thinkingmachines.ai logo
thinkingmachines.ai

explore-blogs

Installation

Adds this website's skill for your agents

 

Summary

Enumerate the research blog posts on thinkingmachines.ai (the "Connectionism" blog at /blog/), returning each post's title, URL, publication date, and author. Read-only; optionally enriches posts with description, full text, and word count via RSS or individual post pages.

FIG. 01
FIG. 02
SKILL.md
236 lines

Explore the Thinking Machines Lab Blog (Connectionism)

Purpose

Enumerate the research blog posts published on thinkingmachines.ai — the blog is branded "Connectionism" and lives at /blog/ — and return a structured list of every post with its title, canonical URL, publication date, and author/attribution. Optionally enrich each post with its description, full article text, and word count. Read-only — this skill only reads public pages; it never submits forms, applies to jobs, or clicks "Join us".

When to Use

  • "What's on the Thinking Machines Lab blog?" / "List their latest research posts."
  • Monitoring Connectionism for new posts (poll the index or the RSS feed on a schedule).
  • Building a feed/digest of titles + dates + authors + links.
  • Pulling the full text of a specific post (e.g. "LoRA Without Regret") for summarization.
  • Any flow that would otherwise scrape the blog HTML — fetch is faster, cheaper, and more reliable than driving a browser here.

Workflow

thinkingmachines.ai is a Hugo static site, fully server-rendered. Every blog page — the index, each post, the RSS feed, and the sitemap — returns complete HTML/XML on a plain HTTP GET. There is no client-side rendering and no content JSON API (the only XHR the page fires is a Cloudflare RUM beacon at /cdn-cgi/rum). Cloudflare fronts the site but does not challenge simple fetches: a bare browse cloud fetch (no proxy, no stealth) returns HTTP 200. So the optimal path is a direct fetch + HTML/XML parse — driving a real browser is unnecessary and ~100× more expensive.

1. Enumerate posts from the blog index (primary)

GET https://thinkingmachines.ai/blog/

Each post is a list item shaped like:

<a class="post-item-link" href="/blog/{slug}/">
  <time class="desktop-time">May 11, 2026</time>
  <div class="post-info">
    <div class="post-title">Interaction Models: A Scalable Approach to Human-AI Collaboration</div>
    <div class="author-date"> Thinking Machines </div>
    <time class="mobile-time">May 11, 2026</time>
  </div>
</a>

Extract per post:

  • url — the href (relative /blog/{slug}/); prepend https://thinkingmachines.ai for the absolute URL.
  • title — text of .post-title.
  • published — text of the time element (e.g. "May 11, 2026"). Note there are two <time> nodes per item (desktop-time + mobile-time) with identical text — dedupe.
  • author — text of .author-date (e.g. "Thinking Machines", "John Schulman in collaboration with others at Thinking Machines"); trim whitespace.

The index lists newest-first and currently shows all 5 posts on one page — there is no pagination.

2. (Optional) Get full content / descriptions via RSS

GET https://thinkingmachines.ai/index.xml

A standard RSS 2.0 feed. Each <item> carries <title>, <link>, <guid>, <pubDate> (RFC-822), and <description> containing the full HTML body of the post (the feed is large — ~420 KB — because it inlines complete articles). One request yields every post's full text. Filter <item>s whose <link> contains /blog/ — /index.xml is the site-wide "recent content" feed and may include non-blog (/news/) items in the future; today it happens to contain only the 5 blog posts.

3. (Optional) Enrich a single post

GET https://thinkingmachines.ai/blog/{slug}/

Each post page is server-rendered and exposes clean metadata in <head>:

  • <title> — "{Post Title} - Thinking Machines Lab"
  • <meta name="description" content="..."> — one-line summary
  • <meta itemprop="datePublished" content="2025-09-29T00:00:00+00:00"> — ISO-8601 date
  • <meta itemprop="wordCount" content="5784">
  • <article>...</article> — the full rendered body (math is KaTeX, rendered client-side, but the raw $...$ LaTeX source is present in the server HTML).

4. (Optional) Discover all URLs via sitemap

GET https://thinkingmachines.ai/sitemap.xml

24 <loc> entries covering the whole site; exactly the 5 /blog/{slug}/ URLs plus the /blog/ index are the blog surface. Useful as a cross-check that the index didn't miss a post.

Browser fallback

Only needed if the fetch endpoints ever start returning a Cloudflare interstitial (not observed). Drive a Browserbase session:

sid=$(browse cloud sessions create --keep-alive --proxies | node -e "let s='';process.stdin.on('data',c=>s+=c).on('end',()=>{s=s.slice(s.indexOf('{'));process.stdout.write(JSON.parse(s).id)})")
browse open "https://thinkingmachines.ai/blog/" --remote --session "$sid"
browse wait load --remote --session "$sid"
browse snapshot --remote --session "$sid"   # ~60 a11y refs; post titles/dates/authors + the href map are all present

The accessibility snapshot exposes every post's title, date, author, and link, so a single snapshot after wait load is enough — no scrolling or clicking required.

Site-Specific Gotchas

  • The blog is named "Connectionism" and lives at /blog/. The top-nav label is "Connectionism", not "Blog". Don't confuse it with /news/, which is a separate section for company announcements (Tinker GA, NVIDIA partnership, grants, etc.) — not research blog posts.
  • No content API. The site is static Hugo HTML; the only XHR is the Cloudflare RUM beacon (/cdn-cgi/rum). Don't waste time hunting for a JSON endpoint — parse the HTML or RSS.
  • Proxies / stealth are NOT required. Cloudflare fronts the site (Server: cloudflare, cf-cache-status: DYNAMIC) but does not challenge GETs — bare browse cloud fetch returns HTTP 200 on the homepage, /blog/, every post, /index.xml, and /sitemap.xml. The pre-run probe predicted likelyNeedsProxies: true, but direct testing showed proxies are unnecessary for the fetch path. (The browser-fallback validation happened to run on a --proxies session, but that was belt-and-suspenders, not a requirement.)
  • browse cloud fetch prints an "Update available: 0.7.2 -> 0.8.3" banner to stdout before the JSON envelope. Strip it before parsing: slice the output from the first { (e.g. s = s.slice(s.indexOf('{')); JSON.parse(s)). jq is not installed in the sandbox.
  • /index.xml is site-wide "recent content", not blog-only. Today it contains exactly the 5 blog items, but to stay correct over time, filter <item>s by <link> containing /blog/. For a guaranteed blog-only feed, /blog/index.xml also exists.
  • Two <time> nodes per index item (desktop-time and mobile-time) carry identical text — dedupe so you don't double-count the date.
  • Author strings vary in shape — from a bare "Thinking Machines" to "John Schulman in collaboration with others at Thinking Machines". Treat the whole .author-date string as the attribution; don't try to split out a single name.
  • Date formats differ by surface: index = "May 11, 2026"; RSS <pubDate> = RFC-822 ("Mon, 11 May 2026 00:00:00 +0000"); post-page datePublished = ISO-8601. Normalize if you need a canonical date.
  • Post math is KaTeX. Article bodies contain LaTeX ($...$, $$...$$) rendered client-side by KaTeX. The raw LaTeX source is in the server HTML, so a fetch captures it; a browser screenshot captures the rendered math.
  • Browser-fallback quirk: inside a session already attached in CDP mode (e.g. with a trace client), browse open <url> --remote errors with "already running in cdp mode" — use plain browse open <url> --session <sid> instead.

Expected Output

Primary shape — the blog index enumeration:

{
  "success": true,
  "blog_name": "Connectionism",
  "blog_url": "https://thinkingmachines.ai/blog/",
  "post_count": 5,
  "posts": [
    {
      "title": "Interaction Models: A Scalable Approach to Human-AI Collaboration",
      "url": "https://thinkingmachines.ai/blog/interaction-models/",
      "published": "May 11, 2026",
      "author": "Thinking Machines"
    },
    {
      "title": "On-Policy Distillation",
      "url": "https://thinkingmachines.ai/blog/on-policy-distillation/",
      "published": "Oct 27, 2025",
      "author": "Kevin Lu in collaboration with others at Thinking Machines"
    },
    {
      "title": "LoRA Without Regret",
      "url": "https://thinkingmachines.ai/blog/lora/",
      "published": "Sep 29, 2025",
      "author": "John Schulman in collaboration with others at Thinking Machines"
    },
    {
      "title": "Modular Manifolds",
      "url": "https://thinkingmachines.ai/blog/modular-manifolds/",
      "published": "Sep 26, 2025",
      "author": "Jeremy Bernstein"
    },
    {
      "title": "Defeating Nondeterminism in LLM Inference",
      "url": "https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/",
      "published": "Sep 10, 2025",
      "author": "Horace He in collaboration with others at Thinking Machines"
    }
  ],
  "error_reasoning": null
}

Optional enriched per-post shape (when fetching an individual post page or RSS item):

{
  "title": "LoRA Without Regret",
  "url": "https://thinkingmachines.ai/blog/lora/",
  "published": "2025-09-29T00:00:00+00:00",
  "author": "John Schulman in collaboration with others at Thinking Machines",
  "description": "How LoRA matches full training performance more broadly than expected.",
  "word_count": 5784
}