Dispatches, Tarry Singh

Three press releases in three months told the industrial-AI story more clearly than any analyst report.

A new MIT preprint draws the line between search and discovery in category theory, and one empty set turns out to be the formal wall most agent stacks will never cross.

The demos are real and the peer review is real.

Two energy curves dominate 2026.

Two regulators blinked on the AI question in a single five-month window.

Inside thirteen days Brussels told teachers to adopt AI, delayed the rules governing the riskiest classroom uses, and only then began explaining what those rules mean.

America is writing AI-competency requirements into degree audits without a defensible idea of what competent looks like.

Brussels filed two plans on the same afternoon that quietly cancel each other out: triple the data centres, and please switch off the dishwasher.

Brussels promised to triple Europe's datacentres in five-to-seven years.

Europe builds compute it must import and under-builds the people it already has.

The most expensive line item in your AI budget for the next two years is the one your finance team has not yet named.

The AI story most boards are still being told is a productivity story.

$1 trillion of compute spend is the most concrete thing in technology right now.

Consensus for 2026 hyperscaler capex has roughly doubled in twelve months.

Only 13% of employees can actually work with the agents their firms are deploying.

Profitable firms are cutting staff to fund AI capex they cannot yet show pays off, and freezing the junior hiring that manufactures expertise.

The companies buying the most copilot seats are often the ones whose people are getting faster at being shallow.

Intelligence and energy are deflating toward zero, and automation is crossing from the screen into the world: the firm, the laboratory, the factory and the launch pad remade at once.

The owners are directionally right that they can run with far fewer people.

Three credible measurement traditions, applied to the same workforce in 2026, point in three different directions.

The full argument behind the comprehension gap.

Brussels made digital sovereignty a procurement rule in the last week of May.

A scroll-driven feature in four movements: how system complexity outran human comprehension by four orders of magnitude, and why you close the gap by reshaping the substrate, not enlarging the engineer.

Three weeks of racks, export approvals, and merger announcements turned the sovereignty debate from a slide into a bill of materials.

Any 2026 production agent stack without the three-layer caching pattern (engine prefix cache, API prompt cache, gateway semantic cache) is carrying a 30–60% avoidable inference bill.

1,059 hedge funds have made a record, leveraged, one-directional bet on AI semiconductors.

$700 billion in quarterly capex, and the useful question is which of three walls the number is actually paying down.

The cognitive layer of industrial AI (copilots, agentic optimisation, the operating-system metaphor) is real and mostly in pilot.

The fork between copilot and agent is the right framing for the next two years of industrial AI.

AI is the tool that could compress seismic interpretation from months to days.

Every enterprise is a graph of algorithms, and the bottleneck is the edges between nodes, not the nodes themselves.

Three jurisdictions, zero frameworks designed for the agents already in production.

AI literacy is being legislated as a curriculum addition when the actual problem lives in the exam hall.

The EU made AI literacy a legal duty eighteen months before the educational infrastructure it funded could plausibly deliver at scale.

The regulatory squeeze and the energy squeeze are the same squeeze.

Three rounds of layoffs, each with a grander narrative.

Karpathy's first assignment at Anthropic is to point the current model at the job of building the next one.

Monitoring measures a worker so a manager can manage them.

Four data sets, three methodologies, and no one who matters can yet give you a single defensible figure for AI's effect on knowledge-work productivity.

Geopatriation, the localisation of AI workloads under sovereign pressure, is the dominant pattern in 2026.

The failure was never the bad command.

MCP and A2A now sit under the Linux Foundation, adoption is measured in billions of tool calls, and the standards war is over.

Mozilla's Mythos shipped 423 security fixes in a single month after fifteen months of human-pace trickle.

The AI arc bent in the last four weeks, from four shifts at once rather than one launch: the model race went plural, NVIDIA's Rubin redefined silicon economics, sovereign AI stopped being a slogan, and the agentic enterprise got real numbers.

After thirty years of building, I wanted a home that reads less like a portfolio and more like a working room.

Plate I of the Synaptic Cartography series went through eleven drafts before it stopped feeling like a poster and started feeling like a window.