OpenAI ships open weights for the first time since GPT-2
gpt-oss-120b and gpt-oss-20b under Apache 2.0. A strategic reversal, six years late, and genuinely useful.
page 8
older dispatches from the README archive.
Rate limits are a product decision, not an infrastructure one
When a tool's economics change, the users find out through the limit. There's a better way to do that.
Linux 6.16 and the steady grind of filesystem work
Faster ext4 writes, more Rust bindings, and a release with no headline feature. That's a good sign.
ToolShell: SharePoint gets exploited at scale
A patch bypass turns into mass exploitation of on-prem servers in days. The lesson is about what "patched" means.
ChatGPT Agent, and the browser sandbox as a product
OpenAI merges Operator and deep research into one agent with a virtual computer. The interesting part is the permission model.
Windsurf gets pulled apart in a week
An acquisition collapses, Google licenses the technology and hires the founders, Cognition buys the rest. Everyone learns something about acquihires.
Kimi K2 is a trillion-parameter open weights release
Moonshot ships a 1T-parameter MoE with 32B active, tuned for agentic tool use, with weights you can download.
Grok 4 and the benchmark that ate the discourse
xAI claims frontier results with heavy test-time compute. The number that matters is the one nobody quotes.
Write the README first
The oldest trick in the book, still the best design tool available, and almost nobody does it.
Cloudflare flips the default and starts charging crawlers
AI bots blocked unless allowed, plus a marketplace for per-crawl payment. A fifth of the web changes its robots policy at once.
Rust 1.88 ships let chains, and a decade-old papercut closes
if let A = a && let B = b finally compiles. Also naked functions and automatic cargo cache cleaning.
Gemini CLI puts a free agent in your terminal
Apache 2.0, generous free limits, and a very direct shot at the terminal-agent category.
Cursor at $9 billion and the coding-tool land grab
An editor fork raises at a valuation that only makes sense if you believe the IDE is the control point.
Gemini 2.5 goes generally available with a thinking dial
Pro and Flash hit GA, Flash-Lite arrives, and every tier exposes a thinking budget.
Meta buys half of Scale AI and most of its leadership
$14.3 billion for 49%, and Alexandr Wang moves to run a new superintelligence lab. The data layer just got picked.
WWDC 2025: Liquid Glass, a version number reset, and a local model API
Apple renumbers every OS to 26, redesigns everything, and quietly ships the most developer-relevant thing in years.
The Illusion of Thinking, and the argument about what reasoning is
An Apple paper finds reasoning models collapse past a complexity threshold. The rebuttals are as instructive as the paper.
Your logs are a product and you are shipping a bad one
Structured events, one line per unit of work, and the specific reason your grep-based debugging is slow.
OpenAI buys a hardware company that hasn't shipped anything
$6.5 billion for io, Jony Ive's design studio. The bet is that the phone is the wrong shape for this.
Claude 4 and the agent that runs for hours
Opus 4 and Sonnet 4 ship with a focus on long-horizon work, and Claude Code goes generally available.