Behind the scenes at the AI frontier
Stan van Baarsen
Stan van Baarsen
AI PlanCo-author of the Nationaal AI Deltaplan, the national AI strategy for the Netherlands, 2025
Co-author of a scenario on Europe’s dependence on foreign AI, 2026
Now at Cursive, a new AI lab in London founded by former Google DeepMind researchers
Part 1 of 4
Chat
Reasoning
Multimodal
Code
Terminal and tools
Browser and computer use
AI research
Models that help build the next model
Spatial reasoning
Next: general-purpose robots
Science
Maths, chemistry, biology, health care
Chip design
More efficient chips for the next models
Professional work
Finance, law, office software
What is the most accurate way to find out what the labs are working on today?
Look at what data they are buying.
Google DeepMind, July 2026: one model controls a whole humanoid
Physical Intelligence, November 2025: robots that learn from experience
Figure, 2026: in a pilot on BMW’s factory floor
Genie 3 turns a prompt into a world you can move through
An AI agent practising tasks inside one of those worlds
Waymo, built on Genie 3: an elephant or a tornado on the road, on demand
Part 2 of 4
Gold-medal score at the International Mathematical Olympiad
Erdős’s 1946 unit-distance conjecture disproved
Claude writes a complete machine-checked proof of Fermat’s Last Theorem
OpenAI claims a proof for a version of Navier–Stokes, a Millennium Prize problem
“At this point nothing would surprise me.”
Thomas Bloom, mathematician, Manchester
“The shock is immense. I am paralysed.”
Hugo Duminil-Copin, Fields medallist
“Please leave me alone for an hour, I need to digest this thing.”
José Penadés, Imperial College
Amsterdam
Designs better proteins with AI, for Bayer, Lundbeck and others
Cambridge and Amsterdam
Designs new materials with AI; raised $450 million in July
London
DeepMind’s drug-design spin-off; raised $2.1 billion, first patients expected by the end of 2026
Law
Software engineering
Finance
Science
Customer service
Part 3 of 4
Law
Software engineering
Finance
Science
Customer service
Top models on the Artificial Analysis Intelligence Index
Major model releases. Since September, a new frontier model has come out more than once a week.
Yann LeCun. World models. $1.03 billion
Mira Murati. Customisable AI. $2 billion
David Silver. Learning from experience. $1.1 billion
Pim de Witte. Agents trained on game clips. Valued at $6.2 billion
Yann LeCun. World models. $1.03 billion
Mira Murati. Customisable AI. $2 billion
David Silver. Learning from experience. $1.1 billion
Pim de Witte. Agents trained on game clips. Valued at $6.2 billion
Ex-DeepMind researchers. Custom AI models for specific verticals that keep learning in deployment
Part 4 of 4
AGI-pilled
Timelines
GPU-poor
Post-AGI
NGMI
Artisanal
With AI becoming so capable, we’re actually being asked to redefine what makes humans human.
“We could summarize this as a country of geniuses in a datacenter.”
“…compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years.”
“…likely to vastly improve treatments for, or even cure, most mental illness…”
“…a dream scenario … would be 20% annual GDP growth rate in the developing world…”
“I think many will be literally moved to tears by it.”
mail@stanvanbaarsen.nl linkedin.com/in/stanvanbaarsen
Data centres and chips, planned for 2026
Amazon
Alphabet
Microsoft
Meta
Raised by the labs in 2026
OpenAI, valued at $852bn
Anthropic, valued at $965bn
→
↑
↓
←
Cursor retrains its coding model on live feedback from users, and ships an improved version as often as every five hours, allowing it to learn from how people use it right now.
Practice worlds where a model attempts a task and gets scored
Find the cheapest two-bedroom listing for these dates
Compute the five-year growth rate
Build a pivot table of revenue by region
Fix the bug so the tests pass
What labs pay for them
$200 to $2,000
about $20,000
about $300,000
~$1 billion at Anthropic alone
Claude writes more than 80% of the code merged at Anthropic
OpenAI says it has reached its ‘automated AI research intern’
OpenAI’s target for a fully automated AI researcher
The gap is now about five weeks, the smallest yet
Most of the best open models are now Chinese: Kimi, Qwen, DeepSeek, Zhipu
Meta has gone closed for its frontier models
OpenAI has released no new open model since August 2025
You can run them on your own servers, and keep your data
“I think the next tier of risk is actually AI companies themselves.”
Dario Amodei, CEO of Anthropic, January 2026
The hopeful version: most useful work runs on models that are good enough and cheap, and the frontier labs keep what only the frontier can do.
