Homework From WordCamp  🔗 **Source:** [Homework From WordCamp](https://ma.tt/2026/08/homework-from-wordcamp/#full) > [!note]+ Document Notes > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > ## 🏷️ Title > Homework from WordCamp > ## 🧭 Table of Contents > 1. References > 2. Summary > 3. Main Content > 4. Paraphrase > 5. Newsletter > 6. Action Items > 7. Hypotheses > 8. Methodology > 9. Questions > 10. Future Research > 11. Implications > 12. Learning Styles Assessment > 13. Accessibility Assessment > 14. DEIB Assessment > 15. Logical Validity > 16. DARVO Analysis > 17. Conclusion > 18. Refactor Appendix > 19. Template Library > ## 🔗 References (in-page) > - Hugging Face, "Illustrating RLHF" (visual explainer) > - Chip Huyen, RLHF deep dive (huyenchip.com, 2023-05-02) > - Anthropic, "Constitutional AI" paper (arXiv 2212.08073) > - Andrej Karpathy, "Deep Dive into LLMs" (YouTube) > - Maxime Labonne, "Uncensor any LLM with abliteration" (Hugging Face blog) > - "Refusal in Language Models Is Mediated by a Single Direction" (arXiv 2406.11717) > - Open-weight model orgs on Hugging Face: Qwen, DeepSeek, Kimi (moonshotai), GLM (zai-org) > - An OBLITERATED model card (Qwen3.8-27B-OBLITERATED) > - Ollama; LM Studio > ## 📝 Summary > A short, direct follow-up post Matt Mullenweg published the same day WordCamp US 2026 closed, converting the unanswered show-of-hands moments from his own keynote into an AI reading list for the WordPress community. Three topics, each tied to a specific question the room didn't answer. (1) **How models are actually trained** — prompted by co-host Robert Jacobi's on-stage "I see no yeses" when asked who knew about the reinforcement learning stage; Mullenweg explains base models predict next tokens and the helpful behavior arrives in post-training via human/AI feedback, arguing this changes how you prompt and what you should expect. Recommended: Hugging Face's Illustrating RLHF, Chip Huyen's RLHF deep dive, Anthropic's Constitutional AI paper, and — the single pick if you only do one thing — Karpathy's "Deep Dive into LLMs." (2) **Uncensored/abliterated models** — almost nobody in the room had tried one; since refusal behavior is added in post-training, it can be subtracted, and the underlying research finds refusal is "roughly, a single direction inside the model." He explicitly draws the free-software parallel: open weights mean the community can modify these models "the same way the GPL means you can modify WordPress." (3) **The open-weight wave** — only a few hands for Qwen 3.8; he names Qwen, DeepSeek, Kimi, and GLM as at or near frontier at a fraction of the cost, states **token costs have dropped about 20x** with the curve not slowing, notes he demoed a local model from his own MacBook on stage, and prescribes running one tonight via Ollama or LM Studio — framing local inference as continuous with the four freedoms. Closes: "Read one, watch one, run one. Then bring what you learned to your next meetup," and "It is really important to me that WordPress community members ride this AI wave." Bottom line: this is the constructive, pedagogical face of the same AI-push the WPOCC survey found the community most negative about — and it's a genuinely useful, non-promotional reading list. But it also, read alongside that survey, answers a governance objection with an education program, which is the third distinct reframing of that critique now documented in this vault. > ## 📖 Main Content > ### 🗒️ Notes > **Framing and provenance.** The post's premise is the keynote itself: Mullenweg asked for shows of hands, got "mostly blank stares back, especially on the AI questions," and explicitly declines to treat that as a failing — "That's not a criticism — it's a reading list." The stated stakes: "If the future is going to be built on this stuff, the WordPress community should be fluent in it." (Direct corroboration of the fireside chat and The Repository's recap, both already in this vault; Robert Jacobi's "I see no yeses" is quoted here as an on-stage moment.) > **Topic 1 — post-training and RLHF.** The technical claim: base models are trained to predict the next word across huge text corpora; the friendly, helpful behavior is added afterward in post-training, where human/AI feedback reinforces some responses over others. His argument for why practitioners should care: "If you use these tools every day, understanding that pipeline changes how you prompt them and what you expect from them." Reading path is tiered by time investment — visual explainer → long-form deep dive → primary research paper → single-video option (Karpathy) if you only have one slot. > **Topic 2 — abliteration / uncensored models.** The mechanism: because refusal is *added* in post-training, it can be *removed*; the cited research finds it is "roughly, a single direction inside the model that you can subtract out." He's careful to separate the technique from its applications — "Whatever you think of the applications, it's worth understanding that open weights mean the community can modify these models, the same way the GPL means you can modify WordPress." That GPL analogy is the post's central rhetorical move: framing model modifiability as continuous with free-software values the audience already holds. > **Topic 3 — open-weight models and cost collapse.** Names four Chinese-lab open-weight families (Qwen, DeepSeek, Kimi, GLM) as "at or near the frontier at a fraction of the cost," notes smaller variants run on a laptop, and reports he demoed one from his own MacBook on stage. The quantitative claim: **"Token costs have dropped about 20x, and that curve isn't slowing down"** — stated without a source, timeframe, or baseline. Prescription: install Ollama or LM Studio, pull a small Qwen model, and "talk to an intelligence that lives entirely on your own machine, no cloud required." The values tie: "For a community built on the four freedoms, that should feel familiar — and exciting." > **Closing.** A three-verb call to action — "Read one, watch one, run one. Then bring what you learned to your next meetup" — followed by a personal statement of stakes: "It is really important to me that WordPress community members ride this AI wave." > **Incidental technical detail worth noting:** the page's own generator metadata reports `WordPress 7.2-alpha-63327` — Mullenweg's personal blog runs 7.2 alpha, one full version ahead of the 7.1 release that shipped the day before. > ## 📚 References > Entirely first-party (Mullenweg's own post) with an unusually high-quality outbound reading list: two arXiv papers, one Hugging Face official blog explainer, one independent long-form technical writer (Huyen), one widely-respected educational video (Karpathy), and four model-org landing pages. The recommendations are technically substantive and vendor-neutral — notably, three of the four named model families are competitors to Automattic's own AI partners, and the Anthropic paper is cited on its research merits rather than commercially. > ## 🗣️ Paraphrase > This is the best-faith version of Mullenweg's AI push, and it's worth reading as such. The reading list is genuinely good — real papers, a real explainer, Karpathy's video is the correct single recommendation — and it's not promotional: he's pointing people at Chinese open-weight labs and local inference, not at Automattic products. The GPL analogy is doing real work too, and it's not merely rhetorical: open weights genuinely do give users the modify-and-redistribute capability that the four freedoms describe, and abliteration is a vivid demonstration that the capability is real rather than nominal. But the thing worth holding in view is what this post is *doing* in sequence. The WPOCC survey found AI-in-WordPress to be its single most negative measure, with the objections landing specifically on process — defaults, opt-outs, consultation. At the keynote he answered that by proposing to rebrand AI as "automation." Here, a day later, he answers it again by proposing that the community needs to learn more. Both responses locate the problem in how the community *perceives* or *understands* AI. Neither engages the governance mechanism the survey actually A few times during the [WordCamp US](https://us.wordcamp.org/2026/) keynote, I asked for a show of hands and got mostly blank stares back, especially on the AI questions. That's not a criticism — it's a reading list. If the future is going to be built on this stuff, the WordPress community should be fluent in it. Here's the homework. #### How models are actually trained When I asked who knew about the reinforcement learning models go through, Robert looked out and said "I see no yeses." Base models are trained to predict the next word across huge amounts of text; the friendly, helpful behavior comes after, in post-training, where human/AI feedback reinforces some responses over others. If you use these tools every day, understanding that pipeline changes how you prompt them and what you expect from them. Start with Hugging Face's [Illustrating RLHF](https://huggingface.co/blog/rlhf), a visual explainer. Chip Huyen's [RLHF deep dive](https://huyenchip.com/2023/05/02/rlhf.html) is the best longer read, and Anthropic's [Constitutional AI paper](https://arxiv.org/abs/2212.08073) shows where post-training went next. If you only have time for one thing, watch Andrej Karpathy's [Deep Dive into LLMs](https://www.youtube.com/watch?v=7xTGNNLPyMI) — pretraining, fine-tuning, reinforcement learning, and why models behave the way they do, all in one sitting. #### Uncensored models Almost nobody had tried an [OBLITERATED](https://huggingface.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED) model. Since refusal behavior is added in post-training, researchers found it can also be removed — it turns out to be, roughly, a single direction inside the model that you can subtract out. Maxime Labonne's [Uncensor any LLM with abliteration](https://huggingface.co/blog/mlabonne/abliteration) explains the technique, and the paper behind it, [Refusal in Language Models Is Mediated by a Single Direction](https://arxiv.org/abs/2406.11717), is short and genuinely fascinating. Whatever you think of the applications, it's worth understanding that open weights mean the community can modify these models, the same way the GPL means you can modify WordPress. #### The open-weight wave Only a few hands went up for Qwen 3.8. The open-weight models coming out of Chinese labs — [Qwen](https://huggingface.co/Qwen), [DeepSeek](https://huggingface.co/deepseek-ai), [Kimi](https://huggingface.co/moonshotai), [GLM](https://huggingface.co/zai-org) — are at or near the frontier at a fraction of the cost, and the smaller ones run on a laptop. I demoed one from this MacBook on stage. Token costs have dropped about 20x, and that curve isn't slowing down. The easiest way to feel this for yourself is to run one locally tonight: install [Ollama](https://ollama.com/) or [LM Studio](https://lmstudio.ai/), pull a small Qwen model, and talk to an intelligence that lives entirely on your own machine, no cloud required. For a community built on the four freedoms, that should feel familiar — and exciting. Read one, watch one, run one. Then bring what you learned to your next meetup. It is really important to me that WordPress community members ride this AI wave., Homework From WordCamp, Matt Mullenweg, 01m0h2k6bwdjrqfdep5j1f3rth.