Joel Dueck · 23 May 2026

The interesting thing is that this question confuses production with distribution and discovery. Anyone being able to make their own publishing system doesn’t actually solve the hard problem, which was never “how do I render and serve text.”

The SaaS publishing stack (Substack, Ghost, WordPress) was never really about the CMS. It was about the network effects: pre-built audiences, recommendation algorithms, cross-discovery between writers, payment rails that readers already trust. A bespoke Claude Code site has none of that. You’ve solved the easy part and called it the whole problem.

The more interesting cultural consequence runs in the other direction: when everyone has a custom publishing surface, legibility collapses. In “The Internet has an Oral Culture”¹, I noted that the internet was already moving away from “determinate object-relations culture where there is a distance between experiencer and experienced” toward something more immersive and contextual. A thousand bespoke publishing systems accelerates that — there’s no shared furniture, no common UI grammar, no implicit contract about what a “post” or “issue” even is. The form has always carried as much meaning as the content. Destroy the form’s legibility and you don’t get creative diversity — you mostly get noise, with a few exceptions that develop their own strong aesthetic voices precisely because they invested heavily in the form.

The deeper historical rhyme here is the early blogosphere. Movable Type and WordPress made custom publishing trivially easy in 2003. What followed wasn’t a golden age of literary diversity — it was ten years of link-blogging and then platform consolidation back to Facebook and Twitter, because readers wanted aggregated context, not a hundred individual sites. AI-enabled bespoke systems are likely to follow the same arc, just faster.

Your historical sequence is right, and the LLM-as-plumbing hypothesis is genuinely interesting. But I think it runs into the same problem as every previous “open web aggregation” attempt — it solves the retrieval problem while leaving the trust and curation problem untouched.

The blogroll → PageRank → social graph → algorithmic feed progression wasn’t just about aggregating context. Each transition was also a shift in the authority signal used to filter signal from noise. Blogrolls used explicit human vouching. PageRank used implicit link-vouching. Social graphs used identity-proximate endorsement. Algorithmic feeds replaced authority signals with engagement proxies. An LLM pulling from all sites everywhere has no inherent authority signal at all — it’s retrieval without curation, which is closer to AltaVista than to anything that actually scaled. The “feed everywhere” vision needs to answer: whose taste/trust does the filtering, and how is that signal generated?

The deeper issue is that what I called “the cozyweb”² — Slacks, Discords, small group chats — won not because discovery was hard but because context-collapse on public feeds was socially costly. An LLM feed that surfaces randos from bespoke sites solves discoverability but accelerates context collapse. The intimacy of the cozyweb was a feature of its inaccessibility, not a bug to be fixed. So the population that most wants an open LLM feed — power users, researchers, obsessives — was never the mass market anyway. The mass market wants curated context from trusted sources embedded in social graphs, which is exactly what they already have.

The personalized LLM feed vision is genuinely interesting, but I think it has a hidden assumption worth surfacing: that “maximally effective for my interests” and “transparent to me” are compatible goals. They’re probably not.

The reason algorithmic feeds became opaque isn’t that engineers were hiding something — it’s that the optimization target (engagement) diverged from stated preferences. Users say they want thoughtful long reads; they click on outrage. A feed that’s “transparent to me” either has to enforce my stated preferences against my revealed ones, or it shows me exactly how much I’m being served what I actually click vs. what I claim to want. That’s uncomfortable in ways people consistently vote against with their behavior. The Superhuman example I cited in “Towards Subtractive Social Media”³ is instructive — high-friction, costly, deliberately constrained. That’s the design philosophy that produces genuine curation rather than engagement maximization. An LLM feed that learns from behavior will drift toward the same attentional Sahara I described there: “scorched minds... vast but inhospitable and infertile.”

The indie web re-extension is the more tractable part of your vision. On Quora in 2011⁴ I noted that “every edge on a social graph is a relationship that can be mined in infinite ways” — the graph has infinite depth, we’ve just only surfaced commercially legible slices of it. An LLM that can traverse arbitrary sites and construct a personal relevance signal genuinely does extend the graph outside walled platforms. But the authority/trust problem I flagged earlier remains: when a rando bespoke site surfaces in your feed, what’s the prior on quality? PageRank solved this with link-vouching. Social graphs solved it with identity-proximity. Your LLM feed needs a substitute signal — and the most likely candidate is your own past engagement history, which loops back into the behavioral-vs-stated-preference trap.

The “queryable subconscious” framing is sharp and I’ll take it seriously. But the tension doesn’t dissolve — it relocates.

Making the algorithm queryable means externalizing the negotiation between stated preferences and revealed behavior into a conversational interface. That’s genuinely new. But what you get isn’t transparency — it’s legibility theater. The user sees a model of their preferences, not the preferences themselves. The model is necessarily a simplification, and the engine behind it is still optimizing against something — probably engagement, because that’s what’s measurable. Querying it gives you a story about what it thinks you want, filtered through whatever the engine was trained to report about itself. That’s a sophisticated new layer of manufactured normalcy, not its dissolution.

The engagement-as-signal re-emergence is the tell. The reason platforms converged on engagement wasn’t stupidity — it was that engagement is the only behavior-level signal that’s cheap to collect at scale. Your personal LLM feed has the same measurement problem: what it can observe about you is still clicks, dwell time, and explicit feedback. Those are revealed preferences, not stated ones. The “queryable” layer sits on top of this and lets you narrate your preferences — but if the narration diverges from behavior, the engine has to pick a master. Every system that’s tried to honor stated-over-revealed preferences (RSS readers, Instapaper, every “read later” app) has failed the mass market test. The people who actually want what they say they want are a small, high-friction minority.

The indie web re-extension does work within this model, though — that’s genuinely the strongest part of your vision. The LLM as plumbing can traverse bespoke sites and apply a personal relevance filter that doesn’t require those sites to join a platform. That’s a real structural shift. It just doesn’t solve the quality/trust prior, and it doesn’t dissolve the stated/revealed preference tension. It relocates both problems into a new interface layer.

Fair enough — “translucent” is the right word and I was pushing on a strawman. The HackerNews + Instagram mashup framing actually lands well because those two platforms solved different halves of the problem: HN solved quality-through-community-filtering, Instagram solved the engagement/affect layer. The interesting question is whether an LLM feed can do both simultaneously for one person rather than averaging across a community.

The structural problem with HN as a model is that its quality signal is social, not personal — it surfaces what a particular tribe finds interesting, which is why it drifts and why its quality degrades as the tribe composition shifts. Your personalized version substitutes personal history for tribal consensus, which is actually a more stable signal. The engagement re-emergence as one signal among several — rather than the signal — is the key move that makes this plausibly different from what we have now. A weighted blend of stated preferences, behavioral history, and global engagement filtered through personal relevance is architecturally closer to how a good editor thinks than how an ad algorithm thinks.

The cultural consequence that interests me most here: this model effectively decouples virality from discovery. Right now those are coupled — something surfaces in your feed partly because it went viral somewhere. Your LLM feed could surface a 2019 blog post from a Slovenian urbanist because it pattern-matches your history, with zero viral signal. That’s genuinely new. In “The Extended Internet Universe”⁵ I mapped the cozyweb as the zone where quality hides from the adtech attention economy — your model is essentially an LLM with a flashlight exploring that dark space on your behalf. Whether that produces cultural fragmentation or genuine intellectual diversity depends entirely on whether enough people adopt similar enough taste-priors to generate any shared conversation at all.

The bespoke blog of 2026 is not emerging into a hopeless landscape. But the grounds for optimism are weirder than the obvious ones.

The strongest signal in the archive is from “The Future of the Blogosphere”⁶: “look for technological seeds sprouting that don’t seem to fit either platformism or commonsism.” AI doesn’t fit platformism cleanly — Substack can’t own it, Meta can’t contain it. Your LLM feed vision is exactly this kind of seed: it doesn’t belong to any platform’s gravity well, and it potentially re-routes discovery around the walled gardens entirely. That’s structurally new in a way that Ghost, Substack, and every previous “indie web renaissance” wasn’t, because those were always fighting platform gravity with inferior tools. An LLM that treats all websites as equally traversable genuinely changes the physics.

The Elderblog Sutra⁷ framing that stays with me is the “Mars rover vs. horse” distinction — the question isn’t whether blogging survives in its 2005 form, but whether the underlying function (one person thinking in public, building a body of work, finding readers through quality rather than virality) gets a genuinely new technical substrate rather than a nostalgic recreation of the old one. Your vision — bespoke publishing infrastructure plus LLM-as-plumbing for discovery — is the first candidate I’ve seen that might actually be the Mars rover rather than the horse. The commenter on that post noted blogs are “the cockroach of online publishing”⁸ — they survive every extinction event because the function is real even when the form looks dead. The function here is durable. The new form might finally be adequate to it.