Better systems, not just better models.

4 stories · 1–2 minute read

01 / ENGINEERING

A software factory needs feedback loops

An agent needs a way to check its work. Illustrative engineering loop, not a diagram of OpenAI’s internal architecture. — Illustrative engineering loop, not a diagram of OpenAI’s internal architecture.

The Pragmatic Engineer reports how OpenAI connects Codex to internal workflows, including agents triggered by performance problems. The transferable idea: connect detection, implementation, and verification. Its unusually integrated internal setup matters.

Read the source ↗ newsletter.pragmaticengineer.com

02 / RESEARCH

Game skills transfer only with the right training

Good Start Labs reports that training a tool-using terminal agent on a railroad game improved financial-research performance; its single-turn setup did not. The training workflow mattered as much as the game.

Read the source ↗ latent.space

03 / INFRASTRUCTURE

Count delayed projects, not scary headlines

SemiAnalysis estimates roughly 2.3 GW of capacity is genuinely delayed by the restrictions it analyzed. Its distinction is useful: a project inside a restricted area is not necessarily blocked. These are the firm’s estimates.

Read the source ↗ newsletter.semianalysis.com

04 / MEDIA IDEA

Build a publication around real sales calls

Half Baked proposes a media business built on interviews with working salespeople: actual deals, scripts, mistakes, and follow-ups. Specific practitioner stories give readers something they can use—and give a publication a clear identity.

Read the source ↗ read.gethalfbaked.com