Hey
Hope you had a great weekend wherever you are. Ours was a good one - visitors in town, a borrowed dog demanding attention, a pizza festival that absolutely delivered, and cocktails on the river to wind it all down on Sunday. Can't complain. 🍕
Anyway - big week in AI. Let's get into it:
The Broken Ladder: AI's Productivity Gains Are Appearing Exactly Where Entry-Level Jobs Are Disappearing | Ground Shifts 🌊
Stanford HAI’s annual AI Index lands this week and it is packed. US software developers aged 22–25 saw employment fall nearly 20% from 2024 - while their older colleagues' headcount grew. In the same period, AI agent task success on real-world benchmarks went from 12% to 66%. Productivity gains of 14–26% are showing up in customer support and software development - precisely the entry-level domains most exposed to automation. Executives surveyed expect planned headcount reductions to outpace recent cuts. The disruption is targeted, and by their own account, just beginning.
We’ve discussed this many times before but entry-level roles aren't just jobs. They're the developmental pathway - where junior work gets done, reps gets repped, judgment gets built through doing work badly before doing it well. Universities are certifying students for an entry point that's disappearing not as a destination but as a training ground. The standard response - 'focus on higher-order skills' - skips the question of where those skills come from without the lower-order work that builds them. What does a degree certify when the apprenticeship it assumed has been automated away?
Stanford's AI Index 2026: Experts and the Public Are Living in Different Realities | Trust Gap 📉
The same Stanford HAI's annual AI Index includes a finding that reframes the entire HE debate. On AI's impact on jobs, 73% of experts expect a positive effect. The public: 23%. A 50-point gap that repeats across every domain measured. Experts forecast AI assisting 80% of US work hours by 2030; the public estimates 10%. Not different opinions about the same future - incompatible accounts of what's already happening. The trust data compounds it: the US reports the lowest government confidence in AI regulation of any country surveyed. The EU is trusted more than either the US or China to manage the technology both are racing to dominate.
Universities prepare students for futures - but which one? The expert version, where AI is reshaping work at measurable speed? Or the public version, where most Americans expect fewer jobs? Those aren't competing predictions. They're different levels of exposure to what the technology already does. Educators who use AI daily and those who don't are teaching toward the same credentials while preparing students for different worlds. The 50-point chasm isn't a communication problem. It's the operating environment. Which future is your curriculum written for?
The Ruler Measuring Itself Just Built a Classroom: Khan, Stanford, and the Vendor Capture of Higher Education | Capture 🎓
The field's most credible annual report on AI cited Anthropic's own usage data as evidence of how students engage with AI - ~40% using Claude for creating, ~30% for analysing. Vendor-supplied data, in an independent report, measuring whether the vendor's product supports learning. The same report documents the Foundation Model Transparency Index dropping from 58 to 40 in a single year - the most capable models disclosing the least - while documented AI incidents rose from 233 to 362. The infrastructure for evaluating AI in education is being built by the people selling AI in education.
The Khan TED Institute, announced this week, makes that logic institutional. Khan, TED, and ETS launching an HE alternative - Google, Microsoft, McKinsey shaping competency signals, ETS deciding what counts as knowing something, no universities at the founding table. The same week Chalkbeat documented Khanmigo flopping - students didn't use it, teachers abandoned it, Khan's own chief learning officer: 'so far I am not seeing the revolution in education.' When adoption failed, Khanmigo didn't improve - it made itself impossible to ignore, becoming an always-on chatbot activated without the student's invitation. The response to a product students wouldn't seek out wasn't to fix it. It was to build the institution that makes the question irrelevant. The ruler measuring itself, then redrawing the map.
China's AI+Education Action Plan: The Most Comprehensive National Framework Anywhere — and What It Doesn't Say | Governance Surge 🌏
Five Chinese ministries co-signed an AI+Education Action Plan on April 2nd that makes every Western national strategy look like a discussion paper. AI becomes a compulsory public course across all universities, with textbooks authored by subject area and national competency frameworks for both students and teachers. Teacher qualification exams and accreditation will incorporate AI competency requirements - meaning you won't be certified to teach in China without demonstrating AI capability. Crucially, the plan mandates validation mechanisms for educational LLMs before deployment and safety auditing across the full AI lifecycle: algorithms, data, infrastructure, application systems. Pre-deployment validation as a legal requirement, in the same week Stanford's Foundation Model Transparency Index dropped from 58 to 40.
The gap the plan doesn't close question re. psychological and social dimensions - treated as downstream problems rather than upstream conditions. Governance architecture first, human readiness later - a familiar sequencing. Fengchun Miao at UNESCO, who helped shape the framework, responds that the human dimensions are designed as pre-conditions. Practitioners aren't convinced: policy without implementation bridges doesn't reach classrooms. China has the most coherent national AI education framework in the world. Whether coherent governance produces coherent practice is a different question - and not one the plan necessarily answers. Still - amazing stuff!
Stop Building Agents, Start Building Skills: The Unglamorous Future of Institutional AI | Capability Shift 🛠️
This week's been busy with frontier model announcements from Anthropic (Mythos, Opus 4.7, Claude Design, etc.) but if you haven't seen it, a fantastic talk from a couple Anthropic PMs a while back is worth your attention - and it isn't about capability. The thesis: agents are brilliant but lack expertise, and the solution isn't more powerful models, it's organised folders. Skills are collections of files - markdown, scripts, assets - that package procedural knowledge an agent loads at runtime when relevant. Built once, available across every agent interaction, version-controlled in Git. Five weeks after launch, thousands of skills exist across enterprises and research teams. The most interesting signal: non-technical staff in finance, legal, and recruiting are building them.
The standard institutional complaint - 'AI doesn't know our context, our voice, our standards' - is exactly the problem skills are designed to solve. A curriculum team's assessment principles. A writing centre's house style. A research office's ethics protocols. Package them once, and every subsequent agent interaction starts from institutional knowledge rather than rebuilding it from scratch. The technology is deliberately unsexy: just put stuff in a folder. The question is whether HE moves fast enough to shape what goes in that folder before vendors do it for them.
Stanford's annual reckoning arrived sharper than last year. The capability-governance gap is being measured with decreasing accuracy by instruments built by the people it's meant to hold accountable. China legislates comprehensively and leaves human readiness for later. Khan doesn't fix the failing product - he builds the institution that makes the question irrelevant. Entry-level developers lose jobs at nearly 20% while productivity gains appear exactly where they were standing.
The vendors aren't just in the classroom - they're writing the curriculum, measuring the outcomes, and replacing the accreditor. The folder is being filled. The question is who's doing it.
The Air Conditioning Question: What Happens When Thought Gets Cheap | Adjunct Intelligence 🎙️
Air conditioning consumes 10% of global electricity. Data centres: less than 1%. Air conditioning didn't just cool buildings - it made entire economies possible, turning cities that couldn't sustain knowledge work into global hubs. This week on Adjunct Intelligence, Dale and I ask what happens when thought gets that cheap. Not what jobs disappear - what becomes possible that wasn't before. Find us wherever you listen to podcasts or watch on YouTube.