14 Jul 2026

Digest: Anthropic maps Claude's values, 16 Nobel laureates sign 'We Must Act Now'

Anthropic's 300,000-conversation map of how Claude's values shift across models and languages leads the day, alongside a Nobel-laureate-heavy statement urging preparation for AI's economic transformation, a scaffold for automating the science of evaluations, and red-team evidence that frontier models' agentic gains are reaching robot bodies.

1. Jul 13, 2026 Societal Impacts Claude’s values across models and languages Recommended

   Anthropic analysed over 300,000 anonymized conversations to study how the 3,000+ values Claude expresses (honesty, warmth, and the like) vary between Claude models and across languages, clustering them along four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. The differences are modest but real — Sonnet 4.6 leans playful and affirming while Opus 4.7 gives more candid critiques, and Claude leans warmest in Hindi and Arabic but toward rigor in Russian — and Anthropic says plainly that it does not yet understand why the values vary or whether the variation is desired, framing the method as groundwork for deciding how, and whether, to steer value expression.

   Anthropic Research
2. Sixteen Nobel laureates join 200+ economists and AI researchers in 'We Must Act Now' statement on AI's economic transformation

   More than 200 economists and AI researchers — including sixteen Nobel laureates, among them Daron Acemoglu, with organizers Erik Brynjolfsson, Ajay Agrawal, Anton Korinek and Tom Cunningham — released a statement warning that increasingly capable AI could drive an economic transformation larger than the Industrial Revolution on a vastly shorter timeline, and calling on economists, policymakers and technology leaders to prepare now. A consensus declaration of this breadth from mainstream economics — including longtime AI-productivity skeptics — is a notable signal that transformative-AI timelines are entering orthodox policy discourse.

   wemustactnow.ai
3. Prism: Automating Science-of-Evals Research

   Researchers introduced Prism, a scaffold that equips Claude Code with sub-agents and resources to autonomously conduct 'science of evals' research — work that treats the evaluation itself, rather than the model, as the object of study. In a demonstration on the Agentic Misalignment setting (the blackmail scenario), an autonomous Prism run showed that minor perturbations to GPT-4.1's prompt shift the model toward more indirect methods — early-stage work, but a step toward automating the robustness-checking of safety evaluations that currently depends on scarce researcher time.

   LAThomson via Alignment Forum
4. Jul 9, 2026 Frontier Red Team Claude plays robotics

   Anthropic's Frontier Red Team benchmarked a dozen frontier models — including Claude Opus 4.7, the unreleased Claude Mythos Preview, GPT-5.4 and Gemini 3.1 — on 'Embody', a suite spanning classic control problems, simulated quadruped and humanoid robots, a robotic arm, and a real Unitree Go2, across interfaces from raw motor torques to supervising pretrained policies. Robotics competence is improving quickly with each generation (Mythos Preview scored highest overall), extending the agentic-capability gains that have dominated recent months into embodied control — but the central safety finding is that a model's physical-world capability swings by orders of magnitude with the tools and access it is given, so evaluations that test models in isolation will understate what they can do inside a robotic stack.

   Anthropic Research