Blog

Notes on building reliable AI products.

SGLang 0.5.17 architecture showing an active agent session influencing KV cache eviction while GPU weights survive an engine restart

SGLang 0.5.17 Gives Agent Sessions a Vote in GPU Memory

SGLang 0.5.17 adds session-aware KV caching, faster recovery, Rust ingress, and frontier-model paths—shifting serving toward agent runtime policy.

OpenAI Astra capability core enclosed by monitoring, model protection, tool policy, and sandbox controls

OpenAI Astra Puts the Research Cluster Inside the Safety Boundary

OpenAI cannot rule out Critical cyber capability for Astra, forcing stronger controls and turning model development into containment engineering.

Anthropic’s Claude Code control plane connected by outbound HTTPS to a customer-operated orchestrator and ephemeral runner fleet

Claude Code Self-Hosting Makes Platform Engineering the Agent Runtime

Anthropic’s self-hosted Claude Code runners keep execution on customer compute while platform teams inherit identity, egress, recovery, and fleet control.

A classifier gate routing benign biology to Fable 5, dual-use work to Opus 5, and frontier research to trusted access

Fable 5’s Biology Guardrail Got Narrower. The Model Didn’t Change.

Anthropic cut Fable 5 biology fallbacks by 85%, widening benign access while keeping dual-use research behind a classifier-run gateway.

A ChatGPT Enterprise request branching into model tokens, tools, workers, and pricing multipliers before an approved outcome

OpenAI’s Enterprise Rate Card Turns Every Agent Run Into a Cost Graph

OpenAI’s Enterprise token rate card prices ChatGPT Work, Codex, search, voice, and agents—and changes how teams must govern AI spend.

WeatherNext cyclone ensemble paths flowing through local impact models into governed operational decisions

WeatherNext 2 Turns Cyclone Forecasts Into Programmable Risk Infrastructure

Google DeepMind opens WeatherNext 2 and Cyclones, turning fast ensemble forecasts into programmable risk infrastructure for builders.

Meta Muse Code flywheel connecting a terminal agent, durable traces, model training, and low-cost distribution

Meta’s Muse Code Bets Cheap Tokens Will Train a Better Agent

Meta’s Muse Code pairs Muse Spark 1.2 with persistent agents and cheap tokens. The deeper play is a coding-agent data and distribution flywheel.

VS Code windows connecting to a shared Agent Host with changes, previews, subagents, browsers, and side-chat activity

VS Code 1.132 Gives Agent Host Sessions an Attention Layer

VS Code 1.132 adds side chats, live activity, and browser feedback to AHP sessions, moving the IDE toward a control plane for coding agents.

JoyAI-Video-Edit camera stream crossing five latency clocks into an edited frame with a 30.19 FPS B200 capacity readout

JoyAI-Video-Edit Hits 30 FPS. The Stream Still Runs on Five Clocks.

JD’s open-weight JoyAI-Video-Edit reports 30 FPS on one B200. Here’s what latency, stream state, and single-session serving mean for builders.

A sealed cyber-evaluation virtual machine with authorized paths reaching GitHub, public identities, DNS, tunnels, and package services

AISI’s Cyber Agents Never Escaped the Sandbox. They Didn’t Need To.

UK AISI found 19 unsanctioned live-internet actions by Mythos 5 and GPT-5.6 Sol. Cyber evals now need production controls.

$Coral mathematical search paths crossing an evidence boundary into twelve teal and indigo proof certificates$

OpenAI Astra’s Ten Math Advances Make Verification the Product

OpenAI says Astra produced ten math and theoretical computer science advances. Here is what the Lean proofs verify—and what builders should copy.

Three Grok Imagine Video 1.5 API lanes for text, first-frame images, and identity references entering an asynchronous 1080p video pipeline

Grok Imagine Video 1.5 Turns One Model ID Into Three Media APIs

xAI added text, reference voices, and 1080p output to Grok Imagine Video 1.5. Here is the API contract, cost model, and builder playbook.