AI

Grid operators in several regions have revised demand forecasts upward, citing datacentres. The forecasts vary widely and several previous revisions were later reduced.

AI will consume the grid. Energy demand is exploding without bound.

Demand forecasts created during investment booms have historically overshot. Efficiency per token has fallen sharply year over year.

Which forecasts are based on signed interconnection agreements versus speculative requests? How much capacity is double-counted across utilities? What does per-query energy cost actually trend toward?

Benchmark scores continue to rise, but an increasing share of gains comes from evaluation-specific tuning and prompt scaffolding. Independent replications often show smaller deltas than launch announcements.

AI capability is accelerating without limit. Each new model is a step change.

Progress may be real but uneven. Headline benchmarks may measure familiarity with test formats as much as general capability.

How much of reported gains survive contamination-free evaluation? Which capabilities transfer to unscaffolded real-world tasks? What would a saturated benchmark regime look like?

Code-generation tools are widely adopted, yet job postings for senior engineers remain steady while junior postings decline. The work is shifting toward review, integration and specification.

Software engineering as a profession is ending. AI writes the code now.

Tasks are being replaced faster than roles. The bottleneck may move from writing code to deciding what code should exist.

What happens to the senior pipeline if junior roles disappear? Does verification effort grow faster than generation savings? Which domains resist specification by prompt?

The measured gap between open-weight and frontier proprietary models has narrowed on public evaluations, while the gap in deployed product quality is harder to measure and rarely reported.

Open models have caught up. Paying for frontier access is unnecessary.

Parity on public tests may coexist with meaningful gaps in reliability, tooling and long-tail behaviour that only show up in production.

How large is the gap on private, uncontaminated tasks? Who funds open-weight training runs, and why? Does the gap close, oscillate, or widen with each generation?

Vendors report rapid enterprise agent adoption. Public case studies cluster around pilots and internal tools; fully autonomous customer-facing deployments remain rare in regulated industries.

Autonomous agents are already running core business processes everywhere.

Adoption may be wide but shallow. The distance between a demo, a pilot, and a load-bearing system is consistently underreported.

What fraction of announced deployments survive twelve months? Where does liability sit when an agent acts? What error rate is acceptable per domain?

Datacentre capital expenditure has grown to a visible share of GDP growth in several economies. Utilisation figures for that capacity are rarely disclosed.

Whoever controls compute controls the future. Build at any cost.

Oil is consumed once; compute depreciates whether used or not. Overbuilding is a known failure mode of infrastructure races.

What is actual utilisation of installed AI capacity? How fast does hardware depreciate against algorithmic efficiency gains? Who absorbs losses if demand projections miss?