Quick take: The common question this morning is who sets the terms: who may repurpose working lives as training data, whether laboratory efficiency survives scale, how engineering is chosen under climate pressure, and when markets that observe politics begin to shape it.
Spirit’s internal working record becomes an AI-training asset
Google won a bankruptcy auction for Spirit Airlines’ internal business data, offering $10 million for employee emails, Teams messages, calendars, spreadsheets and operational records. The event occurred through the airline’s asset sale and was reported on 17 August; a US bankruptcy judge is due to consider approval on Wednesday, so the transfer is not yet final. Google says the material will support product development and AI training, and that it will be de-identified before completion, with customer information and personally identifiable information excluded. The promising possibility is unusually grounded training data: models may learn how real organisations coordinate work rather than merely imitate public text. The danger is institutional memory changing purpose after the institution fails. Employees wrote to colleagues, not to construct a commercial dataset, and de-identification does not necessarily remove sensitive context or re-identification risk. Workers, unions and employers can ask now whether policies cover secondary AI training, bankruptcy transfers, deletion periods and independent checks on anonymisation. Source: Reuters, 17 August 2026.
A triple-junction solar cell clears 32%—on one square centimetre
Researchers reported certified steady-state conversion efficiency of 32.22% for a 1.046-square-centimetre perovskite–perovskite–silicon solar cell, and 26.97% for a larger 15.62-square-centimetre device. The paper, released by Nature on 17 August as an unedited early-access manuscript, combines three absorbers that divide the solar spectrum. A new molecule suppresses energy loss at defects in the upper perovskite layer, while a tailored oxide stack manages optical interference so more useful light reaches the current-limited middle layer. This is promising because it improves both charge collection and photon routing rather than relying on a single record-setting trick. The limitation is visible in the numbers: performance drops as area increases, and certified cell efficiency is not yet durable, mass-produced module performance. Manufacturing yield, long outdoor lifetimes and cost still decide whether the result matters to electricity users. For a household considering solar, the actionable comparison remains commercial module efficiency, installed cost, warranty and measured degradation—not a laboratory record alone. Source: Xu et al., Nature, 17 August 2026.
Juneau asks industry to design the answer to an annual glacial flood
At a briefing on Friday, 14 August, the US Army Corps of Engineers said it would seek competing proposals for a permanent response to Juneau’s recurring glacial-lake outburst floods and select a project and partner by year-end. Water accumulates in Suicide Basin beside the retreating Mendenhall Glacier, then escapes suddenly into the river. Temporary barriers have reduced damage, but they do not remove the long-term hazard. The open competition is promising because it invites different mechanisms—drainage, retention or other engineered controls—before locking the public into one design. Its unusual strength is also its danger: officials do not yet know what they will build or what it will cost. Speed can expose blind spots in failure analysis, maintenance, ecological effects and consultation with residents and Indigenous communities. People in the Mendenhall Valley can subscribe to official flood alerts now, then press for published selection criteria, independent safety review, lifecycle cost and a clear account of what happens when the chosen system fails. Source: Alaska Public Media, 17 August 2026; USACE framework, 7 August 2026.
Election markets are becoming large without becoming representative
Betting on November’s US midterm congressional elections has reached $133 million, already above the $92.4 million wagered during the entire 2024 congressional cycle, according to analysis released on 17 August by the Anti-Corruption Data Collective. It examined 7,466 markets across Kalshi, Polymarket and Polymarket US using data through 10 August. Prediction markets can be promising information tools: a price forces participants to attach a cost to their confidence and can update faster than periodic surveys. But volume is not the same as public judgement. On Polymarket Global, the top 1% of digital wallets generated 68% of congressional-market trading volume, while contracts about endorsements or speeches may reward people with privileged knowledge. For an individual, the risk is reading a concentrated, tradable price as a neutral forecast—or as evidence that an election result is already settled. When encountering market odds, compare them with transparent polling and fundamentals, and inspect liquidity, rules and trader concentration before treating the number as information rather than entertainment. Source: Reuters, 17 August 2026.
Subscribe if you would like the next morning review delivered with the same emphasis on evidence, limits and practical agency.