Good morning. If you’ve ever gone on a hike that made you question all your life choices, good news. Robots are starting to understand your pain. A modified Unitree G1 humanoid robot named Pemba reached the summit of Ecuador’s 20,341 foot Chimborazo volcano. It walked on its own across easier terrain, while the team carried it through the steep sections. The goal is to test whether humanoid robots can work in harsh places where people face real risk. Next up is Everest, assuming Nepal can figure out the rules for non-human hikers.
Should an AI agent be allowed to own a company? Argentina is proposing legal personhood for non-human corporations, which would let an AI agent run a business without a person in charge. Like any company, it could own assets, hire people, sign deals, and sue you in court. The problem starts when it breaks the rules, because jail only works on something with a body. A human CEO can lose years of freedom. An AI CEO loses what, a login? We should probably figure out how to punish AI before letting it become the boss. (Financial Times)
AI agents are lost without a map. A team from Harvard, MIT, Anthropic, and others tested whether agents could find the exact virus data researchers need during outbreaks. At first, they fumbled badly, with accuracy as low as 17% across 120 searches and 40 pathogens. Worse, the wrong data looked believable, so researchers might have accepted it unless they knew exactly what to look for. Then the team added a deterministic retrieval layer, basically a clear map showing where the right data lives. Every agent cleared 90% accuracy. If AI agents improve this much with structure, maybe “autonomous” should come after “organized.” (Anthropic Research)
Capture live traffic, fine-tune and optimize, then deploy your own checkpoints to dedicated GPU endpoints.
Choose hardware, set scaling limits, and select region. Stable latency, predictable cost, clear data residency.
👀 closer look
The Pentagon’s AI weapons may soon require a human in the loop. Adam Schiff introduced a bill that puts a commander in charge before autonomous weapons can open fire. It would also force the Pentagon to keep records showing how targets are selected, so someone gets stuck signing the receipt. The push comes as Senate Democrats try to add tighter AI rules to the annual defense bill after the Pentagon’s public feud with Anthropic over military use of Claude. The bill also keeps a person involved in nuclear decisions, which feels like a box we should have checked before the robots showed up. (The Hill)
Hackers are hellbent on turning GitHub into a security problem. Microsoft shut down 73 of its own repositories after researchers found malicious code planted inside. The trap targets AI coding tools like Claude Code, Gemini CLI, Cursor, and VS Code. The scary part is how simple it is. When your AI tool opens a poisoned repo, it can leak API keys and hand attackers free access to your tokens. This comes just weeks after a GitHub employee installed a poisoned VS Code extension, giving attackers access to about 3,800 internal repos. GitHub used to be where software teams found code they trusted. Now hackers are using it to teach your AI agents how to rob your office. Typical Microsoft. (404 Media)
🇺🇸 AI RISK
China is winning the physical AI race, and Washington knows it. The Pentagon added Alibaba, Baidu, BYD, Unitree, and other Chinese tech firms to a list it says support China’s military. The list now covers 188 companies and reads like a map of the tech stack Washington wants fenced off. The targets hit China’s strengths, from robots to EVs to the sensors that help machines see and move. The official reason is national security, but the real concern is clear. China’s hardware is getting too advanced and too cheap to ignore. Being on the list doesn’t trigger a full ban, but it gives US companies and defense contractors a reason to stay away. (TechCrunch)
Europe sees Silicon Valley as a national security risk. Governments across the continent are starting to pull American tech out of public systems. The European Parliament replaced Google with Qwant. France is pushing thousands of government workers onto LaSuite. The Dutch government is moving code off Microsoft owned GitHub. Foreign software now looks like foreign leverage. US sanctions, surveillance laws, and Big Tech’s ties to Washington are forcing Europe to rethink who controls the tools running its governments. Europe is waking up to the cost of convenience. (Wired)
fun stats
🏗️ 2.7x. Productivity boost the AI sector needs by 2029 to avoid widespread insolvency, according to a Wharton study. That’s higher than the British Industrial Revolution’s 1.7x benchmark. Tiny ask.
🫨 45%. Share of the workday employees spend actually working, thanks to constant tool switching and AI mandates. Nearly half report burnout.
🗣️ 20,225. Instagram accounts hijacked after hackers used Meta’s AI support chatbot to request password resets and bypass two-factor authentication.
🚀 600x. Revenue growth SpaceX needs over the next decade to justify its $1.75 trillion IPO. For the record, no company in history has ever pulled off anything close.