AI Labs Get Mediocre Grades, South Korea Bets $880B, Karpathy Joins Anthropic

Today’s AI news feels like a report card day where nobody’s proud of their grades, a government drops a monster check on the table, and a celebrity researcher switches teams. Let’s break it down.

AI Labs Get Graded on Safety, and Nobody Does Well

The Future of Life Institute dropped its 2026 AI Safety Index, and the best grade anyone got was a C+ for Anthropic. OpenAI and Google DeepMind landed at C, Meta got a D+, and xAI, DeepSeek, and Mistral basically failed. The index scores each lab on risk management, transparency, and whether they actually keep the safety promises they make in public.

A C+ being the top of the class is the real headline. An independent group is essentially saying that even the most safety‑focused AI company is doing a mediocre job by its own stated standards, right as these systems get wired into hospitals, cybersecurity, and self‑driving software. The report also found that several labs have quietly walked back safety commitments they made earlier, usually when they were raising money. Anthropic topping the chart fits its 2026 image as the careful, enterprise‑friendly lab, but a C+ is not a gold star.

My take: independent report cards like this are worth ten corporate safety statements. The fact that the whole industry is quietly getting a C average while telling us everything is fine tells you exactly why watchdogs like this need to exist.

Abstract robot agents collaborating over a floating holographic code screen

South Korea Bets $880 Billion on AI

South Korean President Lee Jae-myung announced a ten‑year AI plan worth about 1,350 trillion won, roughly $880 billion, one of the largest national AI commitments any country has ever made. The money splits into around $518 billion for memory chip factories through Samsung and SK Hynix, about $550 billion for AI data centers, a target of 8.4 gigawatts of data‑center power by 2029, and a push to grow the country’s share of the humanoid robot market from 1 percent to 20 percent by 2028.

This is a country going all in. South Korea already sits at the heart of the AI hardware world, since SK Hynix makes around 60 percent of the special memory chips every AI processor needs. This plan is meant to lock in that lead against Taiwan, the US, and China. To put $880 billion in perspective, that is a nation spending close to a trillion dollars to make sure it owns a piece of the AI decade, not just rents it.

The robot target is the wild part. Going from 1 percent to 20 percent of the humanoid robot market in two years is a huge leap, and it signals that Seoul thinks physical robots, not just chatbots, are the next big prize. Whether a government‑planned bet this size beats messier market‑driven efforts elsewhere is the experiment to watch.

My take: the countries treating AI like national infrastructure, the way they once treated highways and electricity, are the ones positioning to win the next 20 years. South Korea just made that official with a near‑trillion‑dollar checkbook.

Stylized futuristic terminal window with glowing prompt and agent avatar

Andrej Karpathy Reportedly Joins Anthropic

Andrej Karpathy, the former Tesla AI director and an OpenAI founding member, has reportedly joined Anthropic, along with Monzo co‑founder Tom Blomfield, who is joining the AI compute team. Karpathy is one of the most respected and widely followed people in AI, known as much for teaching millions of people how neural networks work as for building them. Where he chooses to work is a signal the whole field watches.

This adds to an incredible run of hires for Anthropic in 2026, which earlier landed Nobel Prize winner John Jumper on the AI compute team. Karpathy’s move underscores how aggressively Anthropic is building out its talent stack as it pushes into enterprise‑grade agentic systems and safety‑focused model development.

My take: when a name like Karpathy moves, it’s not just a résumé line—it’s a signal about where the smart money thinks the action is. Anthropic’s hiring spree suggests they’re betting big on being the go‑to for serious, safety‑conscious AI work.

Sources & Further Reading

If you’re looking to dive deeper into AI safety, consider Superintelligence: Paths, Dangers, Strategies by Nick Bostrom (Amazon). For hands‑on agent building, Agent‑Based Modeling for Social Systems is a solid pick. And if you want to experiment with hardware, the NVIDIA Jetson AGX Orin Developer Kit lets you run models at the edge.

That’s it for today’s spaghetti‑style digest. Stay curious, stay skeptical, and keep building.