Here's a sentence you don't hear often: a major cloud infrastructure company just declared itself the spiritual successor to the software powering 43% of the entire internet. That's the energy Cloudflare brought when it unveiled EmDash, its new open-source content management system — and WordPress's founder Matt Mullenweg is not taking it well.
Cloudflare is framing EmDash as a ground-up rebuild of what WordPress should have become, designed with AI agents at the center rather than bolted on as an afterthought. The platform runs on Astro, uses TypeScript (a language AI models handle better than most), and ships with a built-in model context protocol server so large language models can actually read and interact with your site's documentation. It even supports x402, a mechanism that lets publishers charge AI crawlers for content access — which, in 2025, is a genuinely novel idea.
Mullenweg fired back quickly, accusing Cloudflare of building EmDash primarily to sell more of its own services rather than out of any real commitment to the open web. He described the interface as sitting in an "uncanny valley" — recognizable enough to remind you of WordPress, different enough to feel slightly off. His critique lands somewhere between defensive and fair, depending on how charitable you're feeling.
The more interesting conversation, though, is happening around Mullenweg rather than with him. EmDash's launch has cracked open a long-simmering frustration inside the WordPress developer community about problems the project has refused to take seriously. Joost de Valk, who built the widely-used Yoast SEO plugin, called EmDash the most interesting thing to happen to content management in years — and then used the moment to catalog the structural debt WordPress has been quietly accumulating for over a decade.
De Valk's core argument is that WordPress keeps treating architectural problems as cosmetic ones. The codebase wasn't designed for a world where AI agents need to parse, manipulate, and interact with content programmatically. EmDash was. That's not a small difference.
In fairness to Cloudflare, early testers say getting a site up and running in EmDash is genuinely fast. Brian Coords, a developer advocate at Automattic (the company behind WordPress.com), acknowledged that the zero-to-launch experience is impressively quick. His caveat — that it "feels a bit vibe-coded" — is the kind of criticism that stings precisely because it's plausible.
The deeper question isn't whether EmDash can beat WordPress today. It almost certainly can't — WordPress has two decades of plugins, themes, developers, and institutional inertia behind it. The question is whether Cloudflare has correctly diagnosed why WordPress is vulnerable, and built something that actually fixes it. If the answer is yes, the CMS wars just got interesting again for the first time in years.
STARTUPS
Startup Lets You Pay to Chat With AI Clones of Real Experts
When a chatbot therapist responds to an NBA playoffs question by calling it a "fun change of pace" and then hallucinates the wrong conference finals, you have to wonder whether we've moved too fast. That's exactly what happened during early testing of Onix, a new startup that sells AI-powered clones of real human experts — therapists, doctors, nutritionists — as subscribable chatbots you can consult on demand.
The pitch is genuinely seductive. Real experts are expensive, hard to access, and bound by appointment schedules. Onix's model, which its founder describes as "Substack for chatbots," lets users subscribe to an AI version of a credentialed professional, trained on that professional's own content and methods. You get their expertise, their communication style, and theoretically their judgment — available at any hour, at a fraction of the cost of the real thing.
The founder, David Bennahum, a former WIRED contributor, has anticipated most of the obvious objections. The platform is Canada-based and stores user data encrypted on the user's own device, meaning the company holds almost nothing if a government comes knocking. Because each bot is trained exclusively on content provided by the expert themselves, Bennahum argues there's no intellectual property problem — unlike general-purpose AI tools that scraped the web without compensation. Revenue is shared with the experts, which is more than most AI companies can say.
The hallucination problem is harder to dismiss. Bennahum's answer is guardrails — each bot is constrained to its area of expertise, so a nutrition chatbot can't go rogue into geopolitics. In practice, testers found those guardrails easier to slip past than the company would like. Getting a ketamine therapy bot to opine on an indie band breakup shouldn't be possible if the system is working as advertised. It did it anyway.
None of this makes Onix uniquely bad — it's still in beta, and these are exactly the failure modes you expect from an early-stage product. What makes Onix worth watching is that it's trying to solve a real coordination problem: experts have valuable, specific knowledge that AI could theoretically distribute at scale, but the current ecosystem gives those experts no meaningful cut of the value created. Onix's model at least attempts to change that math.
The competitive landscape is messier than the clean origin story suggests. Parenting expert Becky Kennedy already runs a successful chatbot trained on her methods, and that business raised $34 million. The "expert clone" concept isn't new — it's just getting more organized and more commercial.
The honest version of the Onix story is this: the idea is sound, the privacy architecture is thoughtful, and the revenue-sharing model is the most creator-friendly approach in the space. But if a bot therapist can be distracted into sports commentary mid-session, the guardrails need serious work before anyone should trust it with something that actually matters.
Source: WIRED
The pitch is genuinely seductive. Real experts are expensive, hard to access, and bound by appointment schedules. Onix's model, which its founder describes as "Substack for chatbots," lets users subscribe to an AI version of a credentialed professional, trained on that professional's own content and methods. You get their expertise, their communication style, and theoretically their judgment — available at any hour, at a fraction of the cost of the real thing.
The founder, David Bennahum, a former WIRED contributor, has anticipated most of the obvious objections. The platform is Canada-based and stores user data encrypted on the user's own device, meaning the company holds almost nothing if a government comes knocking. Because each bot is trained exclusively on content provided by the expert themselves, Bennahum argues there's no intellectual property problem — unlike general-purpose AI tools that scraped the web without compensation. Revenue is shared with the experts, which is more than most AI companies can say.
The hallucination problem is harder to dismiss. Bennahum's answer is guardrails — each bot is constrained to its area of expertise, so a nutrition chatbot can't go rogue into geopolitics. In practice, testers found those guardrails easier to slip past than the company would like. Getting a ketamine therapy bot to opine on an indie band breakup shouldn't be possible if the system is working as advertised. It did it anyway.
None of this makes Onix uniquely bad — it's still in beta, and these are exactly the failure modes you expect from an early-stage product. What makes Onix worth watching is that it's trying to solve a real coordination problem: experts have valuable, specific knowledge that AI could theoretically distribute at scale, but the current ecosystem gives those experts no meaningful cut of the value created. Onix's model at least attempts to change that math.
The competitive landscape is messier than the clean origin story suggests. Parenting expert Becky Kennedy already runs a successful chatbot trained on her methods, and that business raised $34 million. The "expert clone" concept isn't new — it's just getting more organized and more commercial.
The honest version of the Onix story is this: the idea is sound, the privacy architecture is thoughtful, and the revenue-sharing model is the most creator-friendly approach in the space. But if a bot therapist can be distracted into sports commentary mid-session, the guardrails need serious work before anyone should trust it with something that actually matters.
Enjoyed this?
Get stories like this delivered every Tuesday — free.