Here’s what happened, for anyone who missed the drama unfold on their timeline.

Thariq Shihipar, who works on Claude Code at Anthropic, published a piece with a title that reads less like a blog post and more like a diss track: “The Unreasonable Effectiveness of HTML.” His argument, stripped of the niceties: he’d stopped reading his own AI’s Markdown output once it crossed a hundred lines, and he couldn’t get anyone else in his organisation to read it either. Four million views in sixteen hours later, Andrej Karpathy — a man whose day job has involved teaching computers to see — showed up to agree with him. And he brought a diagram.

Karpathy’s case is basically neuroscience with a grudge. Roughly a third of your brain is wired for processing what you see. He calls vision the human brain’s highest-bandwidth channel — a superhighway, in his words, next to which Markdown’s toolkit of bold text, headers and the occasional table looks like a footpath. His progression runs: plain text (effortful), Markdown (today’s default, tolerable), HTML (early, but already forming the better default) — and eventually, several steps out, something he only half-jokingly calls interactive neural video. We are, apparently, one generation of models away from our status reports dreaming at us.

Now — I’ve lived my working life in the Microsoft stack. I have built SharePoint wikis that a project team opened once, at kickoff, and never again. I have shepherded forty-slide decks whose entire purpose was to survive a steering committee meeting before being quietly filed under Final_v3_ACTUALfinal.pptx. If Markdown has a midlife crisis coming, the consulting deck should already be checking its pulse.

Here’s the part that actually matters for those of us in consulting and enterprise knowledge management, because the format war is entertaining, but the implication underneath it is bigger. Markdown, decks and Word docs are cousins of the same idea: knowledge frozen the moment someone hit save, then mailed around as an attachment you need the right tool to even open. HTML breaks that. It’s a link, not an attachment — it opens in the browser everyone already has running. It can hold a real diagram, a colour-coded risk table, or a slider your client actually drags instead of squinting at a frozen number on slide fourteen. And here’s the bit I didn’t expect: the same HTML file a person reads is the file an AI agent reads back later, as its own working memory of the decision. It stops being a report. It becomes a surface both of you keep working from.

There’s a second reason this isn’t only a human-readability story, and it’s the one that should worry knowledge management teams more than the first. AI doesn’t just write HTML well — it reads HTML back better than almost anything else you hand it. A PDF is, structurally, a page of positioned ink; extracting a table means an agent reverse-engineering the fact that fifty scattered text fragments happen to line up in columns. A PowerPoint slide is XML underneath, but it’s a bag of shapes sitting at x/y coordinates with no defined reading order — nothing tells an agent that this box is the heading and that one’s a footnote somebody forgot to delete. Markdown is flatter still: headings, bullets, bold, and not much else to express a relationship like “this row is the high-severity one.” HTML is the outlier, because it was never built to look pretty first — it was built to be a tree. Every tag nests inside another, every table row is genuinely tied to a header, and screen readers have been parsing exactly that structure for people with visual impairments since long before anyone trained a language model on anything. AI agents didn’t invent a new way to read documents. They just showed up to a format that had already solved the problem.

→ don’t take my word for it

Below is the single most-produced artefact in enterprise consulting: the solution architecture diagram. Same system, four formats — in the order history gave them to us. Watch what happens to the diagram as you move right. Then, on the last tab, click the components. Try doing that to slide 14.

+------------+       +-----------------+      +----------------+
|  Channels  | ----> |  Copilot Studio | ---> |  Azure OpenAI  |
|  (web chat,|       |      Agent      |      |  (UK South)    |
|  Teams,    |       |                 |      +----------------+
|   voice)   |       |                 | ---> | SharePoint KB   |
+------------+       |                 |      +----------------+
                     |                 | --> | Dataverse / D365 |
                     +-----------------+      +---------------+

To be fair to Markdown’s defenders — and there are furious ones — HTML isn’t free. It costs more tokens to generate, it’s messier in version control, and giving an AI agent free rein to write JavaScript into your organisation’s knowledge base is the kind of decision that ages badly if nobody’s reviewing it. This isn’t “burn every deck you own by Friday.” It’s “stop assuming static and unreadable is the safe, neutral choice.” It never was — we just didn’t have an alternative that an AI could produce for the price of asking.

Earlier in my career I’d have called this a UX problem. Turns out it always was one. We just needed the AI to write the interface for us, so we’d stop pretending bullet points were a design system.