Subscribe
Share
Share
Embed
Norway bans AI from classrooms for under-13s
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Saturday, June twentieth.
In today's briefing we see John Jumper, the AlphaFold co-lead and twenty twenty-four Nobel laureate, joining Anthropic to head AI-for-science work, Z.ai's GLM-5.2 outperforming frontier closed models on coding benchmarks at a fraction of the cost, and Hyundai taking full control of Boston Dynamics as the robotics market reprices.
First up - Today in the big model news;
Google + Deepmind / Gemini
DeepMind published the AI Control Roadmap on June eighteenth, the first formalized defense-in-depth framework for agentic deployments from a major lab. The roadmap maps safeguards across detection tiers (D1 through D4) and prevention and response tiers (R1 through R3), explicitly modeling threat scenarios where alignment fails: model exfiltration, rogue internal deployment, work sabotage, direct harm. For product teams planning enterprise agentic deployments, this framework sets a new standard for secure agentic deployment, because DeepMind is treating alignment failure as a planning input rather than an aspirational assumption.
Anthropic - Claude
John Jumper, the AlphaFold co-lead and twenty twenty-four Nobel Chemistry laureate, is leaving Google DeepMind after nine years to join Anthropic. Jumper's protein structure work is used by over two million researchers across one hundred ninety countries. Anthropic has spent twenty twenty-six building an AI-for-science vertical: opening wet labs, publishing agent research in biology, and forging partnerships with the Allen Institute and Howard Hughes Medical Institute. His arrival signals that this vertical is real. For research teams evaluating frontier model deployments for scientific workflows, Jumper's move is a credible signal that Anthropic is investing in deep science infrastructure, because the company is now attracting the most credentialed scientists from the labs that pioneered computational biology.
In the local model developments space;
Open-source models are resetting the market. OpenRouter data shows open-weight models at roughly sixty percent of routed LLM traffic, up from about forty percent three months prior, a shift directly driven by enterprise fallback planning around the Fable 5 suspension. Z.ai's GLM-5.2, a seven hundred forty-four billion parameter mixture of experts with forty billion active parameters, is outperforming GPT-5.5 on seven of twelve shared benchmarks including SWE-Bench Pro (sixty-two point one percent versus fifty-eight point six percent), tool use, and multi-turn planning, at roughly one-sixth the cost. Practitioners report GLM-5.2 as rivaling Opus four point eight for local coding and planning workflows; the gap keeping it off the top slot for mixed-task pipelines is lack of multimodal support. For product teams building agentic systems on open-weight infrastructure, the business case for staying on closed frontier models just narrowed, because what's achievable at commodity pricing has crossed the threshold where feature parity is now the binding constraint, not raw capability.
In the harness, tools and orchestration world;
Two structural patterns are shaping production agent engineering. The first is fan-out: master agents decomposing tasks and spawning five to one hundred child agents in parallel, documented in Cognition and Devin production workflows. The second is loop reliability as a first-class engineering discipline: explicitly designing for client, server, and inference failures rather than letting the agent loop silently degrade. Nous Research released a new Hermes version this cycle with session compression, agent distribution, and iMessage integration, specifically for production builders. For teams shipping agentic automation at scale, the gap between a demo agent and production is increasingly a harness problem, because the fastest-compounding teams are closing harness quality faster than they upgrade models.
In other news;
Norway's government announced a near-total prohibition on generative AI for students in grades one through seven (ages six to thirteen), effective at the start of the school year in late August. Students aged fourteen through sixteen may use AI only under teacher supervision. Prime Minister Jonas Gahr Stoere cited the risk that AI causes children to skip foundational learning steps. This is the sharpest government restriction on AI use by developmental stage from any major democracy to date. For product teams targeting K through twelve markets, this regulation provides a template likely to spread across Europe, because the EU AI Act's risk-based framework doesn't capture age-stage prohibitions, meaning government action now runs ahead of formal regulation.
Hyundai is acquiring SoftBank's remaining nine point sixty-five percent stake in Boston Dynamics for three hundred twenty-five million dollars pending board approval on June twenty-second, making the robotics firm wholly owned. The implied valuation is approximately twenty-two billion dollars, roughly a twenty-times increase from the one point one billion dollar valuation when Hyundai first took controlling interest in twenty twenty-one. Atlas humanoid robot commercial production begins this year with initial units earmarked for Hyundai's own manufacturing facilities and Google DeepMind. For teams building embodied AI or robot-adjacent software, the twenty-times valuation surge over five years signals a fundamentally different capital environment, because the market is repricing physical AI infrastructure as production-ready rather than experimental.
That's the briefing. Have a great day.