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 Do not quote or attribute any other line to the named author. > <!-- deep-read:sections start --> ## 🏷️ 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 identified. That's now three documented reframings of the same critique across three sources in this vault, and the pattern is more informative than any one instance. ## 📰 Newsletter **Matt Mullenweg turned the blank stares from his own WordCamp keynote into an AI reading list — and it's a good one.** Three topics, each tied to a question the room couldn't answer: how models are actually trained (post-training/RLHF, with Karpathy's "Deep Dive into LLMs" as the single pick if you only have one slot), abliterated models (refusal turns out to be roughly one direction inside a model that you can subtract out), and the open-weight wave from Qwen, DeepSeek, Kimi, and GLM. His claim: token costs have dropped ~20x and the curve isn't slowing. The prescription is refreshingly concrete — install Ollama or LM Studio, pull a small Qwen model, and run an LLM entirely on your own laptop tonight. The framing that'll land with this audience: open weights mean you can modify the model "the same way the GPL means you can modify WordPress." Read one, watch one, run one. ## ✅ Action Items - **Core AI:** The reading list is a legitimately good onboarding curriculum for contributors working on AI-adjacent core features — worth adopting or adapting rather than writing something new. - **Training:** Directly actionable — this is a ready-made, vendor-neutral learning path (visual explainer → deep dive → paper → video → hands-on local install) that could become a Learn WordPress module with minimal adaptation. - **Community:** The "bring what you learned to your next meetup" framing is an explicit invitation to make this meetup programming; worth pairing with the WP 101 beginner-track model from WCUS. - **Core / Core Performance / Core Program Team / Mobile / Playground / Marketing / Support / Design / Photos / TV / Polyglots / Accessibility / Meta / CLI / Tide / Openverse / Themes / Plugins:** Not directly applicable — this is an educational post, not a technical or governance change. - **Documentation:** The tiered-by-time-investment structure (one visual, one long read, one paper, one video, one hands-on) is a reusable pattern for any docs team building a learning path on a complex topic. - **Hosting:** The local-inference angle (small models on a laptop, no cloud) is worth tracking as a potential shift in where AI compute for WordPress workflows actually runs. - **The WPOCC:** Directly relevant as the third data point in a sequence — survey finding (governance/consent objection) → keynote response (rebrand as "automation") → this post (community should learn more). Worth noting for WPOCC's own records that none of the three engages the opt-in/default/consultation mechanism the survey identified, while all three engage perception or understanding. - **FAIR:** The open-weight/local-inference emphasis is philosophically adjacent to FAIR's decentralization mission — worth a look at whether there's a real connection or only a rhetorical one. ## 🧪 Hypotheses 1. The "~20x token cost drop" figure, if traced, will prove highly sensitive to the chosen baseline model and timeframe — predicted: true directionally (costs have fallen dramatically) but not reproducible as a precise multiple without specifying which models are being compared across which dates. 2. Because this post answers a governance critique with an education program, community AI sentiment (per the WPOCC's planned 2027 survey) will not improve measurably from educational interventions alone, absent changes to defaults and opt-out mechanisms. 3. The GPL/open-weights analogy will resonate strongly with this specific audience but is technically incomplete — open weights grant modification rights without granting the training data, code, or compute needed to reproduce the model, which is a weaker guarantee than the GPL provides for source code. 4. Local-inference adoption among WordPress developers following this post will be measurable but small, and will concentrate among those already running local dev environments — i.e. the recommendation lowers a knowledge barrier, not an infrastructure one. ## 📊 Methodology A personal blog post, not research — an opinionated curated reading list with brief technical framing. Its factual claims about RLHF and post-training are accurate and uncontroversial; the abliteration description accurately summarizes the cited paper's finding. The one quantitative claim ("token costs have dropped about 20x") is stated without source, baseline, or timeframe and should be treated as directional rather than precise. The show-of-hands observations are anecdotal audience sampling from a single keynote, not survey data — which Mullenweg does not overstate. ## ❓ Questions - What baseline and timeframe produce the "~20x" token-cost figure, and does it hold for frontier models or only for the open-weight tier he's recommending? - Does the GPL analogy survive scrutiny — open weights permit modification, but without training data, training code, or compute, is the practical freedom comparable to what the GPL grants for WordPress source? - Is there a version of this AI push that addresses the WPOCC survey's actual finding (defaults, opt-outs, consultation) rather than community understanding — and does Mullenweg see those as distinct problems? - Why lead the "uncensored models" section at all, given the community-relations context? It's technically fascinating and the free-software argument is coherent, but it's an unusual choice for a post aimed at building trust with a skeptical audience. ## 🔮 Future Research - Trace the token-cost curve with explicit model/date baselines to establish what the defensible multiple actually is. - Track whether Learn WordPress or meetup programming actually adopts this curriculum, and whether participants' AI sentiment shifts (a natural experiment for Hypothesis 2). - Examine the open-weights-vs-GPL analogy rigorously — the "open weights are not open source" debate is live in ML circles and directly relevant to how a free-software community should reason about these models. ## 💡 Implications - **Practice:** The five-resource path is immediately usable by any WordPress developer wanting genuine AI literacy rather than tool familiarity — and the local-model exercise is the highest-value item for building intuition. - **Policy:** Notable that project leadership's response to documented AI skepticism is educational rather than procedural; whether that's sufficient is an open question this vault can now track across multiple sources. - **Security:** The abliteration section has an unaddressed dimension — removing refusal behavior has real safety implications, and the post treats it purely as a demonstration of modifiability without engaging what removed guardrails mean in practice. - **Education:** Genuinely strong curriculum design: tiered by time investment, vendor-neutral, ending in hands-on practice. - **Funding:** Not applicable to this source. ## 🧠 Learning Styles Assessment - **Visual:** Moderate — recommends a visual explainer as the entry point, though the post itself is text-only. - **Aural:** Moderate — the Karpathy video recommendation serves auditory learners; the post is short enough to read aloud. - **Verbal:** Strong — concise, quotable, with clear section headers per topic. - **Kinesthetic:** Very strong — the entire post culminates in a hands-on instruction (install, pull, talk to it tonight). - **Logical:** Strong — three parallel sections, each following the same structure: what the room didn't know → why it matters → what to read. - **Social:** Moderate — explicitly ends by directing learning back into meetup conversation. - **Solitary:** Strong — self-directed reading list requiring no coordination. - **Naturalistic:** Not applicable. Estimated reading level: technical professional, though notably accessible — explains post-training in two sentences without jargon. Short (~600 words). ## 📊 Accessibility Assessment (WCAG 2.2 AA) - **Perceivable:** Text-only post; no images requiring alt text in the main content. - **Operable:** Not applicable — static blog post. - **Understandable:** Genuinely good — explains RLHF and post-training in plain language before using the terms, and tiers recommendations by reader time budget. - **Robust:** Not assessable from this capture. - **Overall:** No accessibility concerns; the plain-language technical explanation is a positive example. ## 🌍 DEIB Assessment - **Cultural breadth:** Notable and worth crediting — the post foregrounds Chinese AI labs (Qwen, DeepSeek, Kimi, GLM) as frontier-competitive on their technical merits, in a Western tech-media environment where that framing is often absent or geopolitically hedged. - **Inclusive language:** No concerns identified. The "that's not a criticism" framing deliberately avoids shaming the audience for not knowing. - **Neurodiversity support:** The explicit time-tiering ("if you only have time for one thing") is a genuinely considerate accommodation for varying capacity. - **Representation:** Not a primary dimension of this source. - **Psychological safety:** The post's opening move — reframing blank stares as a reading list rather than a knowledge deficit — is a deliberate face-saving choice for an audience that had just been publicly quizzed, and worth noting as constructive. ## ✅ Logical Validity **Strong reasoning:** The technical explanations are accurate and appropriately compressed; the pedagogical structure (tiered by time) is sound; recommending competitors' models rather than his own company's partners is a credibility-earning choice. **Fallacies:** One unsupported quantitative claim ("token costs have dropped about 20x") with no baseline or source — directionally defensible but not verifiable as stated. One incomplete analogy: "open weights mean the community can modify these models, the same way the GPL means you can modify WordPress" elides a real disanalogy — the GPL guarantees access to source code sufficient to rebuild and understand the artifact, while open weights provide a compiled artifact without training data, training code, or the compute to reproduce it. The modification freedom is real; the equivalence to GPL is not. This matters specifically because the analogy is doing persuasive work with an audience that holds the four freedoms as a core value. **Tone of voice:** Warm, enthusiastic, teacherly; deliberately non-defensive about the audience's knowledge gaps. Considerably more constructive in register than the keynote's "rebrand" comment. ## 🛡️ DARVO Analysis Not applicable — no deny/attack/reverse-victim-offender dynamic. Worth noting the *opposite*: the post explicitly disclaims criticism of the audience ("That's not a criticism — it's a reading list"), which is a de-escalating move. The pattern worth tracking here is not DARVO but **reframing** — a documented governance objection answered as an education problem, which is the third such reframing across three sources now in this vault (survey → rebrand-as-automation → learn-more). Each individual response is reasonable in isolation; the pattern is what's informative. ## 🎯 Conclusion The most constructive artifact to come out of WordCamp US 2026 so far, and genuinely useful on its own terms — the reading list is well-chosen, vendor-neutral, tiered by time investment, and ends in hands-on practice rather than more reading. The GPL/open-weights analogy is the sharpest thing in it rhetorically and the weakest thing in it technically, and both facts matter. For this vault specifically, its value is as the third point in a now-legible sequence: the WPOCC survey documented an AI objection rooted in governance and consent; the keynote answered it by proposing a rebrand; this post answers it by proposing education. All three are now captured, and the throughline — that the response consistently addresses perception rather than mechanism — is the thing worth carrying into any future synthesis on WordPress AI governance. ## 🧱 Refactor Appendix ### Literature Split Pad (H3) #### Open weights are framed as continuous with the four freedoms > [!quote] ID:rw_atom1 > 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. Where/why: the post's central rhetorical move, mapping model modifiability onto a value the WordPress audience already holds — persuasive, and technically incomplete in a way worth naming. #### Refusal behavior is a removable post-training artifact, not an intrinsic model property > [!quote] ID:rw_atom2 > 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. Where/why: a compressed but accurate summary of the abliteration finding; important because it establishes that safety behavior in open-weight models is a coating rather than a structure. #### The cost curve claim > [!quote] ID:rw_atom3 > Token costs have dropped about 20x, and that curve isn't slowing down. Where/why: the post's only quantitative claim, offered without baseline or source — directionally defensible, not verifiable as stated, and load-bearing for the local-inference argument. #### The stated personal stake > [!quote] ID:rw_atom4 > It is really important to me that WordPress community members ride this AI wave. Where/why: the closing line, and the clearest statement of why the AI push persists despite documented community resistance — personal conviction about a missed-wave risk, not a product strategy. ### Atomic Split Pad (H4) **Answering a governance objection with an education program locates the problem in the objector rather than the process.** Claim: when a community's documented objection is procedural — about defaults, consent, and who was consulted — responding with a curriculum implies the objection stems from insufficient understanding, which leaves the identified mechanism untouched no matter how good the curriculum is. Evidence: the WPOCC State of the Community survey located AI-in-WordPress objections in opt-out and process design; Mullenweg's keynote response proposed rebranding AI as "automation"; this post, published the following day, proposes a reading list so the community can "be fluent in it" — three responses, all addressing perception or comprehension, none addressing defaults or consultation. Scope: applies where the objecting party has articulated a specific procedural mechanism; does not apply where an objection genuinely does stem from misunderstanding, in which case education is the correct response. The distinguishing test is whether the objectors' own stated reasons are procedural or epistemic. <!-- atomicity: 1 idea ✓ --> **Open weights grant modification rights without granting reproduction capability, making the GPL analogy directionally right and materially incomplete.** Claim: describing open-weight models as free-software-equivalent overstates the guarantee — the GPL provides source sufficient to understand, rebuild, and audit an artifact, while open weights provide a compiled artifact modifiable at the margins without the training data, training code, or compute required to reproduce or fully audit it. Evidence: Mullenweg's explicit framing that open weights let you modify models "the same way the GPL means you can modify WordPress," set against the abliteration example he cites in the same post — which demonstrates modification of behavior, not access to the process that produced it. Scope: applies when open-weight models are argued for on free-software grounds to an audience that holds those values; the practical freedoms open weights do grant (local execution, fine-tuning, behavioral modification, redistribution) are real and substantial — the claim is about equivalence, not about value. <!-- atomicity: 1 idea ✓ --> ### Fleeting Queue - Cross-link to the fireside chat (`01m0fk9689vfd849j1811w1g5n`), The Repository's WCUS recap (`01m0g8zg86npkqsw5zskd4xh8h`), and both WPOCC survey docs (`01m0danc26nyddhj3m1vvzh1dw`, `01m0d9sydk9d95mtfgpcv26es8`) — these five now form a single traceable sequence on WordPress AI governance and should be linked as a cluster at synthesis. - Actually do the homework: Karpathy's "Deep Dive into LLMs" is a strong candidate for this vault's own deep-read pipeline, and running a local Qwen via Ollama is a cheap, high-information experiment. - Look into the "open weights are not open source" debate for a rigorous treatment of the GPL analogy's limits. <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Homework From WordCamp.md|Homework From WordCamp by Matt Mullenweg]] ## Highlights - 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. ([View Highlight](https://read.readwise.io/read/01m0h2m6z23vp2ax0x91sf4rjj)) - Tags: atom/literature/h1 - 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. ([View Highlight](https://read.readwise.io/read/01m0h2m9z5z54nne2psqb1ya05)) - Tags: provenance/unsourced-figure, atom/literature/h2 - It is really important to me that WordPress community members ride this AI wave. ([View Highlight](https://read.readwise.io/read/01m0h2mdrx9qh6b9sdt85mn2ka)) - Tags: atom/literature/h3