🟣 EV Daily: The weird nostalgia for coal

Five things to know today + Future of work

Lead story: The weird nostalgia for coal

The US Department of Energy warns that by 2030, the risk of blackouts in the US could increase a hundredfold as power demand grows partly because of AI. It quantifies an urgent need for roughly 100 GW of new peak supply by 2030. Then the report becomes ideological pushing the administration’s recidivist love of thermal coal.

As we (and others) have long-argued, resilient grids are possible in the short term without coal or oil, and in the medium term without natural gas. CrusoeAI recently demonstrated a data center that ran around the clock on solar power and batteries, showing that continuous renewable supply is technically feasible. Carbon Briefing found that near-constant solar power can be delivered in many cities for about $100 per megawatt-hour—cheaper than new coal plants or nuclear power—and those costs are still falling. Time to build is also double the speed for renewables than fossil plants. Although renewables alone cannot meet one hundred percent of demand, dispatchable gas plays a necessary role; framing this as a failure of renewables misrepresents the true challenge of grid flexibility and investment.

Ultimately, the US remains reliant on 19th-century fuels at a time when the rest of the world is building energy systems for the 21st century. Policy should be focused on building a balanced mix of clean power and dispatchable resources, rather than nostalgia for coal. US Department of Energy #Energy

Key signals, quick scan

A 30-second scan of four secondary signals that hint at where the curve is bending.

• Google-backed Isomorphic Labs is gearing up to start human trials of a drug discovered by AlphaFold3. It will be the first time DeepMind’s breakthrough drug-discovery system is applied in actual patients. While AI is accelerating early discovery, Phase I–III trials still typically take six or more years, so the real payoff from today’s breakthroughs will likely emerge next decade. Fast Company #AIapplications

• An impostor posing as Secretary of State Marco Rubio used AI-generated voice and text to contact at least five senior officials via Signal and SMS. It took as little as 15 to 20 seconds of audio to clone Rubio’s voice. Washington Post #Society

• Global startup funding hit $91 billion in Q2 2025—down 20% from Q1’s spike but still up 11% year-on-year. Nearly $40 billion went to AI, and just 16 mega-rounds soaked up a third of all capital. VC is rebounding, but the money is piling into a few big, AI-heavy bets. Crunchbase #Economics

• Meta has acquired a €3 billion stake in EssilorLuxottica, the maker of Ray-Ban. Mark Zuckerberg appears to be doubling down on his belief that the future AI form factor will be spectacles. Bloomberg #Devices

Future of work: Expertise matters

David Autor—one of the world’s most-cited labour economists at MIT—and Neil Thompson, director of MIT’s FutureTech group, analyse how different kinds of task automation ripple through wages and head-counts.

Automating high-skill tasks can democratise access to well-paid jobs, but at the cost of compressing wage differentials and eroding expert rents. But automating routine tasks, increases wages for specialists while reducing employment. A one-standard-deviation decline in the complexity of an occupations tasks¹ — equivalent to the decline in expertise required by a telephone operator from 1980 to 2018 — is linked to an 18 % wage decline and a 40 % increase in employment per decade; the mirror image holds when expertise rises.² They back this with 40 years of task data across 303 US occupations (1980–2018).

Firms deciding where to deploy AI need to choose: widen the talent pool (automate the hard bits) or deepen specialist capability (automate the drudge). Education and re-training must target the remaining task mix, not broad occupational labels—which speaks to a need for greater coordination between firms and educational systems.

The real question is not “will AI kill this job” but “which tasks will AI impact and how will that move the expertise bar?”

---

Was this email useful?

• 👍 That helped me. Thanks

• 😐 Meh - was ok.

• 👎 Not interesting to me.

Thanks! Want to share this with someone like you? Forward the email.

---

Want to opt out of the daily emails? Follow these steps.

1. Autor and Thompson quantify expertise using a Standard Frequency Index (SFI). Each task’s written description is scored by combining the rarity of its words (how infrequently they appear across all task descriptions) with their predictability in context (how well language models anticipate them). High-SFI tasks, such as “design and architect a distributed database,” demand specialist knowledge, while low-SFI tasks, such as “enter customer data,” are routine and generic.

2. Although the largest increase in an occupations expertise was only 0.5 standard deviations.