Home / Blog
Blog
Field notes on engineering superintelligence.
AI security, custom LLMs, agentic systems and the infrastructure under them.
Infrastructure
GPU spend: measure cost per useful outcome
A busy GPU is not the same as a productive one. The metric that matters is what each verified, useful result costs, and how that changes as you tune the system.
October 13, 2026 · 4 min
Agentic systems
Evaluating agents: verifiable outcomes over vibes
An agent that looks good in a demo can still fail quietly in production. Evaluation should check what the agent actually changed in the world, not how convincing it sounded.
October 10, 2026 · 4 min
AI security
Model supply-chain security: weights, data, provenance
Models arrive with the same supply-chain risks as software, plus some of their own. Here is how to know what you are running and where it came from.
October 8, 2026 · 4 min
Custom LLMs
Custom LLM or frontier API: a decision framework
The build-versus-buy question for language models is really six smaller questions. Answer them per workload, not once for the whole organization.
October 6, 2026 · 4 min
AI security
Prompt injection and tool abuse: a defense checklist
No single filter stops prompt injection. Defending an agent means assuming some injected text will get through and limiting what it can make the agent do.
October 3, 2026 · 4 min
Strategy
Superintelligence, engineered: what it means for you
Superintelligence is a loaded word. For an enterprise, it should mean capability you can bound, govern and measure, not a promise you have to take on faith.
October 1, 2026 · 5 min