Diving Deeper
….things get murkier
Deeper dive into the ecosystem around Nathan Lambert’s note.
The short note (“Knowledge wants to be free” as Interconnects’ unofficial mission statement, credited to Florian Brand, linking the RLHF course) is not an isolated slogan. It is a concise manifesto that sits at the center of Lambert’s (and Interconnects’) ongoing work: technical transparency, open-model advocacy, nuanced critique of closed-lab opacity, and active construction of publicly accessible AI knowledge resources.
Here is a structured exploration:
1. What the RLHF Book & Post-Training Course Actually Covers
The free online resource (rlhfbook.com, with companion course videos, slides, GitHub code, and a comparison library) is positioned as “a one-stop shop for learning all foundational concepts for post-training language models.” It targets readers with some quantitative background who want to understand how frontier models are actually turned from base pretrained models into useful, aligned systems.
Core structure (organized around the “canonical RLHF recipe”):
• What RLHF does and why it was invented (seminal milestones, InstructGPT lineage).
• Reinforcement learning primer (policy gradients, PPO, REINFORCE, GAE, GRPO, etc.).
• The pipeline stages:
• Instruction tuning / supervised fine-tuning (SFT).
• Reward modeling (training a model to score outputs according to human preferences).
• Rejection sampling / best-of-N.
• Reinforcement learning (online RL with the reward model).
• On-policy distillation.
• Direct alignment algorithms (especially DPO and variants — surprisingly powerful for broad gains in math, code, reasoning).
• Broader origins: convergence of ideas from economics, philosophy, and optimal control.
• Advanced and current topics: synthetic data, tool use, character training / persona, evaluation challenges, overoptimization / reward hacking, open questions, RL with verifiable rewards (RLVR), reasoning-focused training, and product implications.
What makes it distinctive:
• It distills intuitions from frontier labs (acknowledgments include John Schulman and others from the RL sphere). Much of this knowledge has historically been tacit or only partially documented in technical reports.
• Practical artifacts: runnable code for the algorithms, a library for comparing model completions across post-training stages, lecture videos, and slides.
• Free and open (book content, code, course intro). A print edition is forthcoming via Manning.
In short, the course embodies the note’s ethos by taking techniques that were once mostly internal to labs (or scattered across papers) and making them legible, reproducible, and teachable. It lowers the barrier for researchers, engineers, and independent thinkers to understand and build upon post-training methods rather than treating them as black-box magic.
2. How This Fits Alongside Recent Interconnects Pieces
Interconnects has a consistent pattern: high-signal technical tracking of open models, deep dives into post-training realities, and skeptical-but-constructive commentary on evaluation and governance. The note’s emphasis on freely shared knowledge directly fuels and is reinforced by this output.
Key recent examples (mid-2026):
• Jun 16 – “Frontier post-training recipe review” (interview/podcast with Finbarr Timbers, former AI2 colleague): This is perhaps the closest companion piece. It traces the rapid evolution of post-training recipes from 2022–2026. Early open/academic recipes (e.g., OLMo, Tülu) were relatively simple and rigid (SFT → DPO → RLVR). Frontier/closed labs have moved to far more complex, industrial-scale approaches: training many domain-specialist teachers via targeted RL, then using Multi-teacher On-Policy Distillation (MOPD) to merge them into a general student without catastrophic capability trade-offs. Examples include MiMo Flash V2, DeepSeek V4, Nemotron 3 Ultra (>10 teachers), and others. The interview highlights how organizational scaling and compute enable this complexity, while open efforts often lag due to capacity limits. RL became expensive and conflict-prone when trying to optimize everything at once; specialization + distillation is the current pragmatic response.
• Benchmarks and evals coverage: Recurring theme of nuance and caution. Benchmarks are increasingly viewed as marketing-influenced, less reliable correlates of real-world/agentic performance, and plagued by reproducibility issues, prompt sensitivity, and out-of-distribution weaknesses. CAISI evaluations of open models (e.g., DeepSeek V4) often show widening gaps on certain metrics (Elo via Item Response Theory across multiple benchmarks including ARC-AGI-2 and private ones), but Interconnects notes that open models can be surprisingly strong on others and that “vibe tests” and actual usage matter more in the emerging “post-benchmark era.” Nathan has long worked on evaluation tools (e.g., RewardBench) and critiques how closed labs tune evals internally while releasing only favorable numbers.
• Open models tracking and open/closed dynamics: Monthly “Latest open artifacts” recaps, posts on specific releases (e.g., GLM-5.2 as a step change for open agents), and analysis that open and closed models are on “different exponentials.” Open efforts provide crucial competition, education, and innovation pressure; closed labs currently lead in certain complex post-training regimes but benefit from the broader ecosystem open work helps sustain.
The note sits perfectly in this ecosystem: by declaring knowledge-sharing as core identity and linking the RLHF course, Lambert signals that Interconnects will continue documenting, critiquing, and democratizing the very techniques (post-training recipes, evals) that determine capability. It is not naive boosterism for “open everything”; it is pragmatic transparency — make the maps public so more people can navigate, improve, and scrutinize the territory.
3. The Open Models Debate in Context
The Jun 19 co-authored piece “Banning Open Source AI Would Be A Mistake” (with Kevin Xu) is a direct policy-facing expression of the same ethos.
Core arguments:
• Open source (especially open weights) is safe, secure, and economically powerful (citing trillions in value from open source software historically).
• It drives education (tools with academic roots, now accessible), innovation (underdogs build on shared foundations, as with Linux or Android), and competition (counters concentration in closed duopolies like OpenAI/Anthropic).
• Safety via transparency: “Given enough eyeballs, all bugs are shallow” (Linus’s Law). Self-hosted models improve privacy (no data exfiltration to labs). Risks of open models are real but overstated relative to benefits; banning would concentrate power and potentially push global adoption toward less-scrutinized Chinese models.
• Historical parallels: Free software movement (MIT 1983 onward), Linux vs. Windows monopoly fears, Android enabling smartphone competition.
• Conclusion: Open source is “sunlight” (Brandeis) in technology. Banning or heavily restricting it would be anti-competitive, anti-educational, and ultimately counterproductive for U.S. leadership and safety.
This piece and the note reinforce each other. The mission statement provides the cultural north star; the policy argument applies it to real regulatory pressures (executive orders, congressional proposals, export-style controls on models). Interconnects consistently tracks open releases while acknowledging where closed labs currently pull ahead in post-training sophistication — a mature, non-dogmatic stance.
4. Parallels with Broader Knowledge-Commons Thinking
“Knowledge wants to be free” is a deliberate echo of Stewart Brand’s 1984 formulation (“Information wants to be free… Information also wants to be expensive”). The tweak to “Knowledge” emphasizes synthesized understanding and actionable technique over raw data — exactly what the RLHF course and post-training analyses provide.
Historical and conceptual parallels:
• Open source software as successful commons: Governed by licenses, norms, and contribution incentives rather than pure enclosure or tragedy-of-the-commons collapse. Linux, Apache, etc., scaled massively through shared knowledge.
• Scientific and educational commons: Open access publishing, preprints, shared datasets, and reproducible research traditions. AI open weights and evals extend this.
• Hacker / Whole Earth lineage: Brand’s broader project (Whole Earth Catalog, WELL, etc.) treated tools and knowledge as empowering for individuals and communities to shape their futures. The note and Interconnects work continue this in the AI domain.
• Governance lessons (Elinor Ostrom and others): Successful commons often have clear boundaries, monitoring, graduated responses to misuse, and nested institutions. Open source communities demonstrate versions of this; AI open efforts are still evolving governance around safety evals, responsible release norms, and compute access.
AI-specific tensions and opportunities:
• Enclosure vs. commons: Closed frontier labs create high-value proprietary knowledge (complex post-training recipes, internal evals). Open efforts (OLMo, Tülu, community datasets, shared benchmarks) push back by making foundations and insights public. The note explicitly chooses the commons side.
• Acceleration + scrutiny: Freely shared technical knowledge speeds collective progress and enables more eyes on safety/alignment questions. It also surfaces limitations (e.g., open recipes lagging in specialist distillation scale).
• Risks: Capability proliferation, misuse potential, sustainability (open work often under-resourced vs. well-funded labs). The debate piece addresses these directly rather than ignoring them.
• Philosophical resonance: In systems-thinking and “good ancestor” frames, curating and transmitting high-quality knowledge (rather than privatizing it) is a form of long-term stewardship. It supports distributed intelligence, reduces single-point-of-failure risks in understanding powerful technology, and aligns with transparency values in complex socio-technical systems. It also echoes themes of collective awakening or consciousness — better shared maps of the territory allow more people to participate wisely rather than relying on opaque oracles.
Bottom line: The note is small in form but large in implication. It names the operating principle behind Lambert’s book/course, the technical deep dives (post-training recipes, eval critiques), the open-model tracking, and the policy advocacy. Interconnects is actively building and defending a knowledge commons in AI at a moment when the default trajectory favors enclosure. That choice has downstream effects on innovation speed, competition, safety scrutiny, education, and who gets to understand and shape these systems.