#114: Impact Signals #114 — One Nepali Gauge Recorded Nine Metres in Thirty Minutes, MIT Generates the Storm That Never Happened, $16M Lands on AI Evaluation
AI for Impact Daily Briefing, August 27, 2026
Top Stories
Nepal: ICIMOD Finds the Bhote Koshi Early Warning System Had the Wrong Geometry, Not No Coverage
The International Centre for Integrated Mountain Development published its assessment on August 26 of the flash flood that swept Rasuwa district, Nepal. ICIMOD scientists suspect a large volume of ice and rock entered the Lhende Khola, a tributary of the Bhote Koshi on the Tibetan side roughly 20 km northeast of the Rasuwagadhi border point; the cause remains unconfirmed, and the nearest seismic station is 12 km away at Jilong. Several of Nepal's hydrological monitoring stations in the flood path were damaged or destroyed by the flood itself. The Galchchi station stayed up and recorded a rise of 9 metres in about 30 minutes; Malekhu recorded 7 metres in a comparable window. Nepal's National Disaster Risk Reduction and Management Authority put the toll at 165 dead as of August 27 and rising, with 123 people airlifted out of Timure (94 Nepali, 29 foreign nationals); the Nepal Tourism Board's preliminary tally lists 579 travellers unaccounted for. The Kathmandu Post reported one hydropower project washed away entirely, damage to the Chilime project, and the Timure helipad swept away, which is why the first rescue helicopter could not land. Why it matters: The Bhote Koshi has an early warning system, and the measurable finding here is that it was built to watch the river while the trigger sat in an ice-rock collapse across an international boundary, with the sensors closest to the impact destroyed by the event. What is newly on the record is the physical envelope any future system has to beat: 9 metres at Galchchi in 30 minutes, from a source 20 km inside another country, with no operational channel moving upstream data downstream in the minutes such a surge allows. ICIMOD's named team, led by director general Pema Gyamtsho with Basanta Raj Adhikari of Tribhuvan University's Centre for Disaster Studies, states the gap as cross-border; Neera Shrestha Pradhan's added point is that a warning is only half of it. The same stretch of river flooded on July 8, 2025, killing at least nine, and casualty figures for this event remain provisional.
MIT's Eta-Learning Generates a 300-Millimetre New York Storm With No Match in the Record
MIT announced on August 24 a generative method called Extreme Event Aware learning, or eta-learning, published in Nature Communications on August 20 by Kai Chang, a graduate student in mechanical engineering, and Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering. Machine learning models of extreme weather are trained on the historical record, which contains almost no examples of the events planners most need to size infrastructure against, so those models reproduce the middle of the distribution and smooth off the tail. Eta-learning does not need to have seen an extreme event to generate a statistically plausible one. The team trained on 25 years of hourly precipitation data, using the first six months of records as the training set, then prompted the system with questions of the form "what could a once-in-a-century storm look like in New York City". It returns thousands of possible realisations as maps, each carrying a storm's size, area of coverage, duration and rainfall intensity. New York City's highest recorded rainfall is 200 millimetres; the method generates plausible maps of a 300-millimetre storm. Why it matters: Flood defences, drainage capacity and evacuation plans are all sized against a design storm drawn from what has already happened, which is the wrong reference class in a shifting climate. What is newly true is that a peer-reviewed method now exists for producing a defensible tail event without waiting for one to occur, trained on ordinary precipitation records rather than a disaster archive. Sapsis frames the use case as testing a seawall against a surge beyond anything recorded, and Chang notes the approach is not weather-specific, extending to robotic navigation and market-crash dynamics. What is verifiably next is whether the generated scenarios hold up against events after the training window, and whether a hydraulic modelling group adopts eta-learning outputs as design inputs rather than as illustration.
Anthropic Puts $5 Million Into Wellbeing Evaluations It Does Not Control, Deadline September 21
Anthropic announced on August 25 a $5 million grant programme funding independent research into how AI systems affect user wellbeing. The stated deliverable is not papers but evaluations: open-source measurement tools any developer can run against a model. Grantees receive direct funding, model access and technical support, work fully independently, and publish as open source. Applications close September 21, with shortlisted applicants notified by October 5 and invited to submit full proposals. Eligibility is written around expertise rather than institution, naming clinicians, psychologists and methodologists explicitly. The criteria published by Anthropic's safeguards team ask for clinical experts on the team, real-world scenario testing rather than static benchmarks, and assessments covering both risks and the safeguards meant to catch them, with specific encouragement for proposals covering multi-turn conversations. The programme extends an earlier Anthropic initiative called Inviting Hard Questions. Why it matters: Wellbeing is the category where AI deployment into schools, benefits offices, crisis lines and clinics is expanding fastest and where the measurement layer is thinnest, because the standard benchmark format scores one response at a time while the harm accumulates across a conversation. What is newly true is that a funded, dated route now exists for clinicians and methodologists outside the AI industry to build that instrument and publish it openly, rather than each vendor grading its own homework in private. The independence condition is the load-bearing term and the thing the record will settle: the test is whether published grantee findings unflattering to the funder survive. The deadline is September 21, 2026, at anthropic.com.
Two Ex-DeepMind Researchers Launch Sampura Research in London With $11 Million for AI Oversight
Rishub Jain and Josh Jacob, both formerly of Google DeepMind, launched Sampura Research on August 25. It is a London-based nonprofit with $11 million in initial funding from Coefficient Giving, structured as $7 million for the first year and a further $4 million pledged. Jain and Jacob are joined by Alex Adams as founding team, with a plan to recruit at least six more researchers in London. The organisation's stated object is oversight infrastructure, and its unit of work is what it calls judges: assemblies of human evaluators, AI components, or hybrids of both, deployed across the AI lifecycle from training through evaluation to real-world deployment. The argument is that fully automated oversight and fully human oversight each miss a distinct class of failure, and that the hybrid is where the catchable problems live. Bloomberg reported the launch on August 25, alongside independent coverage from the Future of Life Institute and Crypto Briefing. Why it matters: This is a second, independently funded bet on the same missing layer as the Anthropic programme above, and the two are structured differently in a way worth watching: Anthropic is buying open-source evaluations from outside researchers against a deadline, while Sampura is a standing institution with a four-person founding team and a hiring plan. Together they put $16 million into evaluation infrastructure in a single week, a small sum against AI capability spending and a large one against what has previously been spent on judging outputs. What is verifiably next for Sampura is the London hiring round and its first published judge design.
Florida Launches CORE at FIU, Routing Eleven Universities and K-12 Into Emergency Management
Florida launched the Coalition for Operational Readiness and Education, CORE, at Florida International University on August 26. It is a workforce pipeline rather than a technology programme: the Florida Division of Emergency Management, the Florida Department of Education, FloridaCommerce, universities and private-sector partners building shared curriculum and training routes into emergency management jobs. FIU is the first campus. Ten further universities, including Florida State, the University of Florida and the University of South Florida, plan to adapt CORE into their curricula, and the programme is stated to extend into K-12 schools, state colleges and technical schools. CORE is not restricted to emergency management majors: engineering students and students in adjacent fields can take the coursework and carry those skills into their primary careers, a bet on distributed competence across a workforce rather than a specialist cadre. Officials did not publish an enrolment target, a budget figure or a start date for the K-12 extension at launch. Why it matters: Every early-warning system, hazard model and response platform in this briefing assumes a trained person on the receiving end, and that assumption is the least funded part of the stack; the Nepal assessment makes the same point from the failure side, that a warning nobody can act on is not a warning. What is newly true is that the state with the most hurricane exposure in the country has classified emergency management as a workforce-supply problem and committed eleven university curricula to it. The verifiable next steps are the enrolment numbers Florida has not yet published and whether the ten additional universities actually adopt the curriculum for the coming academic year.
Upcoming Events & Opportunities
Anthropic Wellbeing Research Grants (Funding)
- Amount: $5 million total programme; per-grant size not published. Includes model access and technical support alongside cash.
- Deadline: September 21, 2026. Shortlisted applicants notified by October 5, 2026 and invited to submit full proposals.
- Eligibility: independent research teams with relevant expertise, explicitly including clinicians, psychologists and methodologists. Outputs must be published as open-source evaluations, and grantees work independently of the funder.
- Apply: anthropic.com
- Africa Food Systems Forum, GoGettaz Agripreneur Prize Grand Finale, Kigali, Rwanda, September 2026. US$160,000 total across two $50,000 grand prizes and four $15,000 Impact Awards. Exact dates and entry deadline are not confirmed against the organiser programme page. Verify at gogettaz.africa.
- Cayuga County Nonprofit Summit (AI and leadership sessions), Cayuga County, New York, listed as October 8, 2026 per aggregator metadata and not verified against an organiser page. Registration details unconfirmed.
- Adaptation Fund Climate Innovation Accelerator (AFCIA) 2026, via WFP: surfaced August 25, deadline could not be confirmed against an Adaptation Fund or WFP page. Verify at adaptation-fund.org.
Active Disaster Monitoring (GDACS/OCHA)
- Nepal (Rasuwa district):** Bhote Koshi flash flood, onset August 26 around 09:00 local; 165 dead per NDRRMA as of August 27 and rising, 579 travellers unaccounted for per preliminary Nepal Tourism Board data, and 85 security personnel out of contact. Active search and rescue, 123 airlifted from Timure; affected areas declared disaster crisis zones by cabinet decision on August 26.
- Tropical Cyclone SAUDEL-26:** GDACS Orange notification issued August 18, still listed as active, with 11.482 million people exposed to Category 1 winds of 120 km/h or higher per the current GDACS record.
- Peru:** M6.7 earthquake 31 km NW of Aniso on August 20 per USGS; impact assessment not yet in the feed. Monitoring.
- Japan (Greater Tokyo):** M5.8 earthquake 4 km N of Toride on August 22 per USGS, a shallow event under a dense population centre; no major damage reported in the feed. A separate M6.0 struck 33 km SSW of Honcho on August 23. Monitoring.
- Timor-Leste:** M6.0 earthquake 215 km NNE of Lospalos on August 24 per USGS, offshore, with no significant population impact reported. Monitoring.
- Lao PDR:** floods ongoing since July 2026; ReliefWeb situation page last updated August 26, no consolidated casualty figure available in the feed.
- Note: only major or ongoing-major events are listed; GDACS Green alerts and an uninhabited-ocean M6.2 in the Scotia Sea are excluded per the major-only bar.
Sources: See individual stories above for full attribution.