Bookmarks

Signal Sources — November 2026

Research & Labs

EleutherAI Blog
This is the research collective behind GPT-NeoX and the Pile, and the posts read like lab notebooks rather than announcements. A late-October entry traces how subliminal learning survives changes in initialization, data order, and architecture — the kind of negative-result digging that rarely makes it into a paper abstract. **Cadence:** irregular, one to three posts a month. **Feed:** none.
Prime Intellect Blog
Prime Intellect builds decentralized infrastructure for RL training, and its engineers write up the kernel and systems work as they ship it: a late-October post on fused MoE kernels, another on rewriting their agent runtime in Rust. Several posts a week, almost none of it roadmap talk. **Cadence:** multiple posts weekly. **Feed:** none.
Goodfire Blog
Founded by researchers who left Anthropic and OpenAI to work on interpretability full time, Goodfire just announced a $10 million initiative it calls a "Human Genome Project for AI." An earlier post walking through sparse autoencoders works better as a primer than most of the papers it cites. **Cadence:** weekly to biweekly. **Feed:** none.

Engineering Blogs

ClickHouse Blog
The database vendor's own engineers publish the internals work, not the marketing copy: a recent post on the new cost-based join optimizer gets into reordering logic for non-inner joins, and another describes an autonomous coding agent that has merged more than 3,000 pull requests into the codebase in under seven months. **Cadence:** several posts a week. **Feed:** https://clickhouse.com/rss.xml
Modal Blog
Modal runs serverless GPU infrastructure, and the posts worth reading here are the scheduler and scaling stories — one on launching a trillion-parameter model for Sakana AI, another on scaling to a million concurrent sandboxes. Skip the pure product posts; the infra ones carry the subscription. **Cadence:** two to three posts a week. **Feed:** https://modal.com/blog/atom.xml
PlanetScale Blog
PlanetScale's team names database tradeoffs most vendors avoid saying out loud — "I don't trust ON DELETE CASCADE" is a representative title, and a recent piece walks through designing a planned primary switchover with query buffering and fast rollback built in from the start. **Cadence:** multiple posts a week. **Feed:** https://planetscale.com/blog/rss.xml

Newsletters

The Sequence
Jesus Rodriguez runs several rotating tracks here instead of one generic roundup — Radar, Opinion, Knowledge, a dedicated robotics issue — and a recent Radar edition on AI financing broke down Microsoft's capital expenditure against its operating cash flow in more detail than most financial press managed. **Cadence:** three to four times a week. **Feed:** https://thesequence.substack.com/feed
Gradient Flow
Ben Lorica has been writing about applied ML and data infrastructure since before most of the current AI newsletter crowd existed, and it shows in posts like a recent one where he admits he changed his mind about world models after a robotics result surprised him. **Cadence:** weekly. **Feed:** https://gradientflow.substack.com/feed
Understanding AI
Timothy B. Lee left tech journalism at Ars Technica and Vox to run this solo, and it reads like beat reporting rather than opinion: a recent piece works through all 421 pages of Anthropic's S-1 filing just to check whether the company's financials actually hold up. **Cadence:** several posts a month. **Feed:** https://www.understandingai.org/feed

Changelogs & Release Feeds

DuckDB Release Notes
Every release post is written by the core maintainers explaining why a decision got made, not just what changed — the DuckDB 2.0 LTS announcement shipped alongside a separate "Gems of 2.0.0" post digging into the smaller changes that didn't make the headline. **Cadence:** on release, roughly monthly. **Feed:** https://duckdb.org/feed.xml
Qdrant Releases
The vector database's GitHub release notes carry real performance numbers instead of asserted ones — the latest added BM25 scoring directly on full-text payload indexes and two new quantized datatypes, benchmarks attached. **Cadence:** every two to four weeks. **Feed:** https://github.com/qdrant/qdrant/releases.atom
LiteLLM Release Notes
The proxy layer most self-hosted LLM setups route through ships point releases almost daily, and the notes name the actual production bugs getting fixed — a gzip compression regression quietly slowing down inference routes got its own release note the day it was patched. **Cadence:** near-daily. **Feed:** https://github.com/BerriAI/litellm/releases.atom

Practitioners & Independents

Hamel Husain
Hamel consults on LLM evaluation for a living, and his posts skip the vibes-based advice that fills most of this space — a recent one tests decision models as an alternative to LLM-judge scoring and finds that most teams never bother tuning the default 0.50 threshold, leaving real accuracy sitting on the table. **Cadence:** every two to four weeks. **Feed:** https://hamel.dev/index.xml
Citation Needed (Molly White)
Molly White built a reputation debunking crypto before most people were paying attention, and she's still doing sourced accountability reporting now that AI money has tangled with the same industry — a recent post traces how North Korea laundered proceeds from the Binance hack through exchanges that barely check who's withdrawing. **Cadence:** roughly weekly. **Feed:** https://www.citationneeded.news/rss/
Tim Dettmers
Tim built bitsandbytes and QLoRA, so when he writes about running frontier-scale models on consumer hardware through extreme quantization, it comes from the person who wrote the quantization code everyone else now uses. His most recent post argues academic labs are entering a renaissance precisely because they can't compete on raw compute. **Cadence:** every one to three months, each post substantial. **Feed:** https://timdettmers.com/feed/

Video & Podcasts

No Priors
Sarah Guo and Elad Gil, both investors who've actually built things, get founders and researchers talking shop instead of pitching — a recent episode with Serval's CEO gets into the deterministic approval systems that let enterprises hand coding agents access to production without someone babysitting every commit. **Cadence:** weekly. **Feed:** https://feeds.megaphone.fm/nopriors
Last Week in AI
Andrey Kurenkov and Jeremie Harris run a denser, more technical recap than most AI news shows, structured so you can skip straight to the research segment if the policy drama doesn't interest you — episode 260 covers Meta's Muse download numbers and OpenAI's misalignment disclosures in the same breath. **Cadence:** weekly. **Feed:** https://rss.art19.com/last-week-in-ai

Benchmarks & Leaderboards

METR Blog
METR is the nonprofit that runs pre-deployment safety evaluations for the frontier labs, and the blog is where the actual findings surface before anyone else gets to characterize them — a recent post describes a three-week pilot where METR and Redwood Research stress-tested OpenAI's internal agent-monitoring system and turned up real coverage gaps. **Cadence:** weekly to biweekly. **Feed:** https://metr.substack.com/feed
Vectara Hallucination Leaderboard
Most leaderboards chase general capability; this one tracks exactly one failure mode, how often a model invents facts while summarizing a short document, and adds new models within days of release. The latest update scored both DeepSeek-V4.1-Flash and Claude Opus 5.5 against the same test set. **Cadence:** roughly weekly. **Feed:** https://github.com/vectara/hallucination-leaderboard/commits/main.atom
Terminal-Bench
Stanford, Harbor, and the Laude Institute run this one jointly, which keeps it clear of any single vendor's incentives — real command-line tasks that measure whether an agent can actually operate a terminal instead of just describing how it would. **Cadence:** updated every one to two weeks as new models get run. **Feed:** none. --- *20 sources added this month · 79 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.*