Short notes on AI systems, engineering, and ideas worth revisiting. Each entry links back to its original source.
Josh Rosen’s infrastructure roundup examines how gateways change the interface between MCP servers and agents. The important design question is what the agent actually sees after routing, filtering, and infrastructure policies have been applied.
This reply links to research on black-box, on-policy distillation. The paper’s Generative Adversarial Distillation approach pairs a student model with a discriminator that learns to distinguish student and teacher responses, providing adaptive feedback without access to the teacher’s internal parameters.
Chris Short points to Alibaba’s open-code-review project. Its hybrid approach combines deterministic checks with language-model analysis, a useful division of labor between rules that can be enforced consistently and issues that need contextual judgment.
This bookmark introduces hermes-bot-kit, combining a chat interface with live remote desktop access. The interesting interaction pattern is the ability to move between conversation and a visible execution environment when text alone is not enough.
TrueFoundry describes routing requests into simple, medium, and complex tiers rather than sending everything to the same model. Its reported savings illustrate a hypothesis worth testing against a workload’s own quality targets, latency requirements, and routing errors.
This security research examines risks introduced by intermediary AI routers. The authors report attacks and failures in the routing ecosystem. The architectural takeaway is to include intermediaries in the threat model instead of assuming a familiar model name guarantees an unchanged execution path.
Google Research describes ToolGrad, which starts from successful tool-use chains and then generates matching prompts. This reverses the common prompt-first process and uses execution feedback to improve the quality of training examples.
Andrew Chen’s announcement highlights SF Tech Week on October 5–11 and the scale of independently organized events across San Francisco and Los Angeles. This is an event-discovery bookmark; individual event details belong with their organizers.
This article proposes capturing corrections, decisions, and useful workflows as shared company knowledge. The central idea is a learning loop: improvements discovered in one person’s AI workflow should become reusable context for the rest of the organization.
Alex Lieberman’s notes on agent evaluation emphasize the combination of the agent, its system, tasks, and a verifier. Rubrics and reproducible environments matter because a plausible answer is not the same thing as a successfully completed task.
