🔬 Research & Labs
Connectionism (Thinking Machines Lab)
Mira Murati's lab writes in long form and takes its time: a July post on open-weight safety worked through the actual tradeoffs instead of picking a side, and a May research preview described models built for continuous audio-video-text interaction rather than the usual chat turn-taking. Ex-OpenAI researchers including John Schulman are among the authors, so posts often read like people arguing with their own prior work.
Transformer Circuits Thread
Anthropic's interpretability team publishes here instead of on the main company blog, and the difference shows: an August piece characterizes interference weights inside a deliberately tiny language model, and a July one found evidence that LLMs keep a privileged, reportable slice of their internal state separate from everything else they represent. This is where the mechanistic-interpretability papers actually live, not where they get summarized.
Sakana AI Blog
The Tokyo lab founded by former Google Brain researchers moves fast enough that September alone brought a new scaling result for in-context imitation learning, a new Frontier Intelligence Group, and word that Jürgen Schmidhuber had joined as a researcher. Posts mix technical writeups with lab news at a pace closer to a startup than an academic group.
🛠️ Engineering Blogs
Oxide Computer Blog
Bryan Cantrill's team builds server hardware and the software stack underneath it from scratch, and an August post on running Kubernetes atop that stack names three different provisioning paths they tried, the custom cloud-controller-manager workaround they shipped, and the disk hot-plug support that still isn't done. Few vendor blogs admit the gaps this plainly.
turbopuffer Blog
The vector-search startup announced in late September that it's tearing out the assumption most vector databases are built on, that the ANN index should be primary, after storage and write amplification caught up with them at scale (a trillion-plus documents, tens of millions of writes per second in production). It's a rare case of a vendor publishing the reasoning behind a core architecture reversal, numbers included.
Jane Street Tech Blog
Jane Street runs a crop of intern projects every summer and publishes the writeups each September; this year's standout walks through an activation-checkpointing planner for PyTorch that beats the framework's own built-in algorithm across memory budgets. The trading firm's blog has covered OCaml internals and formal methods for years, but the ML systems posts are what earn it a slot here.
📮 Newsletters
Don't Worry About the Vase
Zvi Mowshowitz puts out a comprehensive numbered "AI" roundup most weeks, and the September 30 edition on the White House's AI safety accord is typical: heavy original argument laid over the links, not just a list of them. Thirty-eight thousand free subscribers and near-universal citation inside the AI-safety world say this reads more like a beat reporter's notebook than a newsletter.
One Useful Thing
Ethan Mollick teaches at Wharton and writes about what AI actually changes in how people work and learn, grounded in his own research rather than vendor demos. His October 1 piece on self-organizing agent swarms (including a case where thousands of coordinating agents worked out a mathematical proof) is the kind of practitioner-facing analysis that doesn't show up in product blogs.
Transformer
A former Economist editor runs this like an actual newsroom for AI governance: opinionated, sourced, and willing to take a side, as in a late-September piece arguing OpenAI's decision to scrap its Astra 6.1 release was the right call but shouldn't have been OpenAI's to make alone. It moves faster and argues harder than the institutional policy trackers covering the same ground.
📦 Changelogs & Release Feeds
Claude Code Releases
Anthropic's own coding agent ships release notes as terse bullet lists on GitHub rather than a blog, and they land fast: five versions went out in six days at the end of September. If you use the tool, this is the actual changelog; the docs page just redirects here.
Hugging Face Transformers Releases
Every minor release lists exactly which newly supported model architectures just became loadable off the shelf; the September 30 release alone added support for four new model families. For anyone tracking which open-weight models are actually usable yet, this beats waiting for a blog post to notice.
SGLang Releases
The inference engine competing with vLLM for serving LLMs at scale ships roughly every two weeks, and the release notes report contributor counts alongside the technical changes; one recent cycle logged 713 pull requests from 237 people. It's a useful pulse check on how much of the serving-stack work is happening outside the big labs.
✍️ Practitioners & Independents
Armin Ronacher
The creator of Flask spends as much time on AI policy opinions as on Rust internals these days, and his September post arguing that open-weight models reduce risk better than lab-controlled pacing proposals reads like someone who builds things arguing with people who mostly don't. A later post on rethinking Rust serialization is proof he hasn't fully switched lanes.
Marc Brooker
An AWS distinguished engineer who normally writes about distributed systems spent late September describing a model he built himself: a two-billion-parameter classifier with a custom scoring head bolted on in place of the usual language-model head, tuned to a Brier score of 0.009 on easy tasks. It's a rare firsthand account of a systems person learning ML the hard way instead of just deploying someone else's model.
Dan Luu
Dan Luu tests claims instead of repeating them. A September post ran coding agents against the same compression-library task under 26 different prompted methodologies (formal verification, property-based testing, fuzzing, and more) and found the agents mostly ignored whatever technique they'd been told to use. No AI-hype framing anywhere in it, just the data.
📚 Docs & Reference Hubs
Model Context Protocol Specification
The changelog for the protocol most agent runtimes, including Claude, now implement for tool calling. Each entry links back to the GitHub discussion that produced it, so a line like "removal of protocol-level sessions" comes with the actual argument behind it, not just the decision. Worth checking after any agent framework starts behaving differently.
AI Incident Database
A nonprofit-run catalog of real-world AI harms and near-misses, over 1,500 entries deep and sourced from more than 3,000 news, academic, and government reports going back to 2015. A late-September entry on a coding assistant uploading code without consent is a reminder this updates faster than most people check it.
EU AI Act Explorer
The Future of Life Institute keeps the full annotated text of the EU AI Act searchable by article and recital, paired with a compliance checker, and updates its explainers each time a new phase-in deadline hits. It's a more reliably maintained reference than the European Commission's own service-desk page, which carries no visible update date.
🏛️ Policy & Industry
CSET (Center for Security and Emerging Technology)
Georgetown's policy shop translates Chinese-language AI policy documents that most English coverage never touches and runs a standing tracker on chip export controls and compute costs. A late-September post breaks down what AI chips actually cost to track, not just where they're banned from going. Near-daily output from a team that includes former government and industry staff.
AI Now Institute
This research institute takes no corporate or tech-industry funding, which shows in its focus: who controls AI markets, what surveillance it enables, and who's accountable when it fails. A late-September post connects an AI-related close call between the US and China to the broader argument that nobody's watching the systems that matter most, a sharper, more adversarial read than the policy-tracker norm.