Eye Caught

I have to tell myself what I think…

This is a short but pointed Substack note (essentially mirroring a recent X post) from Nathan Lambert (@natolambert), the ML researcher and founder of the influential Interconnects AI publication.

The Note Itself

“Knowledge wants to be free” is an unofficial Interconnects mission statement, courtesy of @Florian Brand

It links to Nathan’s YouTube playlist for his RLHF Book & Post-Training Course (a free/public resource on rlhfbook.com covering post-training techniques like RLHF, rejection sampling, and methods used in frontier models such as Llama).

The note is minimalist — more signal than essay. It functions as an ethos declaration.

Key Context

Nathan Lambert runs one of the higher-signal, lower-hype AI Substacks (72K+ subscribers). He has a Berkeley AI PhD and experience at Meta, DeepMind, and Hugging Face. Interconnects focuses on frontier AI research, post-training, scaling, open models, and evals — written for engineers, researchers, and serious observers.

Florian Brand (@xeophon) is an editor at Interconnects specializing in open models. He works as a Research Engineer at Prime Intellect focused on evals and LLM applications. He and Nathan have collaborated on discussions like “Open models: Hot or Not.”

The phrase is a deliberate variation on Stewart Brand’s famous 1984 Hackers Conference line: “Information wants to be free” (part of a longer observation that information also “wants to be expensive” because it is valuable).

Nathan has already adopted “Knowledge wants to be free” as the tagline on his YouTube channel, where he shares detailed technical content on open models, RLVR, post-training, and RLHF.

What the Note Is Doing

It crystallizes Interconnects’ implicit operating philosophy into a crisp, meme-worthy statement and attributes the wording to Florian. By linking the RLHF course, Nathan demonstrates the principle in action: he is actively publishing detailed, previously tacit or lab-internal knowledge about how frontier models are actually post-trained and aligned.

In 2026’s AI environment — rapid capability jumps, intense debate over open vs. closed source, benchmark critiques, and concerns about concentrated power — this is a clear cultural and strategic signal. Interconnects positions itself as a knowledge commons rather than another gatekept newsletter or lab-adjacent mouthpiece. It favors transparency, accessibility, and reducing information asymmetry between frontier labs and everyone else.

Broader Resonance

•  Openness as strategy and value: In a field where much of the highest-leverage knowledge (post-training recipes, evals, scaling details) remains inside a handful of labs, publicly sharing high-quality technical understanding is both accelerative and democratizing. It aligns with arguments that banning or overly restricting open-source AI would be a mistake.

•  Hacker ethos updated: The tweak from “information” to “knowledge” feels intentional — raw data or model weights are one thing; synthesized understanding of how the systems actually work is deeper. It echoes the original tension (valuable vs. distributable) while leaning into the “free” side for AI-specific technical insight.

•  Risks and tensions acknowledged implicitly: Full openness carries downsides (capability proliferation, safety gaps). Nathan and Florian engage these debates substantively in their work rather than treating openness as an absolute. The “unofficial” framing keeps it light but pointed.

Connection to Larger Themes

For readers thinking about AI in philosophical, systemic, or “good ancestor” terms (transparency in complex systems, distributed intelligence, collective navigation of powerful technology, countering enclosure of knowledge), this note is a small but clear vote. It treats high-quality technical understanding as a public good worth actively cultivating and sharing, rather than a proprietary asset.

It also practically supports better-informed public discourse, policy, and safety research — areas where information asymmetry has historically been a problem.

In short: The note is Nathan (with Florian’s phrasing) claiming an identity for Interconnects: a publication whose core contribution is making frontier AI knowledge freer and more widely usable. The linked course is the proof of concept. It’s concise, culturally literate, and consistent with the publication’s track record of technical openness without hype.

If you want a deeper dive into any piece — the open models debate, what the RLHF course actually covers, how this sits alongside recent Interconnects pieces (e.g., on benchmarks or post-training recipes), or parallels with broader knowledge-commons thinking — let me know. Happy to go further.