The Inference Report

The infrastructure layer is reasserting itself. Trivy dominates vulnerability scanning across containers and code with straightforward threat detection, while systemd and protobuf remain the unglamorous backbone that everything else depends on. These aren't trending because they're new; they're trending because developers keep discovering they need them. The practical win here is consolidation: Trivy does secrets, misconfigurations, and SBOM generation in one tool rather than requiring a chain of specialized scanners. When infrastructure tools gain traction, it usually means the previous generation stopped working or became a bottleneck.

The secondary pattern is tooling for AI operations and observability. Phoenix and Claude HUD address a real friction point: models and agents are now complex enough that you need visibility into what's actually happening inside them. Arize and the terminal agent frameworks like gptme treat AI as something you deploy and monitor, not just experiment with in notebooks. Meanwhile, the data side is catching up with opendataloader and Clawith solving the unglamorous problem of getting messy PDFs and enterprise data into formats that models can actually use. LLMs-from-scratch sits apart with its enormous star count because it's educational infrastructure, not a production tool, but it signals that understanding how these systems work matters enough that people want to build them from first principles. The gap between what's trendy and what's useful is narrowing: the viral projects (MoneyPrinterV2, arnis) are novelties; the ones gaining sustained traction solve deployment, observability, and data pipeline problems that don't disappear.

Jack Ridley