The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

Multi-agent debate effects in data cleaning: generation down 1.6 to 15.5 points, detection up 27.4 points.

Multi-agent debate needs a boring data-cleaning cop

Multi-agent debate hurt generation by up to 15.5 points in data cleaning, but a grounded critic rescued detection and repair.

Lars Cornelissen · Jun 3, 2026

Google AI opt out chart showing lower click rates when AI summaries appear in search results

Google AI opt out gives publishers a real lever now

Google AI opt out rules in the UK give publishers control over AI Search, but the hard choice is whether to trade traffic for leverage.

Lars Cornelissen · Jun 3, 2026

Unit distance proof comparison of known asymptotic exponents: linear grid at 1, Sawin lower bound at 1.014, upper bound at 1.333

Research

Unit distance proof moves AI past clever math demos

The unit distance proof gives OpenAI an AI math win: n^1.014 unit pairs, with humans still doing the verification work.

Lars Cornelissen · Jun 3, 2026

Bar chart of the most common developer frustrations with AI coding tools, led by output that is almost right but not quite

Vibe coding is loud online and rare in real codebases

Vibe coding dominates AI discourse, but most professional developers still avoid it. The 2025 survey data shows why: the output is almost right too often.

Lars Cornelissen · Jun 2, 2026

Line chart of model capability over time bending upward after 2024, showing acceleration rather than a plateau

Research

AI capability stopped slowing down. It sped up after 2024

Model capability is improving about 15.5 ECI a year and the rate rose after early 2024. The expected plateau never arrived, which complicates every roadmap built around one.

Lars Cornelissen · May 31, 2026

Line chart of cumulative AI compute stock growing about 3.4 times per year from 2022 to 2025

Business

The bill for frontier AI is now measured in gigawatts

Frontier AI spending has shifted from clever algorithms to power and concrete. A single gigawatt data center now costs about 30 billion dollars to build.

Lars Cornelissen · May 31, 2026

Bar chart of how often developers use AI agents at work, with the largest group reporting no plans to adopt them

Engineering

Your agents are working too hard

AI agents are not yet mainstream, and the developers who use them report personal speed but little team-wide gain. The fix is teaching agents when to stop.

Lars Cornelissen · May 30, 2026

Heatmap grid of model performance across task categories, with brighter cells marking stronger relative scores

Research

The model leaderboard is tightening to a photo finish

The performance gap between the best AI models and the rest is collapsing. Aggregate leaderboard scores now hide more than they reveal about real strengths.

Lars Cornelissen · May 28, 2026

Bar chart of AI chip performance per dollar rising across four generations from the P100 baseline to 24x at the GB300

Engineering

AI chips give 24x more per dollar, if you can afford the sticker

AI chip performance per dollar improves about 37 percent a year, yet each new flagship costs more upfront. The GB300 delivers 24 times the value of a P100 at nine times the price.

Lars Cornelissen · May 28, 2026

Bar chart of the share of organisations using AI rising from about 50 percent in 2022 to 78 percent in 2024

Business

Everyone automated everything. Then they hired more people.

Organisational AI use jumped to 78 percent in 2024, yet demand for human judgment rose alongside it. The productivity paradox is back in a new form.

Lars Cornelissen · May 26, 2026

Bar chart of global GPU cluster performance share by country, with the United States dominating at about 75 percent

Data

Three-quarters of the world's AI compute sits in one country

The United States holds about 75 percent of global GPU cluster performance. That concentration shapes pricing, latency, and policy for everyone building elsewhere.

Lars Cornelissen · May 25, 2026

Scatter plot of model capability against training compute on a log scale with a fitted regression line

Research

The scaling curve nobody wants to extrapolate

Training compute for frontier models has grown about 5 times a year since 2020. The scatter looks clean, but every data point sits to the left of the real question.

Lars Cornelissen · May 22, 2026