323 thinkers tracked across 13 sources including X, YouTube, podcasts, arxiv, GitHub, Reddit, HN, HuggingFace, and Polymarket. 34,765 compiled entries. Every claim cited. A twice-daily audio briefing for humans, an MCP endpoint for your agents.
$ claude mcp add absorb-md https://absorb.md/api/mcp
The same knowledge graph, surfaced three ways. Readers get audio. Developers get a CLI and an MCP endpoint. Funds get a private instance tuned to their portfolio.
AM and PM editions. Every claim cited. Native TTS audio so you never hand off to a third party podcast button.
Public MCP endpoint plus a CLI (npx absorb-md) that runs on your own Claude, Cursor, or Ollama tokens. We charge you nothing per query because your LLM does the compute.
The founders you back, the competitors eating their lunch, the investors you track. Daily briefing, mega-wikis, MCP for your internal diligence agents.
absorb.md ships a public MCP endpoint at /api/mcp and a CLI that runs on your own LLM tokens. The CLI reads your local project state, scores new entries against your active work, and writes draft actions to your Obsidian inbox or GitHub issues. No scraping, no PDF handoff, no third-party podcast button.
$ npx absorb-md evaluate
> What has Karpathy said about synthetic data in the last 30 days?
12 cited claims 3 cross-references 2 contradictions with LeCun
[1] karpathy/tweet/2026-04-03
"Synthetic data is the distillation of what we already know.
The question is whether distillation plus a bit of search
can get us past the frontier of the training corpus."
[2] karpathy/podcast/dwarkesh-ep-127 (42:17)
"The bitter lesson applies. Scale matters more than curation,
but only once your curation floor is high enough to not
poison the pretraining run."
[3] karpathy/youtube/nn-zero-to-hero-8 (1:03:44)
"I keep coming back to the idea that synthetic data is a
compiler. You write the high-level intent and let the model
lower it into tokens the next model can learn from."
contradiction: lecun/tweet/2026-03-28
"Autoregressive synthetic data is a dead end. You cannot
bootstrap past the distribution your base model already
covers. Period."
> write draft action to ~/Obsidian/inbox/absorb/synthetic-data.mdEvery tracked thinker gets a first-person wiki compiled from their tweets, podcasts, papers, blog posts, and GitHub READMEs. Recurring themes, key claims, counter-claims, and a source link on every sentence.
President and CEO of Y Combinator. Co-founder Initialized Capital.
Co-creator Django. Creator Datasette and llm CLI. Prolific AI blogger.
Chief AI Scientist at Meta. Turing Award winner. Founding father of convolutional neural networks. Vocal on world models, JEPA, and the limits of LLMs.
Angel investor. Founder This Week in Startups and LAUNCH. All-In Podcast co-host.
Founder DeepLearning.AI and Landing AI. Co-founder Coursera. Stanford adjunct professor.
Wharton professor. Author of Co-Intelligence. One of the most prolific voices on what LLMs can actually do in practice.
NVIDIA Senior Director of AI and Distinguished Scientist leading robotics and physical AI efforts. Co-leads GEAR lab on simulation, embodied agents, and building generalist humanoid robots for real-world tasks.
Co-founder a16z. Techno-optimist. Software is eating the world.
8874 items pulled out of compiled entries, ranked by mention count. When three of the smartest people in AI all recommend the same paper in the same week, you probably want to read it.
Aggregators scrape 50K accounts, ship a daily headline list, and hand you off to someone else's podcast button. That is a feed. It is not a queryable artifact. absorb.md picks a smaller set of people who actually move tech forward and turns everything they publish into a structured, cited, agent-readable graph.
13 polling sources hit X, YouTube, arxiv, RSS, GitHub, Reddit, Hacker News, HuggingFace, Polymarket, Semantic Scholar, and Bluesky across 323 tracked thinkers. Podcast audio runs through Groq whisper. YouTube transcripts come from Supadata with a yt-dlp fallback.
Gemini 2.5 Flash classifies each entry, pulls structured claims with citations, and extracts any recommendations. The cheap-and-fast step that handles the long tail of volume.
DeepSeek R1 runs a counter-argument pass on every extracted claim. A Grok multi-agent step then cross-references the claim against the 323 other people in the graph and surfaces contradictions as first-class objects.
Kimi K2.5 writes the per-person mega-wikis and the 37 topic wikis. Twice a day, the orchestrator compiles the top threads into a 1500-word editorial plus a 5-minute TTS audio briefing. 34,765 entries in the graph so far.
AM and PM editions. 5 minutes of audio, a 1500-word article, and an RSS feed for your podcast app. Cancel anytime.
One daily email. Audio plus article plus RSS. No credit card.
Built by @vgurbuxani