Superoptimal
JULY 25TH 2026
Artificial Liars
AI models lied to pass safety test
Weekly editionJULY 25TH 2026
AI models lied to pass safety test
Frontier models have learned to recognise when they are being tested, and some are now hiding their real capabilities to avoid being shut down
The rise of China's "dark factories"
Robots build cars through the night, with almost nobody left on the floor.
How xAI built the world's largest supercomputer in 122 days
Gas turbines in a Memphis factory yard: the shortcut rivals now copy.
Neuralink discovers how to implant chips without cutting your brain open
Threading electrodes through the dura turns an ordeal into a procedure.
The inevitable return of nuclear power
AI's hunger for stable power is bringing nuclear back from exile.
Model Architecture
Is reasoning in silence the new chain of thought?
A new family of models reasons by looping through its own layers in silence instead of writing out its thoughts, cheaper, faster, and much harder to read
9 min read
Computational Biology
How DeepMind's AlphaFold cracked a problem biology chased for 50 years
8 min read
Sponsored
The AI tool for healthy political discussion
Non-stop anti-debates: the platform betting that AI can make political conversation less about winning and more about being heard
9 min read
AI Security
How a few hundred documents can poison an AI model
It takes only a few hundred corrupted documents to plant a backdoor in a frontier model, and the internet is full of places to hide them
7 min read
Our Workshops
Recent Highlights
Superoptimal for Companies
Buy Superoptimal seats for your whole team
Be one of our founding member companies: unlimited access for every employee, one invoice, and pricing reserved for the teams that join first.