Research
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 is an enterprise data engineer and the lead writer at Data Today. He builds cloud data platforms at Alliander on Snowflake, Matillion, Python, and AWS, and runs the data and AI studio Datastudy. He turns primary AI and data sources into charts and operator-grade conclusions for the people who build and ship with the technology.
Articles by Lars Cornelissen
AI
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.
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.
AI
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.
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.
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.
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.
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.
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.
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.
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.
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.
Research
Training costs rise 3.5x a year. Efficiency is the only brake
Frontier training costs climb about 3.5 times a year while algorithms get 3 times more efficient. The two trends are racing, and the gap decides who can still compete.
Opinion
How cheap inference made the one-person studio viable
Building a one-person data studio used to be impossible on the economics alone. Inference at a fixed quality level fell roughly 280 times in two years, and that changed the math.
Business
Most agentic AI projects will be cancelled before they ship
Gartner expects over 40 percent of agentic AI projects to be cancelled by end of 2027. The drop-off from demo to production is where the budgets quietly die.
Business
The gigawatt build-out is the real AI race now
The largest AI data center already rivals 700,000 H100 chips, and a 5-million-equivalent campus is due by 2027. Power and concrete, not chips, set the new pace.
Data
Open-weight models closed the gap to 1.7 points
Open-weight models went from a budget compromise to a default choice. On some benchmarks the gap to the best closed models shrank from 8 points to 1.7 in a year.
AI
Last year's frontier model now runs on a laptop
Frontier AI capability reaches consumer hardware in about eight months. The shrinking gap turns today's hosted-only features into tomorrow's on-device default.
Engineering
Context windows grew 30x a year. Your retrieval stack noticed
LLM context windows have expanded about 30 times a year since 2023, from a few thousand tokens to over a million. The change quietly rewrites how RAG systems should be built.
Engineering
The world's AI compute now doubles every seven months
The installed stock of AI chips is growing 3.4 times a year, doubling every seven months. The capacity question has quietly shifted from chips to power and buildings.
Engineering
Inference got cheap, but not for the work that pays
LLM inference costs fell between 9 and 900 times a year depending on the task. The cheapest gains landed on easy work, while frontier reasoning barely moved.