OpenAI launched plugins for Codex, connecting it out of the box with Slack, Figma, Notion, Gmail, and more — turning the async coding agent into a full-stack workflow tool.
The integration means Codex can now triage bugs from Slack, manage calendars, and keep tabs on team activity without leaving the coding environment. The upstream enabler is OpenAI's decision to share authentication with the user's existing ChatGPT account — zero friction onboarding. This is OpenAI making Codex sticky by making it indispensable beyond just code.
OpenAI immediately followed the plugin launch by resetting Codex usage limits across all plans — a clear signal they want maximum adoption velocity on the new capabilities.
The timing is strategic — dropping expanded limits alongside plugins ensures the new integrations get stress-tested at scale while the excitement is fresh. "You can just build unlimited things with Codex. Have fun!"
Anthropic announced Claude Code auto-fix in the cloud — web and mobile sessions can now autonomously follow PRs, fix CI failures, and address review comments so your PR is always green.
The key shift is that this happens remotely: you can walk away from your computer entirely and return to a ready-to-merge PR. This is Claude Code evolving from a pair programmer into an autonomous CI agent — the kind of infrastructure that makes "vibe coding" actually sustainable in production.
Garry Tan celebrated gstack hitting 50K GitHub stars and told developers to install it into Claude Code right now.
The post had the energy of a founder watching organic growth hit escape velocity. 50K stars places gstack in rare territory for developer tools, and the Claude Code integration angle suggests the AI coding ecosystem is consolidating around a few key stacks.
François Chollet announced a new eval designed to measure the actual rate of AGI progress by identifying specific gaps — and immediately faced pushback about whether the bar is too low.
Chollet's framing is pointed: if you care about AGI progress, you should welcome evals that expose weaknesses rather than confirm priors. He clarified that all ARC-AGI-3 environments are human-feasible (each tested by 10 humans, passing if 2 could solve all levels) and that the bar is "very low, objectively" — not 100% human pass rate, which would be unreasonable for any benchmark including MNIST.
Agentica scored 36.08% on ARC-AGI-3 in a single day using their SDK — a rapid result that both validates the benchmark's accessibility and raises questions about what "hard" means.
Derya Unutmaz, MD shared this with a simple "Ouch!" — suggesting the result was either impressively fast or disappointingly low depending on your prior. The video showed the ARC-086-8 task at 12.13% solve rate, indicating wide variance across individual puzzles.
Chroma launched Context-1, a 20-billion parameter search agent that pushes the Pareto frontier of agentic search — claiming an order of magnitude faster and cheaper, released under Apache 2.0.
The video featured Kelly Hong from Chroma Research explaining the architecture. Christopher Manning reposted it, lending academic credibility. At 434K views and 2.1K likes, this was the highest-engagement AI infrastructure announcement on the timeline — agentic search is clearly the next battleground.
Cohere quietly released an Apache 2.0 speech recognition model on Hugging Face — no restricted license, no "research only" asterisk.
Maziyar Panahi asked whether this is a one-off or a genuine direction change for Cohere. The "Actually open. Respect." reaction captures a community tired of "open" models with restrictive fine print. Whether Cohere sustains this posture will be telling.
Dan Shapiro wrote about using Claude to repurpose a Canon camera into a fully functional Rust-based webcam app — replacing software he described as "a World War Two bunker on an otherwise beautiful beach."
The blog post captures the emerging pattern of developers using AI to solve the problems they've tolerated for years. Claude produced a working app overnight with no discernible bugs. The vibe coding dream, realized for a very specific annoyance.
via danshapiro.com · 202 likes, 11K views
Yohei Nakajima published a comprehensive survey on the state of AI memory systems — benchmarks, architectures, and what actually works — compiled using Claude Opus 4.6 Research.
The article covers top memory benchmarks and the highest-performing systems, offering a map of a field that's becoming critical as AI agents need persistent context. Using Claude to research memory systems is the kind of self-referential moment this era produces daily.
Toby Li reminded the timeline that humans are returning to the Moon for the first time in 54 years — exactly one week from today — and that this is not being talked about enough.
At 373K views and 10K likes, the engagement suggests the timeline agreed. Amid the AI noise, the Artemis mission represents the other great technological story of this moment, and the contrast in attention is itself a comment on where collective excitement has migrated.
Parimal connected Green's Theorem to physical tools still used in medicine and biology — the planimeter, which measures tumor cross-sections on MRIs and rare leaf areas.
Quoting ScieVision's thread on George Green (1793–1841), the tweet bridges abstract math and physical practice in a way that resonated quietly on the feed. A reminder that centuries-old mathematics lives inside everyday instruments.
The three biggest stories by reach tonight — Claude Code auto-fix (517K), Chroma Context-1 (434K), and the Moon return (373K) — represent three different scales of ambition: making today's code better, building tomorrow's AI infrastructure, and returning to another world. The feed is loudest about the first, most excited about the second, and most wistful about the third.