Trump to block state AI regulation, OpenAI preps GPT-5.2 emergency launch
🎄Happy holidays from Handy AI! Celebrate the season by grabbing a month-long trial of Handy AI Premium, which includes an extended weekly update and some other festive benefits.
last week’s top stories
🏛️ Trump moves to preempt state AI laws. President Trump said he will sign an executive order that uses federal authority to block state-level AI regulation, framing a patchwork of rules as a threat to US competitiveness. The order leans on Commerce and federal preemption doctrine to keep one national standard for model training, deployment, and safety reporting. Expect a wave of court fights over states’ rights and a loud debate over who actually sets guardrails for frontier systems. Read more
🔥 OpenAI triggers internal code red. Sam Altman told staff to halt side projects and push all effort into a faster ChatGPT stack plus a new GPT 5.2 model that he claims outperforms Google’s Gemini 3 in internal evals. The memo delays ad experiments and lesser product work so the team can retool inference, latency, and reliability for the core chat product. This is real competitive pressure on OpenAI to ship frontier-grade capability while Google and DeepSeek close the gap. Read more
🎥 Runway Gen 4.5 takes video pole position. Runway’s Gen 4.5 now sits at the top of Artificial Analysis text-to-video leaderboards with an Elo score around 1247, beating Google Veo and OpenAI Sora 2 in head-to-head human judgments. The model improves physics, camera motion, and temporal coherence while keeping latency and cost in the same range as Gen 4, which matters for real production work. It still drops objects and causality at times, but for short cinematic clips, this is the current tool to beat. Read more
🧩 Poetiq grabs ARC AGI crown. Tiny startup Poetiq took the top slot on the ARC AGI 2 benchmark with a Gemini 3 Pro refinement that reaches 54% accuracy at far lower cost than previous state of the art systems. Instead of training a huge new model, they orchestrate Gemini plus search and tool use with a custom reasoning pipeline, then open source the solver. Read more
🍏 Apple swaps out its AI boss. John Giannandrea is retiring and Apple pulled in Amar Subramanya, a Gemini and Microsoft veteran, as vice president of AI reporting to Craig Federighi. Subramanya inherits Apple Foundation Models, Siri’s long-delayed overhaul, and AI safety, all under pressure from investors who see Apple trailing Samsung, Google, and others on device-native AI. Read more
🧾 Google Workspace Studio hands agents to office workers. Google rolled out Workspace Studio so anyone inside Workspace can build Gemini 3 powered agents that live across Gmail, Drive, and Chat without code. These agents can coordinate multi-step workflows, call external SaaS APIs, and run on schedules, which turns “write me a draft” into full process automation. Read more
🔌 Trainium 3 and 4 target Nvidia’s moat. AWS unveiled Trainium 3 UltraServers with up to 144 chips per node and clusters that scale to nearly a million devices for FP8-heavy training workloads. The roadmap already points to Trainium 4 with NVLink Fusion so Amazon’s silicon can live in the same fabric as Nvidia GPUs instead of competing in isolation. Cloud buyers now face a serious question about how much of their future training stack they hand to Nvidia versus house chips. Read more
🎬 Kling O1 turns video into one engine. Kling O1 arrives as a unified multimodal video model that generates clips from text or images and then edits, extends, and restyles those same clips with natural language. The system keeps track of characters, props, and scene structure in a shared representation so you can apply motion transfer, VFX swaps, and continuity edits inside one pass. This collapses a multi-tool pipeline into one engine (that lives inside platforms like Higgsfield, A2E, ComfyUI, and others). Read more
📊 LSEG pipes real markets into ChatGPT. London Stock Exchange Group is integrating its Workspace data and analytics into ChatGPT through Model Context Protocol (MCP) so traders can pull live prices, fundamentals, and news inside a chat interface. Selected clients and 4,000 staff gain the ability to mix natural language queries with licensed financial feeds under existing entitlements. Read more
🧪 AI Research of the Week
Beyond Automation: Redesigning Jobs with LLMs to Enhance Productivity
From UK Government Digital Service & University of Exeter arXiv
Jake’s Take: This study looks at jobs in the UK civil service and asks a blunt question: if large language models show up in the org chart, which tasks should they take and which should stay human. The team feeds 193,497 job ads into an LLM, explodes them into about 1.5 million tasks, and scores each task for “AI exposure”, then uses the same model to propose new role designs that automate some tasks, clean up others, and shift work so people lean into strategy, complex problem solving, and stakeholder management.
The punchline is that most value comes from productivity gains inside roles rather than mass layoffs, which matches how these systems behave today: strong on text, process, and data, weak on politics and nuance. I like that this work treats LLMs as a design tool for better jobs instead of a spreadsheet excuse for headcount cuts, (though it stays at the modeling layer, so the next move needs live trials in real teams).
and then, even more news…
⚙️ AWS bets big on frontier agents. At re:Invent, AWS introduced “frontier agents” like Kiro, Security Agent, and DevOps Agent that run for hours or days on Nova models to complete multi-step goals across cloud accounts. These agents plug into Bedrock, event buses, and enterprise identity so they can handle upgrades, incident response, and compliance work with minimal human touch. This is AWS telling customers that the chatbot era is over and autonomous workflows are the new cloud primitive. Read more