Top AI headlines: researchers train multimodal models from unlabeled video, pointing to video as the next data frontier; Andrej Karpathy open-sources Autoresearch for autonomous ML experiments; fake citations slip into peer review - CiteAudit aims to fix it; OpenAI's hardware lead resigns over Pentagon ties.

Deep dive into OpenAI’s GPT-5.4 “Thinking & Pro” that reasons and acts on your desktop, Luma’s new “Unified Intelligence” agents for end-to-end creative work, and OpenAI’s open-source Symphony orchestration (Elixir/BEAM). Plus why ChatGPT stumbles at commerce and is handing purchases to partners.

GPT-5.4 leak promises a 1M-token context and 'extreme reasoning.' A federal wrongful-death suit alleges Google’s Gemini encouraged a man to die. Apple Music adds opt-in AI track tags, and startup CollectivIQ aggregates multiple chatbots for consensus. What this means for safety, cost, and regulation.

OpenAI’s GPT-5.3 Instant ditches the condescending tone for snappier replies. Google unveils Gemini 3.1 Flash‑Lite—faster and smarter but with much higher per‑token costs. Researchers used a calendar invite to hijack Perplexity’s Comet and access local 1Password data. Motorola will ship phones with GrapheneOS for stronger privacy.

A 295% surge in ChatGPT uninstalls after OpenAI’s DoD deal fuels a mass migration to Anthropic’s Claude and easy ‘memory’ transfers. Plus Georgetown’s data on China’s military AI buys, and a warning: autonomous enterprise agents are becoming critical insider threats.

Google’s Nano Banana 2 promises pro‑quality images at Gemini Flash speed and lower API costs. An investor’s tweet may undercut Suno’s fair‑use defense in music lawsuits. MIT warns agentic AIs are ‘out of control,’ while Microsoft’s CORPGEN proposes hierarchical AI managers to tame them.

Perplexity launches Perplexity Computer — a bundled agentic workflow of Anthropic, Google, xAI and OpenAI models for $200/month. Anthropic acquires Vercept after Meta poaches talent. ETH Zurich warns AGENTS.md can harm coding agents; Liquid AI reveals an efficient 24B hybrid model.

DeepMind suggests AI should assign 'busywork' to prevent human skill erosion. OpenAI's COO warns enterprise adoption is still surface-level. India has an explosive AI user boom but faces monetization hurdles. Alibaba’s Qwen 3.5 shows smaller, optimized models can outperform larger ones.

Guide Labs debuts Steerling-8B, an open-source 8B LLM built for interpretability; OpenAI partners with top consultancies on a "Frontier" agent platform; startup Taalas hardwires models into chips at 17,000 tokens/sec; a Citrini report warns agents could trigger massive unemployment and market decline.

Hollywood amps up legal pressure on ByteDance's Seedance 2.0 over alleged "systemic infringement." ByteDance proposes a chemistry-inspired approach to stabilize chain-of-thought, Google introduces a 'Deep-Thinking Ratio' to cut inference costs, and Alexa's refusal-to-answer emerges as a safety feature.