Research
Deep dives into the topics that matter. AI, automation, engineering, and the things most people get wrong about all three.
February 25, 2026
Sweden's AI Strategy Is a PDF, Not a Plan — And the Clock Is Running Out
Sweden's national AI strategy, critiqued by Arash Gilan in Breakit, suffers from three fatal flaws: a budget smaller than a Serie B round, zero mention of autonomous AI agents, and a complete absence of urgency. The strategy was written by people who have never shipped AI in production, and it shows. While Sweden has world-class AI talent and digital infrastructure, its policy apparatus is operating on government time while the technology advances on GPU time — a mismatch that risks turning one of Europe's most digitally advanced nations into an AI laggard.
February 23, 2026
When Brute Compute Meets Black Art: GPU-Powered Antenna Design Without the Intuition
A maker with no antenna design experience repurposed a GPU-accelerated FDTD photonic nanostructure simulator into a brute-force RF antenna optimizer, using LLM-generated code and openEMS validation. The resulting designs are geometrically strange and underperform expert work, but critically match their simulated predictions — demonstrating that the simulation-to-reality pipeline is sound. The real story isn't AI replacing RF engineers; it's that a single non-expert can now assemble a computational design pipeline using open-source solvers, consumer GPUs, and LLM coding assistants that produces physically valid (if mediocre) results. This pattern — brute-force search plus cheap compute plus LLM glue code equals viable design tools for non-experts — is repeating across every intuition-heavy engineering domain.
February 18, 2026
When the Machine Conjectured: GPT-5.2's Gluon Amplitude Formula and the End of AI-as-Tool
OpenAI's GPT-5.2 derived a genuinely novel formula for multi-gluon scattering amplitudes in QCD — a result that human physicists had not previously found. The formula was subsequently verified through formal mathematical proof by collaborating researchers at the Institute for Advanced Study and other institutions. This marks a categorical shift from AI as a tool that accelerates known work to AI as a collaborator that generates original conjectures in fundamental science. The result is symbolic, exact, and was validated by proof rather than experiment — the gold standard of mathematical truth.