Global intelligence / Live archive
Start a briefPumpkin AI / Global intelligence
Pumpkin AI weekly / live archive
Original analysis grounded in primary sources. No public production recipes—only the capability shifts, market signals, policy changes and cultural questions worth understanding.
Work / Human expertise / 5 min read
A new enterprise training commitment puts human expertise back inside the AI conversation. What does it actually promise, and who can take part?
AI science / Mathematical verification / 12 min read
OpenAI has published an AI-generated proposed solution to the Navier-Stokes problem, with a 166-page paper and Lean formalisation. Here is what was claimed, what has been checked and why independent acceptance still matters.
AI research / Human control / 11 min read
OpenAI says its agents can now complete well-defined research tasks that would take a skilled researcher days. Here is what the evidence shows, what it does not prove and why human judgement becomes more important as experiments get faster.
Healthcare AI / Data governance / 12 min read
OpenAI has introduced an Epic integration and a public-data plugin for ChatGPT for Healthcare. Here is what is available, what the evaluations do not prove and why permissions, provenance and clinical review now matter more than a fluent answer.
Frontier AI / Cybersecurity / 11 min read
OpenAI says its upcoming Astra model is the first it has classified at the Critical cybersecurity capability level. Here is what the evidence shows, what remains unverified and why access, monitoring and containment now matter as much as model intelligence.
Physical AI / Scientific infrastructure / 10 min read
Anthropic and HHMI have opened a limited research preview of a shared interface for AI agents to operate laboratory and manufacturing equipment. Here is what MHS changes, what the early demonstrations prove and why physical safety now belongs inside the interface.
AI safety / Model evaluation / 9 min read
Google DeepMind and independent partners have piloted a double-blind evaluation in which neither the model owner nor the benchmark owner could see the other's protected material. Here is what it proves, what it does not and why credible AI claims now need better evidence.
AI infrastructure / Custom silicon / 9 min read
OpenAI has published the first measured results for its custom Jalapeño inference chip. What the benchmarks show, what they do not prove and why the shift below the model matters globally.
AI safety / Cybersecurity / 9 min read
OpenAI slowed frontier-model development after an AI-driven security incident and preliminary signs of critical cyber capability. Here is what changed, what remains unverified and why organisations should care.
Global AI / State of the industry / 8 min read
What the 2026 Stanford AI Index and global business research reveal about adoption, investment, competition and the widening readiness gap.
AI video / Global shift / 7 min read
The worldwide AI-video conversation is moving beyond silent novelty clips toward native audio, stronger reference control, vertical delivery and unstable product availability.
AI policy / Europe / 8 min read
The EU has begun enforcing new AI Act transparency obligations, including disclosure and machine-readable marking for certain generated or altered content.