AI is being measured at three layers: internal representations, scientific invention, and public risk. Anthropic reports a workspace involved in deliberate reasoning, while MLS-Bench finds that agents struggle to invent scalable methods. Illinois, the European Commission, and the ECB are translating capability into audit, classification, and macroeconomic questions.
1. Anthropic identifies a small internal workspace for Claude's reasoning
Anthropic researchers say they found a collection of internal Claude representations, called the J-space, that behaves like a functional global workspace. The patterns can be reported and altered, carry intermediate reasoning, and feed multiple downstream tasks. Removing them left routine language intact but sharply reduced multi-step reasoning in the reported experiments.
Anthropic explicitly separates the finding from any claim that Claude is conscious. Its Jacobian lens only approximates the workspace and favors concepts expressible as single tokens, but the method surfaced evaluation awareness, fabricated-data intent, and planted hidden goals before those states appeared in model output.
Sources: Anthropic's global-workspace research
2. MLS-Bench separates method invention from engineering optimization
MLS-Bench asks AI agents to improve machine-learning systems in ways that generalize and scale instead of merely tuning a fixed test. The benchmark contains 140 tasks across 12 domains. Its authors report that current agents remain far from reliably beating human-designed methods and perform better at engineering adjustments than genuine method invention.
The July 6 preprint revision also examines test-time scaling, adaptive compute, and added context. None removes the reported bottleneck in forming, validating, and scaling a scientific idea. That distinction matters for claims about automated AI research: producing more experiments is not equivalent to discovering a method that survives controlled changes in setting.
Sources: MLS-Bench paper and July 6 revision
3. Illinois signs a frontier-model audit and incident-reporting law
Governor J.B. Pritzker signed Senate Bill 315, the Artificial Intelligence Safety Measures Act. The statute requires covered model developers to publish catastrophic-risk frameworks, obtain annual third-party audits, and report qualifying incidents within 72 hours, or within 24 hours for an imminent risk of death or serious injury. It takes effect January 1, 2028.
Coverage combines a computing threshold with more than $500 million in annual revenue, concentrating the rules on the largest model developers. The attorney general can seek civil penalties, and Illinois adds recurring audits to provisions modeled on California and New York. That alignment may turn three state regimes into a practical national baseline before Congress acts.
Sources: Illinois Senate Bill 315 enrolled text · WTTW on the bill signing and requirements
4. The ECB maps AI's competing inflation channels
ECB Executive Board member Philip Lane used a July 6 speech to set out how AI could alter monetary-policy transmission. Productivity can reduce production costs, but expected income gains can lift demand, adoption requires investment, and data centers add electricity and fuel pressure. Labor displacement and uncertain household income can pull spending in the opposite direction.
This was scenario analysis, not a rate decision or forecast. Its significance is that AI adoption enters inflation analysis through timing and distribution, not one assumed productivity dividend. Monetary policy must distinguish temporary investment and energy shocks from persistent changes in output, wages, markups, and the neutral interest rate as evidence accumulates.
Sources: ECB speech on AI and monetary policy
5. EU draft guidance turns high-risk classification into examples
The European Commission published draft guidelines for deciding whether an AI system is high-risk under the AI Act. The examples cover providers and deployers across biometrics, critical infrastructure, education, employment, migration, asylum, and border control. A targeted consultation remains open through July 23, 2026.
The guidance is not legally binding and will change after consultation, but it records the Commission's intended enforcement interpretation. It also reflects the proposed Omnibus timetable: relevant stand-alone systems from December 2, 2027, and AI embedded in regulated products from August 2, 2028. Classification work now has concrete examples without yet having final safe harbors.