The archive

Latest articles

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

SentinelBench related success data showing baseline agents at 5.6 percent on 1 hour and 2 hour monitoring tasks, while SentinelStep reached 33.3 percent at 1 hour and 38.9 percent at 2 hours.

SentinelBench tests whether AI agents can wait well

SentinelBench is a 100-task benchmark for monitoring agents. Use it to test patience, latency and tool spend before you ship.

Lars Cornelissen · Jun 6, 2026

Donut chart of where AI compute energy goes, with inference at about 85 percent in red and training at about 15 percent in blue.

What an AI prompt really costs in energy, water and CO2

A typical AI text prompt uses roughly 0.24 to 0.34 watt-hours and a fraction of a milliliter of water, not a bottle per email. Training is huge but rare, and inference is what adds up.

Lars Cornelissen · Jun 5, 2026

AI persuasion rates on Reddit: personalization 18.0 percent, generic 16.8 percent, community aligned 9.0 percent, and human baseline 2.7 percent.

AI persuasion got its Reddit field test, and it was ugly

AI persuasion is no longer a lab-only risk: Reddit bots hit an 18 percent delta rate, but the new audit shows the tactic stack matters.

Lars Cornelissen · Jun 5, 2026

EVA-Bench Data 2.0 scenario split: Healthcare HRSD has 83 scenarios, Enterprise ITSM has 80 scenarios, and Airline CSM has 50 scenarios.

EVA-Bench Data 2.0 makes voice agents measurable

EVA-Bench Data 2.0 is a 213 scenario dataset for testing enterprise voice agents across airline, IT, and healthcare HR workflows.

Lars Cornelissen · Jun 5, 2026

AI agent security benchmark comparison showing Mythos Preview at 83.1 percent versus Opus 4.6 at 66.6 percent on CyberGym, and 77.8 percent versus 53.4 percent on SWE-bench Pro.

AI agent security failed in Meta’s support desk test

AI agent security is privileged access control for LLMs. Meta’s Instagram hack shows one support bot can turn account recovery into takeover.

Lars Cornelissen · Jun 5, 2026

Area chart of Palantir annual revenue climbing from 1.1 billion dollars in 2020 to 4.5 billion dollars in 2025

Why Palantir's data software is worth $400 billion

Palantir builds data integration software that fuses an organization's siloed records into one model its AI can act on. Here is why that business now carries a market value near $400 billion.

Lars Cornelissen · Jun 5, 2026

Donut chart splitting seven TeamPCP waves by period: two in 2025, one in early 2026, and four in the May to June 2026 sprint.

TeamPCP: the crew behind the npm worm hitting Red Hat

TeamPCP is the cybercrime crew behind the Shai-Hulud npm worm. It open-sourced the malware in May 2026, then poisoned Red Hat's packages and blurred the question of who to blame.

Lars Cornelissen · Jun 5, 2026

Bar chart of FIFA World Cup revenue per edition rising from $3.3 billion in 2006 to a projected $8.9 billion in 2026, with Qatar 2022 at $7 billion.

World Cup 2026 tickets: dynamic pricing takes over

Dynamic pricing means FIFA moves World Cup 2026 ticket prices with demand, like an airline. Seats run from $60 to $6,730, helping push the tournament toward a record $8.9 billion.

Lars Cornelissen · Jun 5, 2026

Audit then score benchmark accuracy rises from 60.8 percent in Round 0 to 80.4 percent in Round 1, 85.3 percent in Round 2, and 90.9 percent in Round 3.

Audit then score turns AI ground truth into a process

Audit then score is a benchmark protocol that revises labels before grading. Amazon says it lifted expert accuracy from 60.8% to 90.9%.

Lars Cornelissen · Jun 4, 2026

Donut visualization for generalist agents in data curation showing a scaffolded agent using 10 percent of the published baseline data budget and avoiding 90 percent.

Generalist agents hit data curation’s hard budget wall

Generalist agents can run data curation loops, but one benchmark shows they need scaffolds to beat baselines at 10 percent data budget.

Lars Cornelissen · Jun 4, 2026

Virtual power plants capacity scale for AI data centers showing Texas large load threshold at 75 megawatts, Google Voltus VPP at 100 megawatts, and Duke flexible load estimate at 100,000 megawatts.

Virtual power plants meet AI’s data center power wall

Virtual power plants are flexible loads coordinated as capacity. Google’s 100 MW Voltus deal tests whether homes will flex for AI.

Lars Cornelissen · Jun 4, 2026

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