Sources worth following
New releases arrive weekly. A short list of reliable sources beats a feed of everything.
Words on this page (3)
- repo
- : Repository: a project folder tracked by git, usually hosted on GitHub.
- trace
- : The full record of one run: inputs, model calls, tool calls, outputs, timing and cost.
- recall
- : Of the truly positive cases, the share the system found.
How to keep up without drowning
Pick three to five sources and ignore the rest. Prefer people who build and show evidence over accounts that repost excitement. Treat viral threads as pointers: find the primary source (the paper, the docs, the ) and read that. Check the date, since advice from a year ago may be outdated. When something claims a big win, ask what was measured and on what.
A starter set
Lab engineering blogs for how things are built. A practitioner blog with daily links and notes. One weekly newsletter for research. One paper aggregator. A community for local models.
A worked example: a weekly routine
Monday, 20 minutes: skim one newsletter for the week's research. Wednesday, 10 minutes: read one engineering blog post fully. Friday, 15 minutes: check the model release tracker and the radar page here. Total under an hour. Anything bigger gets a deeper read, from the primary source.
Tracing a claim to its source
A thread says: "New model beats the old one by 20 percent." Follow these steps.
- Look for a link. If there is none, search the lab's blog or model card for the release.
- Find the table behind the claim and note which test it uses and the settings.
- Check who ran it: the lab itself, or an independent group.
- Check the date, in case the claim describes an older version.
Suppose the table shows 10 points before and 12 points after. That is a 20 percent relative gain, since 2 divided by 10 is 0.2. It is only a 2 point absolute gain. Both numbers are true, and only the first makes a good headline.
What goes wrong: five sources that are really one
Your five newsletters all report the same new result in nearly the same words, with the same number. It feels like strong confirmation. It is often five summaries of one post.
To confirm, follow each link backward. If all of them end at the same page, you have one source, repeated. Count independent sources, not mentions. The lesson is to read the page they all point to.
How to tell your set is working
At the end of each month, write down three things you learned and where each came from. Mark each one as primary (paper, docs, repo) or secondary (someone's summary). Replace a source if:
- You cannot learning anything from it.
- It never links to primary sources.
- You skim it without reading.
A set that passes keeps you able to say, in one sentence, what changed and what evidence shows it.
Common mistake
Following everything. Sources you do not read are noise.
A viral post says a new model is "insane". A careful practitioner posts a test the same day: it improved on one benchmark by two points and got worse on another. The person who follows five trusted sources knows the real picture by lunchtime. The person who follows five hundred accounts knows only that something happened.
Check yourself
A viral thread claims a huge result. Best next step?
Not quite. Threads are pointers. The paper or docs hold the evidence.
How many sources should you follow closely?
Not quite. Depth on a few beats skimming everything.
A post says "this changes everything" but links nothing. What does that suggest?
Not quite. A claim without a source is a lead, not evidence.
You got 0 of 3 right.
Go deeper
- ArticleAnthropic EngineeringArticle
Anthropic · How Claude features and agents are built.
- ArticleSimon Willison's WeblogArticle
Simon Willison · Daily, careful notes on tools and models.
- ArticleLatent SpaceArticle
swyx and Alessio · Podcast and newsletter for AI engineers.
- ArticleAhead of AIArticle
Sebastian Raschka · Clear research explainers and architecture comparisons.
- ArticleInterconnectsArticle
Nathan Lambert · Post-training, open models and policy.
- ArticleImport AIArticle
Jack Clark · Weekly research summary.
- Tool
Hugging Face · Daily trending papers.
Spotted a mistake or something unclear? Report it on GitHub.