cloud-only AI is gone. 2 min 03 Prompt Injection Surged 340% in 2026's OWASP Report OWASP says prompt injection is now the fastest-growing AI attack, seen in 73% of deployments. The risk scales with how much an agent can do unsupervised. 2 min 04 Siri AI at WWDC26: generic replies or your trained voice? Apple unveiled Siri AI at WWDC26: a rebuilt conversational Siri with reply suggestions in Messages, shipping with iOS 27. What it means for your own voice. 2 min 05 App Intents replaces SiriKit for the new Siri Developer press reports Apple made App Intents the mandatory surface for the new Siri, deprecating SiriKit and adding schemas for Spotlight's semantic index. 2 min 06 iOS 27 may let chatbots plug into Siri. What moves with you Press coverage says iOS 27 will let users plug chatbots like Gemini or Claude into Siri via Extensions. Swappable assistants, and what does not move with them. 2 min 07 Reports say a Gemini model powers the new Siri. Our read Per press reports, the rebuilt Siri runs partly on a custom Google Gemini model inside Private Cloud Compute. What that means for where your text travels. 2 min 08 Apple opens up the Foundation Models framework At WWDC26, Apple added image input, a free Private Cloud Compute model with 32K context, and a protocol that swaps third-party models behind one Swift API. 2 min 09 Apple introduces Core AI for custom on-device models Per MacRumors, WWDC26 brought Core AI: a framework for running custom models on Apple silicon, with ahead-of-time compilation and PyTorch conversion tooling. 2 min 10 Xcode 27 ships an on-device coding agent Per TechTimes, Xcode 27 pairs a local Neural Engine model that completes code without sending source anywhere with optional cloud routing for heavier work. 2 min 11 Build 2026: Windows Bets on Local AI Agents Microsoft centered Build 2026 on local AI agents, OS-enforced sandboxes, and 120B-class desktop hardware. The on-device shift now spans every platform. 2 min 12 Gemma 4 12B Unified: Open Weights Go Multimodal Google's Gemma 4 12B Unified puts multimodal open weights on laptops. What the release means for private AI, and for on-device writing on iPhone. 2 min 13 Google I/O 2026: small models now rival last year's flagship At I/O 2026, Google announced Gemini 3.5 Flash and Gemini Omni. The real signal for personal writing: small, fast models now rival last year's flagships. 2 min 14 Gemini Spark: Google's always-on agent lives in the cloud TechCrunch reports Gemini Spark works around the clock on Google Cloud VMs and takes tasks via Gmail. It is the opposite of the on-device privacy trade. 2 min 15 Perplexity's Hybrid Local-Cloud Answer to Private AI Perplexity demoed hybrid local-cloud inference at Computex 2026, per VentureBeat. How consent-based routing compares with fully on-device writing. 2 min 16 GPT-5.5 Instant makes personalization ChatGPT's default OpenAI released GPT-5.5 Instant as ChatGPT's new default, pitched as smarter, clearer, and more personalized. Per-user voice is now the industry baseline. 2 min 17 OpenAI's Dreaming V3 rebuilds ChatGPT memory OpenAI's Dreaming V3 rebuilds ChatGPT memory: an editable summary, self-updating memories, and a 5x serving-compute cut. The profile still lives server-side. 2 min 18 Grok Custom Skills: The Playbook Idea Spreads xAI shipped Custom Skills for Grok, per InfoQ — reusable instruction bundles that persist across chats. The personal playbook concept goes cloud-side. 2 min 19 What Claude Fable 5 means for everyday on-device writing Anthropic has released Claude Fable 5, with Claude Mythos 5 in limited release. Why frontier cloud models change little for personal, on-device writing. 2 min 20 Claude Opus 4.8 and the Case for 'I Don't Know' Anthropic's Claude Opus 4.8 cuts hallucinations mainly by abstaining when unsure. Why that habit matters more for everyday email than benchmark scores. 2 min 21 Code with Claude 2026: agents that learn from your feedback InfoQ reports Anthropic's Code with Claude 2026 event: multi-agent orchestration, Outcomes success criteria, and a self-improving Dreaming preview. 2 min 22 What the EU's AI labelling code means for your writing The European Commission has published its final Code of Practice on labelling AI-generated content. What it asks for, and what it means for everyday writers. 2 min 23 EU delays AI Act high-risk rules; transparency stays on time EU negotiators agreed to postpone the AI Act's high-risk obligations to late 2027. Transparency duties for AI content still arrive on August 2, 2026. 2 min
Use this hub as training material
AI platform signals can become private agent behavior.
Read the first guide, apply it to one real message, correct what feels wrong, then save the preference so the agent can reuse the behavior next time.
Free to start · No account · Data Not Collected