metir Blog

Bloomberg Intelligence says the US-China AI gap narrowed to about 3% on LiveBench after DeepSeek V4.1 Flash scored 81.1 versus 83.4. What the number means.

IWF found 6,310 AI-generated child sexual abuse images in H1 2026, more than all of 2025. A policy analysis of UK, EU and US law and safety by design.

California SB 1246 lets cities fine robotaxi operators up to $10,000 per vehicle for blocking emergency responders over 30 minutes. What it requires and when.

A PLOS Digital Health audit found only 3 of 1,357 FDA-cleared AI medical devices were evaluated for patient outcomes. Here is what it means for hospitals.

Huawei and Qualcomm announced a multi-year patent license deal covering 5G, compute, AI and networking, with Qualcomm buying some Huawei US patents.

GitLab patched CVE-2026-90970, a CVSS 9.9 AI Gateway flaw letting Duo Agent Platform users escape a prompt template sandbox. What it teaches about agent security.

Munich's RobCo crossed a $1B valuation through a $40M+ secondary share sale, doubling its January price. What it means for physical AI and industrial robots.

The Firmus IPO would give about half its ASX shares to existing holders such as Nvidia and Blackstone. What the deal shows about neocloud listings and AI data center funding.

Nvidia says Anthropic's contracted value across cloud and neocloud providers tops $180 billion, with 2.6 GW of Nvidia capacity through 2028. What the figure covers.

OpenAI will test clearly labeled ChatGPT visual ads during image generation in the US later this month. What is confirmed, what is not, and what to watch.

Jensen Huang says there is a 0% chance AI ends the world by 2030. A neutral guide to the AI risk debate: who says what, the incentives argument, and policy.

ARC-AGI-3 Kaggle scores have jumped past 55% using small open-weight models in harnesses. Here is what the benchmark tests, the prize terms and the caveats.

North Korea says an AI hypersonic missile test changed course at low altitude. Seoul and Tokyo measured shorter flights. What is claimed, disputed and unknown.

Insilico Medicine used AI to design BTK PROTAC degraders for leukemia and lymphoma. How PROTACs work, the rentosertib track record, and AI vs traditional timelines.

Supabase raised $150M led by GIC and agreed to acquire Turso as agents create 70% of its new databases. Funding path, database-per-agent economics, rivals.

Armadin raised $255.5M at a $2.5B+ valuation to build swarms of AI attack agents. The funding math, exploit-timing data and dual-use questions explained.

Meta says Muse Spark helped mathematicians write six papers, five answering open questions. What was claimed, what was checked, and what to weigh.

Local opposition to AI data centers is no longer a US story. Verified figures on power, water and land friction, and what it means for where compute gets built.

Anthropic's Claude Code mods let TypeScript hooks rewrite prompts, gate tool calls and redraw the interface. What is new, how it compares, and the risks.

Google limits the free Gemini app to Flash-Lite on October 9, 2026, removes Pro from AI Plus and adds Deep Think to AI Pro. Tier table and cost analysis.

Project Meridian gives Musk, Luckey and Gingrich 120 days to study future warfare. What the timeline can produce, how it compares, and the conflict questions raised.

OpenAI and Synopsys announced GPT-Synopsys on Sept 30, 2026, a model trained to operate EDA tools. Here is why chip design is a hard test for AI agents.

Treasury Secretary Scott Bessent said AI alarmism without solutions is not leadership. What he said, who said what, and how AI liability is allocated.

Google suspended product vulnerability reports to its OSS VRP on October 1, 2026 after a flood of invalid AI submissions. What it says about bug bounty economics.

Microsoft's Digital Defense Report 2026 finds AI gives attackers the early edge, with exploits weaponized in under 24 hours. What it means for patching and AI agents.