ChatGPT now integrates DoorDash, Spotify, Canva and more, turning into a one‑stop app hub. Inside OpenAI, safety researchers are quitting amid Altman's "vibes" defense. Studies show flattering chatbots can warp reasoning, while millions use ChatGPT for after‑hours medical triage.

Offensive AI cyber capabilities are exploding, raising new security risks—while autonomous coding agents like “Devin” promise to reshape software teams. Plus: rumors of Google Gemini on iPhone and the EU AI Act’s sweeping rules and fines. What changes next?

A look at Sakana’s Ultra Deep Research (Marlin) — an autonomous eight-hour strategy AI; Microsoft’s MAI group releasing three multimodal foundational models including fast MAI-Transcribe-1; Gemini’s tool to import chat memories; and an MIT report forecasting gradual AI impact by 2029.

OpenAI closes a $122B round and unveils a ChatGPT Super App while DeepMind warns autonomous agents are being hijacked by web "traps." Plus: gig workers training humanoid robots via phone headcams and Google's Veo 3.1 Lite cuts generative video costs.

Alibaba’s Qwen 3.5 Omni shows native multimodal zero-shot coding from speech + video. A supply-chain breach tied to LiteLLM led to Mercor’s data extortion. Google’s Veo 3.1 Lite slashes video-gen costs, and Runway backs real-time video startups with $10M.

OpenAI abruptly shutters Sora, sparking privacy and data-harvest fears; image models confidently hallucinate medical diagnoses; Salesforce’s VoiceAgentRAG slashes voice latency by 316×; and AI quietly fuels hit-making in the music industry. A concise look at risks, breakthroughs, and fallout.

AI video hype stumbles as Sora faces shutdown and astronomical compute costs. Viral claims of an AI-designed dog vaccine lack clinical proof. Researchers reveal MetaClaw (training agents during meetings) and A-Evolve, a potential PyTorch-style leap for autonomous AI — with privacy and hype risks.

This week in AI: Apple licenses Google's Gemini to distill on-device Siri models; Google launches Gemini 3.1 Flash Live for real-time multimodal voice/video and tool use; Mistral unveils open-weight Voxtral TTS that clones voices from 3 seconds; OpenAI scraps an erotic ChatGPT project.

OpenAI scraps Sora and loses a billion-dollar Disney deal. Google unveils Lyria 3 Pro — copyright-safe AI music for commercial use. Nvidia's PivotRL slashes agent training compute. Plus Melania Trump's pitch for robot homeschool tutors and the real-world implications.

OpenAI pulled Sora despite Sora 2’s impressive video model; Tasklet’s agentic no-code tool claims to build and deploy apps in minutes; a LiteLLM supply‑chain malware is stealing credentials and spreading through Kubernetes clusters; TinyLoRA proves 13-parameter fine-tuning unlocks strong reasoning.