The people who suffer most in a heatwave, and never say so 🌿

Riding in an autonomous vehicle involves a small, constant act of faith. The car brakes and you do not know why. Over enough trips, people build a private theory of how the car thinks, and that theory is often wrong in ways that matter at exactly the wrong moment.

MIT researchers, working with the autonomous vehicle company Motional, published a system in Nature on September 2 that closes some of that gap. It is called the Concept-Wrapper Network, or CW-Net, and it sits on top of the car's existing planning model and translates what that model is doing into phrases a person can hold in their head. Things like "approaching stopped vehicle" or "close to cyclist." Crucially, it describes the reasoning the car is actually using rather than inventing a plausible-sounding story after the fact.

They then tested whether the explanations changed anything. On a private track with real safety drivers, and in a larger simulation study with non-experts, people who saw the explanations became measurably better at predicting what the vehicle would do next, and the biggest gains came in the moments where the car's behaviour had surprised them. Driving performance stayed within one percent of the unmodified system.

Most of the debate about trusting AI is about whether the system is right. This is about whether the human next to it can anticipate when it will be wrong, which is a different and more useful question.

What we are still uncertain about: a private track and a simulator are not a city street, and the study measures prediction, not safety outcomes. It is also possible that better explanations make people trust the car more than the evidence warrants. That is worth watching.

Sources: MIT News, Nature

A UC Irvine-led team applied language analysis to 24,598 heat-related California social media posts from 2016 to 2022. People living in RVs, vans and boats showed signs of greater difficulty while being far less likely to mention any way of coping. Heat kills more Americans than any other weather event, and the most exposed rarely appear in official data.

Source: UC Irvine News

A policy brief from the UN University Institute for Water, Environment and Health, released September 4, argues that governments are approving electricity infrastructure with 15 to 20-year lifespans against historical weather records that no longer hold. It proposes a narrow, auditable class of model for capacity planning approvals, plus hard caps on the energy AI data centres consume.

Source: TechXplore, Eurasia Review

In an essay on Gates Notes, Gates argued the tax code rewards automation over hiring, since a firm writes off a robot immediately but pays payroll tax on a person. He also floated a "Human Reserved" category barring AI from certain roles, such as delivering a terminal diagnosis. Both proposals would cut into the profits of the largest AI labs.

Source: TechCrunch

Nine in ten executives say three years of AI changed nothing about their headcount

A National Bureau of Economic Research survey found that more than 90 percent of executives reported no effect of AI on employment at their own firm across three years, and 89 percent reported no measurable productivity effect. Separately, University of Pittsburgh professor Mark Ma found that the average stock market reaction to an AI-justified layoff announcement was close to zero, and that employee sentiment toward AI in Glassdoor reviews tracked firm productivity.

Read those two findings together and something uncomfortable falls out. If AI-branded job cuts do not move the share price, then the AI framing is not buying the company anything financially. It is a story companies tell about a decision they were going to make anyway.

Our PM + Risk Manager lens

This is the deployment gap, measured. Capability has run far ahead of integration, and integration is where the value actually lives. Most enterprise AI has been bolted onto workflows that were designed for a pre-AI world, which produces a tool people use occasionally rather than a process that runs differently. The lesson for anyone shipping AI internally is that adoption metrics are vanity unless the underlying work has been redesigned around the new capability. Ma's Glassdoor finding is the useful signal here: firms where employees feel positively about AI are the firms where productivity actually moved. Sentiment is tracking real integration, not enthusiasm.

The exposure runs in both directions. Firms that cut headcount on an AI thesis and then report no productivity gain have taken on real operational risk for a benefit they cannot evidence, which is a governance problem the moment anyone asks for the business case. And a company that publicly attributes layoffs to AI while the survey data says AI did nothing is making a claim to investors and staff it may not be able to defend. Meanwhile, the opposite risk is quietly building. Three years of no measured effect is not proof of no future effect, and firms reading this as permission to wait are making a bet of their own.

The next 12–24 months

The honest position is that we do not yet know whether this is a plateau or a lag. Every general-purpose technology in the historical record has shown a gap of years between arrival and measured productivity. What will separate the two explanations is whether firms that genuinely rebuilt processes start pulling away from firms that only bought licences. Watch for the spread between them to widen. If it does not by late 2027, the sceptics have a much stronger case than they do today.

Source: Futurism

Google DeepMind and Google Research released WeatherNext 3 on September 3. It produces hourly global forecasts at up to 5-kilometre resolution for surface variables like temperature and moisture, and it generates predictions every hour from live satellite data rather than waiting on government datasets that refresh every six hours. Google reports up to 50 percent more accurate precipitation forecasts a day or more out, with the largest gains in regions where forecasting has historically been worst. It is already feeding Search, Maps and the Gemini app, and researchers can query the underlying data through BigQuery and Earth Engine or bulk download it from Google Cloud Storage.

→ Read more: Quartz

If an AI system could explain its reasoning to you in plain language, would you trust it more, or would you just find new things to worry about?

Hit reply. We read every response.