Wednesday, September 23, 2026 · Daily edition
ISO-NE’s not-yet-at-scale load clock, NY Fed postings on AI hiring, and cloud token hardening
Today’s ledger follows New England’s still-empty large-load clock after FERC’s show-cause order, a New York Fed staff blog that finds little distinct AI-driven drop in job postings, joint CISA–NIST guidance on stolen and forged identity tokens, a Nature feature on Google’s multi-agent Co-Scientist workflow, and a UNESCO IESALC census of AI uptake versus strategy across Latin American and Caribbean universities.
ISO-NE large loads: not yet at scale in New England — with a 16 November tariff clock
What happened. ISO New England’s Large Loads Integration Key Project page treats data centers and other large loads as a reliability and interconnection design problem that has not yet materialized in New England at scale. On 18 June 2026, FERC issued a show-cause order under Section 206 (EL26-72-000) on large loads and co-located load arrangements. On 14 August, FERC granted a 90-day abeyance so ISO-NE can run a stakeholder process ahead of a 16 November 2026 Section 205 tariff filing covering five reform categories named in the order. A 20 July informational filing told FERC the operator is still developing resource-adequacy provisions for large-load interconnections and reliability risks. The page’s scale claim is qualitative only — not a queue megawatt census and not a national electricity path. It is a New England stakeholder process and sixth-RTO tariff clock, not MISO’s 120-day path, SPP’s conditional 90-day tools, CAISO’s EL26-71 clock, or PJM ride-through design after nearly 4,000 MW of disconnection.
What to watch. After Midwest speed claims and West Coast tariff clocks, the environment beat is how New England still says the loads have not arrived at scale — with a November filing still ahead.
Read the ISO-NE large loads key project →
NY Fed Liberty Street: postings show little distinct AI hiring drop — the relative decline predates ChatGPT
What happened. Federal Reserve Bank of New York economists Audoly, Guerin, and Topa ask whether job postings already show early labor-market effects of AI. Their 14 May 2026 Liberty Street Economics post links an Anthropic occupational exposure metric to U.S. Lightcast postings in an event study around ChatGPT in late 2022. Less than 10 percent of workers and vacancies sit in occupations with exposure of at least 0.4; about 40 percent of workers are in jobs with zero measured AI exposure. Overall hiring has slowed since ChatGPT, but the authors find little indication of a distinct AI-driven decline in labor demand. The relative drop in vacancies for higher-exposure occupations began before 2022, does not show a clear extra break after ChatGPT, and stabilizes after 2023. Junior and senior postings inside high-exposure occupations move broadly in parallel, so the slowdown is not concentrated in entry-level highly exposed jobs. AI may be contributing, they conclude, but is not the main driver. It is a staff blog on national postings — authors’ views, not FOMC policy, not yesterday’s Board FEDS coder paper, and not a mass-layoff census.
What to watch. After coder-occupation slowdown papers and state automatable-task prints, the jobs beat is a national postings study that puts the relative AI-exposure gap before ChatGPT and then flat — not a pink-slip count.
Read the New York Fed Liberty Street post →
CISA and NIST finalize IR 8587: harden cloud identity tokens against theft and forgery
What happened. CISA and NIST finalized IR 8587, implementation recommendations for agencies and cloud providers on protecting tokens and assertions from forgery, theft, and misuse. The final report folds in nearly 250 public comments on token validation, secrets management, and detection at scale. It builds on NIST SP 800-53 control IA-13 and responds to Executive Order 14306. NIST’s news notes an attack in which foreign actors used forged tokens from one stolen commercial signing key to access agency email and steal more than 60,000 emails from a single agency. The authors also add high-level considerations for AI handling and migration to post-quantum cryptography — without claiming a full toolkit for either. Status is a final interagency implementation report, not a Binding Operational Directive and not a statute. Audience is federal agencies and cloud providers; any organization using tokens in access management can look to it. This is identity-token and SSO hardening with an AI and post-quantum note — not yesterday’s agentic-adoption guide and not a finished agent-control set.
What to watch. After agentic-adoption guidance, the policy beat is how agencies harden the tokens that open cloud identity systems — including a high-level AI and post-quantum paragraph, not a finished control catalog.
Read the CISA–NIST IR 8587 release →
Nature on Google Co-Scientist: more than 700 papers, 108 strategies, one left — glue condensates
What happened. A Nature technology feature describes Google’s Co-Scientist system launching several autonomous agents to search and synthesize papers, evaluate competing explanations, refine and critique hypotheses, and test ideas against published evidence — not a single-pass chatbot. In a Whitehead Institute MYC condensate case, after about an hour of correction for early mistakes, Co-Scientist combed more than 700 papers and generated 108 possible strategies, then rejected all but one. The remaining idea: glue condensates together with click chemistry rather than dissolving them, so MYC DNA can no longer be read. Lab researcher Overholt is quoted that the group “had certainly never thought about anything like this.” Prior feature illustrations include a drug combination that kills leukaemia cells in a dish and a lab treatment that regenerates diseased liver tissue. Limits are plain: journalism feature, not a peer-reviewed trial metric and not a bedside authorization; the hypothesis still needs human verification. This is a hypothesis-generation workflow recap — not yesterday’s closed AISB / OpenFold3 protein–ligand lift.
What to watch. After proprietary structural-biology lifts, the science beat is how multi-agent co-scientist systems can name testable strategies humans had not put on the table — still a feature, not a clinical endpoint.
Read the Nature Co-Scientist feature →
UNESCO IESALC: 87% of surveyed LAC universities use AI — only 26% have a formal strategy
What happened. UNESCO’s International Institute for Higher Education in Latin America and the Caribbean, with UNU-IAS, reports a census of AI implementation across 200 higher-education institutions in 19 countries. Eighty-seven percent use AI in at least one area; 74% in teaching and learning, often with general-purpose tools such as ChatGPT, Copilot, and Gemini; 57% in research. Only 26% have a formal AI strategy, and even fewer have clear policies, governance structures, and monitoring. Adoption is often driven by faculty, researchers, and students rather than institutional policy. Staff training is common; student training lags, with 1 in 2 students reporting confusion about appropriate AI use. Private institutions tend to lead on strategy, staff training, and governance; public institutions face tighter resource and capacity constraints. Limits are plain: institutional self-report census — not a randomized trial, not a K–12 enrollment print, and not PISA. This is uptake-versus-governance evidence, not yesterday’s algorithm-in-the-room think-piece volume.
What to watch. After governance essays and teacher-tool rollouts, the education beat is a regional university census: AI is already in classrooms and labs, while formal strategy and student guidance lag.
Read the UNESCO IESALC study release →
Also in today’s ledger
• A New England not-yet-at-scale clock is not a Midwest 120-day path — and not a terawatt-hour census.
• A pre-ChatGPT relative posting drop is not a coder pink-slip count — and token/SSO hardening is not an agentic-control set.
• A co-scientist hypothesis workflow is not a clinical endpoint — and 87% uptake with 26% strategy is not a learning-outcomes print.
|