Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Illinois creates AI cabinet for $500M model firms
09/23/26, 02:47 AM

Illinois Gov. JB Pritzker signed Executive Order 2026-07 to create an AI Cabinet, while Illinois’s new safety law targets frontier developers above $500 million in annual revenue. The governor’s announcement says the cabinet will prepare for AI incidents and protect public assets; NPR Illinois reports that annual independent audits begin in 2028. Builders selling frontier models into Illinois should put incident reporting and audit evidence on their 2027 compliance calendar, alongside the archive’s frontier-model testing analysis.
Wire

20 countries call for frontier AI controls
09/23/26, 01:54 AM

Twenty countries and the European Commission have endorsed a call to keep frontier AI under human direction, oversight, and control. The official statement signed by 22 leaders from 20 countries asks companies for mandatory pre-deployment testing and independent evaluation, governments for shared incident reporting, and UN members to explore an institution that can verify compliance when capability thresholds are crossed. It is a nonbinding signal rather than a new rule, but builders should file its three asks as the direction of travel for future model-access and incident-reporting requirements; the EU AI Office’s 38-enforcer expansion shows where formal enforcement capacity is already accumulating.
Wire

Fluidstack starts a $4B Texas AI campus
09/23/26, 01:51 AM

Fluidstack says it has begun construction on the $4 billion first phase of an AI data-center campus in Cameron County, Texas, planned to support up to 1.5 GW. In its Cameron County construction announcement, the company says it expects more than 2,500 construction jobs, no local property-tax abatements, and closed-loop cooling; it also commits to paying for its own energy, infrastructure, and water. The scale makes power, water, and permitting inputs—not GPU supply—the first variables builders should stress-test against the broader data-center capex climb.
Wire

OpenAI proposes seven principles for AI assessments
09/23/26, 01:48 AM

OpenAI has published 7 principles and 4 assessment areas for independent third-party reviews of frontier models. The framework for independent AI safety assessments asks evaluators to examine safety cases, safeguard resilience, Preparedness Framework capabilities, and serious misalignment incidents, while requiring pre-registered claims, proportionate access, disclosed conflicts, secure handling, actionable findings, remediation time, and responsible publication. It is a vendor proposal—not a completed audit or certification—but builders should file it as an emerging template for evaluating agent vendors before granting production permissions; Anthropic’s safety-researcher economics show why evaluator access and incentives matter as much as model scores.
Wire

DolphinBench tests agent memory on 600 tasks
09/23/26, 12:56 AM

DolphinBench evaluates agent memory through 600 verified knowledge-work tasks: 200 each across three personas, with roughly 500,000 tokens of history per persona. Its arXiv paper requires evaluations to report accuracy alongside total cost and latency, rather than rewarding recall at any price. The release gives builders a cleaner way to compare memory layers against their operating budget; pair it with the archive’s analysis of shared agent memory before buying a larger context window as a substitute for durable memory.
Wire

Australia's data centers raise A$35B
09/23/26, 12:53 AM

Australian data-centre operators have raised at least A$35 billion in 2026, already 46% above 2025’s A$24 billion, and 85% of this year’s financing is debt, according to ABC’s report on an RBA staff note. About A$25 billion came through syndicated loans, while the estimate excludes some private and single-bank deals. For builders, this is a capital-stack fact, not proof of completed capacity: compare financing with delivery-backed milestones in the archive’s data-center capex analysis before treating an AI campus announcement as available compute.
Wire

UK sets October 13 AI safety hearing
09/23/26, 12:50 AM

Four frontier AI labs — OpenAI, Anthropic, Google DeepMind, and Meta — have been summoned to a UK parliamentary hearing on October 13, with MPs asking whether independent safety testing and serious-incident reporting should become mandatory. City A.M.’s account of the committee’s summons says the hearing will also probe whether regulators should be able to block deployment or require more testing. This is a hearing, not a new law, but builders should file it beside the archive’s analysis of pre-deployment AI testing: voluntary access to frontier models is moving from policy preference toward a deployment-risk variable.
Wire

Firecrawl raises $75M for agent web data
09/22/26, 09:54 PM

Firecrawl raised a $75 million Series B and launched Alexandria, a retrieval layer that combines official data providers, custom connectors, its indexes, and the live web, according to the company’s announcement. In a blind test across 845 tasks, Firecrawl says Alexandria produced 21% higher answer quality than built-in web tools. Agent builders can file the number as a retrieval benchmark—not a guarantee—and compare it with our agent-search API cost analysis before swapping search infrastructure.
Wire

Heidi Health raises $340M for clinical agents
09/22/26, 09:51 PM

Heidi Health announced $340 million in new funding: a $100 million Series C plus $240 million in growth financing, valuing the clinical-AI company at $900 million, according to its funding announcement. The company says it supports 2.8 million patient interactions weekly across 190 countries. For teams weighing ambient documentation, the shadow-clinic analysis of ChatGPT Health is the relevant caution: this capital signals a push from notes toward supervised clinical action, making safety evidence and workflow ownership the purchase test.
Wire

OpenAI halves GPT-6 Sol and Luna API prices
09/22/26, 09:48 PM

OpenAI released GPT-6 Sol and Luna at half the API price of their GPT-5.6 predecessors, according to TechCrunch’s report on the launch. Sol targets coding and complex tasks while Luna handles high-volume clerical work, so builders should rerun cost-per-accepted-task tests rather than promote a model on token price alone; our analysis of Gemini’s scheduled price expiry shows how quickly an attractive tariff can stop being the one your workflow pays.