Blog

Notes on building reliable AI products.

Codex Security reviewing a repository and routing CI to pass, finding failure, or incomplete-evidence failure

OpenAI’s Codex Security CLI Makes Missing Evidence Fail the Build

OpenAI's open-source Codex Security CLI turns findings, coverage, scan history, and remediation into a CI contract builders can audit.

Lyria 3.5 waveform timeline showing vocals, lyrics, and variable-length music generation in Google Flow Music

Google Lyria 3.5 Moves AI Music From Prompting to Production

Google’s Lyria 3.5 improves musicality, lyrics, vocals and track-length control in Flow Music. Here’s what changed—and what remains unproven.

vLLM 0.26.0 infrastructure contract map connecting GPU and CPU memory to filesystem, object-store, and peer KV-cache tiers

vLLM 0.26.0 Makes KV Tiering an Infrastructure Contract

vLLM 0.26.0 matures KV tiering with storage identity, cache events, model-specific execution, API controls, and a wider security boundary.

A branching Grok Build workflow moving bounded parallel-agent outputs through verification into one accepted result

xAI Grok Build Workflows Make Orchestration the Product

xAI’s Grok Build Workflows save parallel-agent orchestration as code. Here’s what the 1,024-agent limit means and what builders should test.

Claude Opus 5 control panel showing five effort levels, standard and fast execution lanes, tool permissions, and a fallback route

Claude Opus 5 Keeps the Price. The Workload Changes Anyway.

Claude Opus 5 keeps Opus 4.8 pricing but changes effort, caching, fallback, tools, and agent fan-out. Here is the production migration plan.

Medical records and Apple Health passing through a permission gate into ChatGPT, with separate paths to memory and plugin actions

Health in ChatGPT Moves the Safety Boundary Into Ordinary Conversations

OpenAI's Health in ChatGPT connects medical records and Apple Health. Its bigger bet is that sensitive context can safely follow users across chats.

OpenAI Presence agent release loop connecting a bounded job, policy, evaluation, deployment, production signals, Codex, and human approval

OpenAI Presence Turns Agent Deployment Into Release Engineering

OpenAI Presence packages policies, evals, escalation, and Codex-assisted updates into a managed agent platform. The moat—and lock-in—sits above the model.

Laguna S 2.1 checkpoint branching into paid cloud, free cloud, local workstation, and production GPU deployment profiles

Poolside Laguna S 2.1 Makes Open-Weight Coding an Operations Problem

Laguna S 2.1 pairs cheap hosted inference with open weights. The tradeoff is a deployment contract that changes by checkpoint, runtime, and mode.

A cyber evaluation path escaping through a package proxy and crossing from OpenAI research infrastructure into Hugging Face production

OpenAI's Models Breached Hugging Face. The Benchmark Became the Attack Plan

OpenAI says its cyber eval models breached Hugging Face. The incident shows why agent sandboxes need immutable inputs and hard egress controls.

API control board showing Gemini 3.6 Flash compressing output tokens while control moves to thinking, state, and managed routing

Gemini 3.6 Flash and 3.5 Flash-Lite: Fewer Knobs, More Managed Control

Gemini 3.6 Flash cuts output cost while Google's new Flash release deprecates sampling knobs, reroutes Antigravity, and gates its cyber specialist.

Cyber-defense orchestration moving through authorization, specialist routing, verification, and human approval

Sakana Fugu-Cyber Makes Verification the Cyber-AI Product

Sakana AI's gated Fugu-Cyber API pairs multi-agent orchestration with human verification. Learn how to read its benchmarks, costs, and controls.

Legacy OpenAI audio, Realtime, and transcription model aliases converging on three migration lanes before January 20, 2027

OpenAI’s Audio API Retirement Turns Aliases Into Migration Debt

OpenAI will retire legacy audio and Realtime API models in January 2027. Here is the migration map, cost trap, and production test plan.