Insights
Notes from the work.
Practical writing on building enterprise AI that reaches production.
Google's grader failed three agent runs that had passed. It was reading the summary, not what the agent did
Google published four evaluation lessons from building an agent plugin: grade the execution trajectory, keep turn structure, compare live and plan-only runs, and baseline on expert golden sets.
Read
The cheapest model is the one whose errors you can afford to catch. A way to price AI by where mistakes land
Recent days brought three ways to cut inference cost: specialize a small model, serve open weights better, finish agent tasks in fewer steps. None prices the errors. A framework for doing that first.
Read
A small model got 85% better at extraction. Microsoft's test is useful because the frontier still won
With 17 training documents, a Python grader and 100 reinforcement steps, Microsoft lifted gpt-oss-20b's extraction F1 from 0.44 to 0.81. Frontier models scored 0.98. How to read the trade.
Read
Announced, contracted, powered: Microsoft's exit from an Essex AI data center shows which capacity counts
Microsoft left Nscale's planned Essex AI data center after the grid operator said full power may not come until the early to mid 2030s. What capacity buyers should contract for.
Read
Your coding agent picked the right library. Your company may not be allowed to use it
GitHub made enterprise-wide open source license policies generally available, enforced at merge and exposed through an API that coding agents can query. The API helps only if agents are made to ask.
Read
DeepSeek's update aims at cheaper finished tasks. Its new search API raises a custody question
DeepSeek V4.1-Flash-1017 claims faster, shorter agent runs under an alias that already changed, and DeepSeek opened its search engine as a per-call API. Cost per task falls; custody questions rise.
Read
AI malware scanners miss what they cannot read. Databricks closed the gap with evidence, not a bigger model
Databricks found LLM scanners struggle with npm packages that hide payloads in compiled binaries. Structured binary-analysis evidence fixed most misses, even for small models.
Read
An engineer asks an Azure agent why an app is slow. Microsoft's threat model asks what else it could do
Microsoft published a threat model for AI assistants connected to Azure through Azure MCP Server. Its best idea: the agent may interpret, Azure must decide, and denial paths get tested first.
Read
Reflection's Beam weights are out. How you serve them now matters as much as the model
Beam shipped on October 16 in three precisions, with day-zero recipes from AWS and Google Cloud. For coding agents, the serving architecture now decides cost as much as the model does.
Read
AWS's sales agents were healthy and financially wrong. The fix was building evaluation into every release
AWS published how it evaluates the multi-agent system its field sellers use: nine stages from defining quality to production canaries. The useful part is where each check sits.
Read
Microsoft's chat model can now answer with an interactive chart. Your app still decides what it may render
GPT-chat-latest in Microsoft Foundry adds GenUI answers and interleaved-thinking output on the existing endpoint. The interface is now part of the response, and needs a rendering contract.
Read
An enterprise agent should survive a restart without paying the invoice twice. AWS's durable Go SDK is GA
AWS made its Lambda durable execution SDK for Go generally available: checkpointed steps, approval callbacks and waits of up to a year. Durability is a decision separate from the model.
Read
Have an AI use case stuck between prototype and production?
Tell us what you’re trying to ship. We’ll reply with honest next steps.
Discuss a use case