Bookmarks

Signal Sources — Full Archive

Research & Labs

BAIR Blog
Berkeley's AI lab publishes its own papers straight from the people who wrote them, no press office in between. A recent post traces a kernel-porting project called K-Search from CUDA to Apple Silicon, the kind of write-up that never makes it into a conference recap. **Cadence:** weekly to biweekly. **Feed:** https://bair.berkeley.edu/blog/feed.xml
SAIL Blog
Stanford's AI lab runs the same format — grad students and faculty writing up their own results — with a recurring "Papers and Talks at [conference]" roundup that works as a cheat sheet for what actually mattered at ICML or CVPR that year. **Cadence:** weekly to biweekly, with spikes around conferences. **Feed:** https://ai.stanford.edu/blog/feed.xml
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. **Cadence:** irregular, roughly monthly to bimonthly. **Feed:** https://thinkingmachines.ai/blog/index.xml
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. **Cadence:** roughly monthly. **Feed:** https://transformer-circuits.pub/feed.xml
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. **Cadence:** multiple posts per week. **Feed:** https://sakana.ai/feed.xml
AI2 Blog
The Allen Institute's own teams write these up, which is why a post about a tutoring model called TutorMoments spends most of its length on when the thing should refuse to help rather than how well it scored. **Cadence:** weekly to biweekly. **Feed:** none
Microsoft Research Blog
Named researchers, not a comms team, describe what they built — pathology foundation models, spatial reasoning in vision-language systems, the occasional detour into physics-flavored ML. Several posts a week, so it's worth skimming rather than reading cover to cover. **Cadence:** several posts a week. **Feed:** https://www.microsoft.com/en-us/research/feed/

Engineering Blogs

Fly.io Blog
Fly.io's own engineers narrate the mess of running a public cloud in first person: GPU fleet rollouts, a Rust gossip protocol called Corrosion, the sandboxing work behind their agent product. It reads like someone explaining a system to a colleague, not a polished case study. **Cadence:** biweekly to monthly. **Feed:** https://fly.io/blog/feed.xml
Tailscale Blog
Alongside the product updates, Tailscale runs genuine postmortems. A recent one walks through hunting down a SQLite corruption bug that had apparently been sitting in the wild for sixteen years — that's the kind of debugging story worth the subscription on its own. **Cadence:** roughly weekly. **Feed:** https://tailscale.com/blog/index.xml
Turso Blog
Glauber Costa's team is rebuilding SQLite in Rust, and they write about the internals as they go: MVCC, async I/O, giving every customer their own database at scale. Primary-source database engineering, updated almost weekly. **Cadence:** weekly. **Feed:** https://turso.tech/blog/feed.xml
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. **Cadence:** monthly to every six weeks. **Feed:** https://oxide.computer/blog/feed
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. **Cadence:** weekly to biweekly in active stretches. **Feed:** https://turbopuffer.com/blog/rss.xml
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. **Cadence:** frequent, with an intern-post surge each September. **Feed:** https://blog.janestreet.com/feed.xml
Cloudflare Blog
Cloudflare posts almost daily, and the good ones go deep on the plumbing: a recent piece walks through saving petabytes of cache storage by switching compression algorithms, numbers included. Skip the product-launch posts and keep the rest. **Cadence:** near-daily. **Feed:** https://blog.cloudflare.com/rss/
Netflix Tech Blog
Real production detail from a company running recommendation and streaming infrastructure at a scale almost nobody else operates at — a recent write-up on multimodal asset personalization reads like an internal design doc that happened to get published. **Cadence:** weekly to biweekly. **Feed:** https://netflixtechblog.com/feed
Shopify Engineering
The team shipping Shopify's AI shopping assistant writes candidly about what didn't work first — a recent post on compressing agent context to cut cost reads more like a postmortem than a feature announcement. **Cadence:** roughly weekly. **Feed:** https://shopify.engineering/blog.atom
Meta Engineering Blog
This is the infrastructure feed, not Meta's PR-heavy AI blog — network engineers describing things like a custom RDMA transport built to move data between GPUs at training scale. Dense, but it's the closest thing to primary sourcing on how the frontier labs actually wire up their clusters. **Cadence:** roughly weekly. **Feed:** https://engineering.fb.com/feed/

Newsletters

ChinAI Newsletter
Jeff Ding translates Chinese-language writing on AI policy and industry that most English-language coverage never touches, including reactions inside China to its own AI companion regulations. No other newsletter is doing this specific job. **Cadence:** weekly. **Feed:** https://chinai.substack.com/feed
Interconnects
Nathan Lambert works on post-training at Ai2, and it shows. His posts on frontier training methods and the open-model ecosystem read like notes from someone who runs the experiments, not someone summarizing a paper abstract. **Cadence:** one to three times a week. **Feed:** https://www.interconnects.ai/feed
AI as Normal Technology
Arvind Narayanan and Sayash Kapoor, the pair behind the "AI Snake Oil" project, bring the same skepticism here, testing what AI agents can and can't actually do instead of assuming it. A useful counterweight if your feed already leans toward hype. **Cadence:** roughly monthly. **Feed:** https://www.normaltech.ai/feed
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. **Cadence:** two to four posts a week. **Feed:** https://thezvi.substack.com/feed
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. **Cadence:** roughly weekly. **Feed:** https://www.oneusefulthing.org/feed
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. **Cadence:** several posts a week plus a weekly briefing. **Feed:** https://www.transformernews.ai/feed
Import AI
Jack Clark has been writing this every week since 2016, longer than most AI newsletters have existed at all, and it still reads like one person's honest reaction to what crossed their desk that week rather than a roundup assembled by committee. **Cadence:** weekly. **Feed:** https://importai.substack.com/feed
The Pragmatic Engineer
Gergely Orosz gets people inside companies to talk on the record about decisions nobody else reports on — Meta's team-size math, a fintech building its own coding agent instead of buying one. Sourced reporting on engineering orgs is rarer than it should be. **Cadence:** about twice a week. **Feed:** https://newsletter.pragmaticengineer.com/feed
Ahead of AI
Sebastian Raschka picks apart one technique per issue — controlling reasoning effort, running coding agents locally — at the level of "here's the mechanism," which makes it slower to read than most AI newsletters and worth it for exactly that reason. **Cadence:** roughly monthly. **Feed:** https://magazine.sebastianraschka.com/feed
Latent Space: AI News
swyx and Alessio compress the day's AI Twitter, Discord servers, and research drops into one digest every weekday morning, which beats doing that scan yourself unless you enjoy it. **Cadence:** daily on weekdays. **Feed:** https://www.latent.space/feed

Changelogs & Release Feeds

vLLM Releases
The changelog for the inference engine most self-hosted LLM deployments run on. New model support and quantization tricks show up here, sometimes several times a week, before anyone gets around to writing a blog post about them. **Cadence:** multiple releases per week. **Feed:** https://github.com/vllm-project/vllm/releases.atom
Ollama Releases
Ollama ships constantly, and the release notes are where new backend support (MLX, for one) and CLI changes land first, often as release candidates before they hit the stable channel. **Cadence:** multiple releases per week. **Feed:** https://github.com/ollama/ollama/releases.atom
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. **Cadence:** near-daily. **Feed:** https://github.com/anthropics/claude-code/releases.atom
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. **Cadence:** roughly every two to three weeks, with patches in between. **Feed:** https://github.com/huggingface/transformers/releases.atom
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. **Cadence:** roughly biweekly. **Feed:** https://github.com/sgl-project/sglang/releases.atom
llama.cpp Releases
The changelog for the runtime most people actually use to run models on a laptop or phone. Builds land multiple times a day, and quantization or backend changes show up here well before anyone writes a blog post explaining them. **Cadence:** multiple releases daily. **Feed:** https://github.com/ggml-org/llama.cpp/releases.atom
Kubernetes Blog
The people who own each feature write its release note, which is why a post about storage version migration going on by default in 1.37 explains the reasoning, not just the flag name. **Cadence:** several posts a week around releases. **Feed:** https://kubernetes.io/feed.xml
PyTorch Blog
Framework-team announcements on what's landing next in the ecosystem — new releases, conference session recaps, the occasional deep dive into a specific optimization. If you deploy models, this is more useful than the release notes alone. **Cadence:** several posts a week. **Feed:** https://pytorch.org/blog/feed.xml

Practitioners & Independents

Eugene Yan
Eugene builds evals and recommendation systems for a living, currently at Anthropic and previously at Amazon and Alibaba. His posts read like internal engineering notes someone decided to publish: heavy on what actually shipped, light on speculation. **Cadence:** roughly monthly. **Feed:** https://eugeneyan.com/rss/
Vicki Boykis
A founding ML engineer with a long resume in embeddings and information retrieval, writing with a craft-first, occasionally contrarian streak. Her recent post on running local models well is more useful than most vendor comparison posts. **Cadence:** roughly monthly. **Feed:** https://vickiboykis.com/index.xml
antirez
Salvatore Sanfilippo built Redis and has spent two decades as a working systems programmer. Now he's turning that same skepticism on AI coding tools and lab governance, posting every few days to weekly with no patience for hype or doom. **Cadence:** every few days to weekly. **Feed:** https://antirez.com/rss
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. **Cadence:** bursty, several posts over a few weeks. **Feed:** https://lucumr.pocoo.org/feed.atom
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. **Cadence:** monthly to every other month. **Feed:** https://brooker.co.za/blog/rss.xml
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. **Cadence:** irregular, with several posts in September. **Feed:** https://danluu.com/atom.xml
Simon Willison's Weblog
Simon posts same-day notes with working code the moment a new model or tool ships, which makes his blog faster than most vendor documentation at answering "does this actually work the way they claim." **Cadence:** near-daily. **Feed:** https://simonwillison.net/atom/everything/
Julia Evans
Julia writes the "how does X actually work" post that should exist for every piece of infrastructure but usually doesn't — a recent one on Django internals is typical: patient, curious, no AI hype anywhere near it. **Cadence:** irregular, roughly monthly. **Feed:** https://jvns.ca/atom.xml

Video & Podcasts

The TWIML AI Podcast
Sam Charrington has been interviewing ML researchers about specific papers and systems since 2016, which makes this one of the longest-running technical interview shows still going. Each episode picks one piece of work and stays with it instead of skimming five. **Cadence:** weekly. **Feed:** https://feeds.megaphone.fm/MLN2155636147
The Cognitive Revolution
Nathan Labenz talks to AI builders and researchers multiple times a week, and occasionally sits down alone to walk through a paper — a recent episode dissects Anthropic's interpretability work on "global workspace" in more depth than the paper's own abstract. **Cadence:** multiple times per week. **Feed:** https://feeds.megaphone.fm/RINTP3108857801
Dwarkesh Podcast
Dwarkesh Patel gets frontier researchers and founders talking for two hours at a stretch, long enough that they stop giving the rehearsed answer. A recent episode on "agent civilizations" gets more technical than most written coverage of the same idea. **Cadence:** roughly weekly. **Feed:** https://www.dwarkesh.com/feed
Practical AI
Less about frontier lab drama, more about what it takes to actually build and ship an agent — Changelog Media's show stays grounded in things a working engineer can use Monday morning. **Cadence:** weekly. **Feed:** https://feeds.transistor.fm/practical-ai-machine-learning-data-science-llm

Benchmarks & Leaderboards

LiveBench
Most benchmarks get gamed within a few months of release. LiveBench refreshes its question sets on a rolling basis specifically to stay ahead of that, and scores against verifiable ground truth rather than another model's judgment. **Cadence:** continuous, new model configs added multiple times a month. **Feed:** none
SWE-bench Verified (Vals AI)
Vals AI runs every model through the same minimal bash-tool harness against the 500-task SWE-bench Verified set, which strips out the scaffolding tricks that can make some agent demos look better than the underlying model actually is. **Cadence:** rolling, roughly every one to two weeks. **Feed:** none
Epoch AI Benchmarking Hub
Epoch pulls FrontierMath, GPQA Diamond, SWE-bench, and a dozen other evals into one place, plus its own composite Capabilities Index. Useful when you want the whole picture instead of checking five separate leaderboards. **Cadence:** continuous, multiple updates a month. **Feed:** https://epochai.substack.com/feed
Artificial Analysis
Independent comparisons of model intelligence, speed, and price across every major provider, updated continuously and transparent about how the numbers are produced — closer to a buyer's guide than a marketing chart. **Cadence:** continuous. **Feed:** none
Arena
The team behind the crowdsourced head-to-head model rankings (formerly LMArena) explains its methodology changes as it makes them — new evaluation modes, coding-specific leaderboards — instead of just publishing a score and moving on. **Cadence:** roughly biweekly. **Feed:** none --- *19 sources added this month · 39 on the list. The full archive imports as [bookmarks](../../signal-sources-all-bookmarks.html) or as an [OPML feed bundle](../../signal-sources.opml). Found a dead feed? Reply to the email.*

Communities & Policy

AI Alignment Forum
A researcher-run discussion board for technical alignment work, where the comment threads are often as substantive as the posts — people like Paul Christiano showing up to argue about failure modes in reward hacking and debate training. **Cadence:** several posts a week. **Feed:** https://www.alignmentforum.org/feed.xml
Tech Policy Press
A nonprofit newsroom covering the actual mechanics of AI regulation and platform accountability, plus a standalone tracker for pending legislation and litigation. Closer to a beat reporter's notebook than a think-tank press release. **Cadence:** multiple articles daily. **Feed:** https://techpolicy.press/rss/feed.xml --- *20 sources added this month · 20 on the list. The full archive imports as [bookmarks](../../signal-sources-all-bookmarks.html) or as an [OPML feed bundle](../../signal-sources.opml). Found a dead feed? Reply to the email.*

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. **Cadence:** dated spec revisions every few months, blog posts more often between them. **Feed:** https://blog.modelcontextprotocol.io/index.xml
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. **Cadence:** several new entries a week. **Feed:** https://incidentdatabase.ai/rss.xml
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. **Cadence:** roughly monthly, timed to AI Act deadlines. **Feed:** https://artificialintelligenceact.eu/feed/

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. **Cadence:** near-daily to every other day. **Feed:** https://cset.georgetown.edu/feed/
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. **Cadence:** two to three posts a week. **Feed:** https://ainowinstitute.org/feed --- *20 sources added this month · TBD on the list. The full archive imports as [bookmarks](../../signal-sources-all-bookmarks.html) or as an [OPML feed bundle](../../signal-sources.opml). Found a dead feed? Reply to the email.*