MISO’s large-load path: approval within 120 days with studies and GIAs signed — ZGIA co-location, ERAS, and EPR after FERC’s June direction.
What happened: Midcontinent Independent System Operator’s Large Load Additions page frames an unprecedented surge in large-load interconnection requests from data centers, advanced manufacturing, AI computing facilities, and other energy-intensive industries. Many customers want to connect within 18–36 months (also stated as 1.5–3 years), faster than traditional generation and transmission build timelines. Clock locked: on 18 June 2026, FERC directed all RTOs/ISOs, including MISO, to update large-load tariff provisions (195 FERC ¶ 61,212). Framework locked in three parts: (1) Connect quickly — large loads can get MISO approval within 120 days with all studies and generator interconnection agreements signed; named tools include co-located resources (ZGIA), expedited generation (ERAS), fast load studies (EPR), and a new study process still under development. (2) Transition and scale with flexibility — ZGIA, non-firm service, and firm service step-up options, with needs around curtailment capability and resource-adequacy coordination. (3) Operate with long-term reliability — withdraw or inject based on conditions; needs include real-time telemetry, curtailment, ride-through, stability control, resilience, and security. Scale language locked only as potential to add thousands of megawatts — not a queue census and not a TWh path. This is a Midwest stakeholder process and tariff-response framework — not SPP’s CHILL/HILLGA 90-day conditional path, not CAISO’s FERC show-cause clock (EL26-71 / draft final 24 September), not PJM ride-through after nearly 4,000 MW of disconnection, not ERCOT Batch Zero queue megawatts, and not IEA or LBNL national TWh shares.
Why it matters: After Plains conditional service, West-coast tariff clocks, and ride-through design, the environment beat is how the Midwest operator sells speed via a 120-day approval claim and co-located generation tools — reliability tradeoffs, not another metered growth print.
Fed FEDS 2026-018: coder employment decelerated sharply since ChatGPT — an occupation-specific shock, still growing but far slower than pre-2022.
What happened: Board of Governors staff working paper FEDS 2026-018 by Crane and Soto, AI and Coder Employment: Compiling the Evidence (PDF date 20 March 2026), asks whether large language models have had any discernible aggregate labor-market impact so far, focusing on computer programming-intensive occupations because coding is among the most LLM-exposed tasks. From the Board HTML abstract and PDF title/intro only: linking O*NET to CPS, aggregate employment of coders has decelerated sharply since the introduction of ChatGPT. A novel industry-level shock control shows the slowdown is not explained by coders’ exposure to slowing industries — it is an occupation-specific shock around ChatGPT. Coder employment has continued to grow in recent years, but much more slowly than pre-2022. Authors say results suggest AI is having a measurable and potentially consequential impact on employment for some groups, while noting they cannot address general-equilibrium demand effects. No independently extracted percentage-point gap or coder headcount is locked here. This is a staff working paper on one occupational group — authors’ views, not FOMC policy, not yesterday’s Soto–Thieu–Allen FEDS Note on buildout versus broad displacement, not NY Fed regional 61%/51% adoption packages, not Philadelphia Fed LIFE, not Minneapolis Fed Freund–Mann task clusters, and not a mass-layoff census.
Why it matters: After buildout dashboards and whole-market perception splits, the jobs beat is a Board staff paper that isolates coder slowdown as occupation-specific — still growing, just not like the pre-ChatGPT path.
CISA and partners: careful adoption of agentic AI — start low-risk, avoid unrestricted access, account for attack surface and privilege creep.
What happened: CISA’s resource page Careful Adoption of Agentic AI Services (dated 1 May 2026) and a same-day news release with Australia’s ACSC and other U.S. and international partners warn that critical infrastructure and defense are increasingly deploying agentic AI. Extra risks named on the pages: expanded attack surface, privilege creep, behavioral misalignment, and obscure event records. Action locked from the news page: avoid granting broad or unrestricted access, especially to sensitive data or critical systems; begin with low-risk, non-sensitive use cases; account for agentic AI security in the organization’s security model and risk posture. Status locked: joint guidance — not a Binding Operational Directive in this extract and not a statute. This is an agentic-AI adoption security guide — not NIST’s AI RMF Critical Infrastructure profile-in-development, not NIST AI 200-2, not SP 800-239 campus/data-center comments due 25 September, not the NVD modernization RFI due 13 October, and not a restatement of day-51 EU Article 50 duties (calendar only).
Why it matters: After CI profile concept notes and vulnerability-database modernization clocks, the policy beat is how operational cybersecurity agencies tell infrastructure operators to adopt agents without handing them the keys on day one.
Nature on AISB / OpenFold3: five drug firms’ private protein–ligand data lifts held-out accuracy above public models — not peer-reviewed, model not public.
What happened: A Nature news recap describes the AI Structural Biology (AISB) Network — AbbVie, Astex, and several other drug companies — fine-tuning OpenFold3 (an open-source AlphaFold 3 replication previously trained on the PDB) on a further 20,167 structures of proteins bound to potential drugs or ligands from five companies, with proprietary data kept private. On 1,056 held-out protein–ligand structures, the AISB model predicted more than half to a high level of accuracy; publicly available OpenFold3 hit that bar on just one-third; Boltz-2 about 40%. The pooled model also beat tools trained only on each company’s individual data. Limits locked: described in a blog post; not peer-reviewed; the model is not publicly available; the team plans to submit a paper. Public-data gap named: the PDB has more than 200,000 protein structures but relatively few drug-like interactions (maybe just 10,000, attributed to Astex’s Mortenson). UK OpenBind is named as a public-data effort supported by up to £8 million (US$10.8 million), with hundreds of new structures released last month. This is an industry-consortium structural-biology model recap — not WHO’s 21 September ethics-oversight report, not Virtual Biotech’s multi-thousand-agent discovery swarm, and not a bedside clinical RCT.
Why it matters: After ethics-oversight guidance and agent-swarm discovery headlines, the science beat is how pooled proprietary binding data can lift protein–ligand models — with the honesty that the lift is still unpublished and closed.
UNESCO’s “algorithm in the room”: 26 think pieces and six questions on agency, pedagogy, and education as a common good.
What happened: UNESCO article The algorithm in the room: Seeing and confronting the implications of AI for the future of education (dateline 8 September 2026; last update 9 September 2026) introduces a compilation of 26 think pieces confronting “elephant in the room” questions about education in an era of AI and algorithms. Six urgent questions locked: (1) who will hold agency, autonomy and oversight; (2) how young people will develop relationally, emotionally and cognitively; (3) what will be meaningful for pedagogy and assessment; (4) how we will think, question and express ourselves; (5) how we will govern education as a common good accountable to the public; and (6) how AI might affect imagination about education’s future. Limit locked: think-piece volume — not an RCT, not an enrollment census, and no independently extracted student n, country count, or uptake percent on the HTML. This is a governance-and-agency essay layer — not World Bank Mongolia TVA / Medle lesson-plan counts, not UNICEF’s child-uptake snapshot, not World Bank The Job I Hope For, and not a PISA print (still not independently extracted here).
Why it matters: After national teacher-tool architecture and child-protection snapshots, the education beat is the harder public questions — who keeps agency when algorithms enter the room.
A 120-day Midwest path is not a Plains 90-day clock.
What happened: Keep MISO as FERC 18 June 2026 / 195 FERC ¶ 61,212 tariff-response framing; 120-day approval with studies and GIAs; ZGIA / ERAS / EPR; 18–36 month connect window; “thousands of megawatts” as potential only. No queue MW census and no TWh locked. Keep separate from SPP CHILL/HILLGA 90-day, CAISO EL26-71 / 24 September draft final, PJM ride-through ~4,000 MW, ERCOT Batch Zero, IEA +17% 2025, IEA Grids queue totals, and LBNL share.
Why it matters: Regional interconnection design is a different evidence class than metered growth rates or single-event disconnection reports.
Coder slowdown is not a whole-market pink-slip count.
What happened: Keep FEDS 2026-018 as occupation-specific coder deceleration since ChatGPT; not industry mix; still growing, much slower than pre-2022; measurable impact for some groups; no pp or headcount locked from the abstract. Keep separate from yesterday’s Soto–Thieu–Allen buildout FEDS Note, NY Fed 61%/51%, Philadelphia Fed LIFE, Minneapolis Fed Freund–Mann, Boston Fed CPP 26-8, and St. Louis RPS shares.
Why it matters: An occupation-shock working paper is a different evidence class than buildout dashboards or household fear surveys.
Joint agentic-AI guidance is not a finished CI control set.
What happened: Keep CISA’s May 1 guide as careful adoption with partners including ASD’s ACSC; start low-risk; no unrestricted access; named risks of attack surface, privilege creep, misalignment, and obscure logs. Day-51 EU Article 50 remains calendar only. Standing clocks only: Canada ISED still open through 23 September; FTC personalized pricing and NIST SP 800-239 still due 25 September; NIST AI 200-2 still due 6 October; NVD modernization RFI still due 13 October; UNESCO consultation still due 15 October; FDA GenAI-device comments still due 19 October; CAISO Board still 28 October.
A proprietary-data lift is not a clinical endpoint trial.
What happened: Keep AISB/OpenFold3 as 20,167 structures from five companies; 1,056 held-out; more than half vs one-third vs ~40%; pooling beat single-company training; not peer-reviewed; model not public; OpenBind up to £8 million / $10.8 million. Keep separate from WHO ethics oversight and from Virtual Biotech agent-swarm discovery.
Why it matters: Closed structural-biology model gains are a different evidence class than ethics-review guidance or bedside RCTs.
Twenty-six think pieces are not a learning-outcomes census.
What happened: Keep UNESCO’s algorithm-in-the-room volume as 26 pieces and six questions on agency, development, pedagogy, expression, public governance, and future imagination; 8–9 September 2026 dateline. No student n locked. Standing context only: UNESCO consultation comments due 15 October. Keep separate from Mongolia TVA lesson-plan counts and UNICEF child-uptake snapshots.
Why it matters: Governance essays are a different evidence class than platform rollouts or child-protection uptake prints.
The full daily ledger keeps broader source-linked coverage organized by topic. Story dates are shown separately from the September 22 edition date.
September 22 · MISO 120-day / ZGIA
MISO Large Load Additions: FERC 18 June 2026 / 195 FERC ¶ 61,212; large loads can get MISO approval within 120 days with studies and GIAs signed; ZGIA co-location, ERAS, EPR; connect window 18–36 months; potential thousands of megawatts — not a census.
Midwest tariff-response framework — not SPP 90-day CHILL/HILLGA; not CAISO EL26-71; not PJM ride-through.
Surge driven by data centers, advanced manufacturing, AI computing facilities, and other energy-intensive industries seeking connection faster than traditional generation and transmission builds.
Process design — not ERCOT Batch Zero; not LBNL TWh share; not IEA +17% 2025.
Board FEDS Working Paper 2026-018 (Crane–Soto, 20 March 2026): coder employment decelerated sharply since ChatGPT; occupation-specific shock, not industry mix; still growing, much slower than pre-2022.
Staff working paper — not Soto–Thieu–Allen buildout note; not NY Fed 61%/51%; not FOMC.
Authors: coding is a predominant generative-AI use; results suggest measurable and potentially consequential employment impact for some groups; cannot address general-equilibrium demand effects. No pp or headcount locked from the abstract.
Occupation-shock evidence — not Philadelphia Fed LIFE; not Minneapolis Fed Freund–Mann.
Nature news on AISB Network OpenFold3 fine-tune: 20,167 proprietary protein–ligand structures from five companies; 1,056 held-out; more than half high accuracy vs one-third public OpenFold3 vs Boltz-2 ~40%; not peer-reviewed; model not public.
Industry-consortium model recap — not WHO ethics oversight; not Virtual Biotech ~37k agents; not a bedside RCT.
UNESCO (8 September 2026): 26 think pieces; six questions on agency/oversight, relational-cognitive development, pedagogy/assessment, expression, education as a common good, and future imagination.
Governance-and-agency essay layer — not Mongolia TVA >700/4,900; not UNICEF ≥20 million / >3×; not PISA.
Day-51 EU Article 50 calendar only; Canada ISED closes 23 September; FTC personalized pricing and NIST SP 800-239 still due 25 September; NIST AI 200-2 still due 6 October; NVD RFI still due 13 October; UNESCO consultation and SPP Q&A still due 15 October; FDA GenAI-device comments still due 19 October; CAISO Board 28 October.
