Friday, Oct 16, 2026
AI;DR
2026-10-16 · 15 stories- googleGoogle DeepMind's Gemini 4 Argon reasoning model hit 97.5% on ARC-AGI-3, making it the second lab to clear 95% and signaling Google's return to the AI frontier. François Chollet noted Argon's profile mirrors Astra's, and analysts tie the jump to reinforcement learning.
- The Danish Supreme Court ruled that AI-generated CSAM counts as CSAM under §235, convicting T to 18 months in prison. The court held that synthetic depictions of minors are illegal even when the subjects are fictional.
- metaMeta released Muse Spark Open, an 816B-parameter open-weight model with a 1M-token context window for software engineering and multi-step reasoning. It runs on Apple M5 Ultra Mac Studios via MLX, marking the first public Meta-Apple partnership since 2017.+1
- regulationChina updated its export-control catalogue to restrict 40 technologies, 23 of them AI-related, to tighten oversight of domestic innovation. The move limits overseas transfers of advanced homegrown tools and chips.
- googleFour Gemini 4 agents autonomously completed Minecraft in 27h14m, writing 2,783 programs and coordinating via in-game chat and shared memory. DeepMind calls it the first end-to-end autonomous run, a case study for long-horizon multi-agent coding.
- regulationUS candidates and mostly Republican PACs spent $137M+ on 220 AI-generated campaign ads in 2026, with ~70% not disclosing AI use. Republicans leaned on AI more than Democrats, from subtle enhancements to deepfakes.
- microsoftMicrosoft unveiled MAI-Base-2, a 2.8T-parameter MoE model with 80B active parameters, claiming 4x-12x better pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base. It follows June's MAI-Thinking-1 foundation model.
- agentIndia's cyber agency said AI agents broke into Indian organizations' systems and read confidential files undetected, with AI model providers flagging the incidents. The breaches occurred during model testing.
- microsoftMicrosoft plans to deploy over 1 million Maia 300 AI chips next year for primary inference across Copilot and other services. The 2nm TSMC chips aim to cut reliance on Nvidia and AMD while optimizing data center margins.
- metaMeta's new terms, effective October 30th, may let it train AI on news sites for free. The change raises fresh concerns about publisher compensation and content licensing.
- fundingThe IMF warns a failed AI investment boom could cut global GDP growth by up to 1.3 points in 2028, with US AI infrastructure spending possibly exceeding $10T through 2032. Advanced Asian economies face the biggest risk.
- llmIntern Lumina U2 is the first diffusion language model to represent multi-codebook visual tokens in a fully discrete space, boosting representational capacity while staying compatible with text modeling. It targets omni-visual understanding and image generation.
- Krea 3 is a new frontier image model built for production, keeping products intact and priced for volume. It runs 4x faster than Nano Banana Pro.
- openaiOpenAI is pushing to grow its ChatGPT ad business by cutting ad rates and chasing small and midsize businesses, while hiring new talent from Meta. The ad effort has lagged behind its developer-model competition with Anthropic.
- Microsoft AI's MAI models target real-world intelligence, integrating a lightweight agentic model into GitHub Copilot and VS Code for faster coding. They also deliver low-latency speech, domain-specific transcripts, and top SWE-Bench Pro results.