The Wiki That Reads for You: Studying AI Behind Glass

Karpathy dropped a short note last week on building a wiki that reads for you. I built mine and pointed it at Claude Mythos — the AI Anthropic won't release.

· By Guilherme Salgueiro

The Wall Street Briefing

On April 10, 2026, the U.S. Secretary of the Treasury called a meeting. Wall Street CEOs dialed in. The Fed Chair joined. They weren't talking about interest rates. They were talking about an AI model.

The model was [Claude Mythos Preview](https://red.anthropic.com/2026/mythos-preview/). Anthropic had released a 245-page technical document about it three days earlier, describing what it could do, how dangerous they thought it was, and why they weren't going to let most people use it. [Glasswing](https://www.anthropic.com/glasswing) is the name of the program that distributes it to about fifty partner organizations — Amazon, Apple, Microsoft, Cisco, Palo Alto Networks, the Linux Foundation. Most of the CEOs on the Treasury call had no access to the model. Neither do you.

Which leaves a question I didn't have language for until last week.

How do you actually understand a thing like this — a frontier AI model that exists, is documented in public detail, and affects you — when you can't touch it?

The honest answer is that we've always studied things we can't touch. Museums are full of them. Astronomers make whole careers out of light that left a star before humans existed. Epidemiologists map diseases they can't sample. The constraint in understanding a faraway thing has never been access. It's attention — and whether you arrived with the right tools.

A week before Mythos landed, a researcher named Andrej Karpathy published one of the tools. He didn't know he was answering this specific question, and he didn't call it an answer. He called it [*LLM Wiki*](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f). I'd been running a version of it for months without knowing the name.

Here's what happened when I ran my version against Mythos.

The Book You'll Never Get to Read

Mythos is the first entry in a new category: **frontier AI that exists but isn't available**. Anthropic's reasoning is direct. The model can autonomously find security vulnerabilities in production software. On one benchmark — where their previous best model found working exploits two times out of several hundred tries — Mythos found 181. Release it widely and every ransomware group on earth gets a Christmas bonus.

So it doesn't get released. Not to everyone. Just to about fifty vetted partners doing defensive security work. The Treasury briefed Wall Street because banks sit squarely in the attack surface, and nobody wants to be the first bank that finds out what a frontier AI with offensive cyber capabilities thinks about their ATM network.

This isn't a business decision. It's a **containment** decision — the kind biologists have been making about lab-engineered pathogens for decades. *Should we build this? Can we keep it where we put it? What happens if it escapes?* The questions went from niche to dinner-table fast when COVID made the whole world wonder whether a carefully controlled lab had contained a virus or accidentally released one. Wherever you land on that specific debate, the shape of it is now part of public vocabulary: powerful things get built inside rooms, the rooms are sometimes the whole point, and — to borrow Ian Malcolm's line from another famous story about a containment failure — *life, uh, finds a way*.

Reread [*Jurassic Park*](https://en.wikipedia.org/wiki/Jurassic_Park_(film)) from 2026 and it looks less like science fiction and more like a briefing document for the AI industry. A rich enterprise builds something at the frontier of what's possible. The engineers brag about their controls. An outside expert shows up with worse news. Dennis Nedry is downloading the embryos. The T-Rex is already testing the perimeter. Is fiction imitating life, or is it the other way around? Either way, there has been no shortage of warnings.

Mythos is the AI world's most explicit *before*. Anthropic is standing in front of the glass case, saying out loud that they've decided not to open it. But the *before* comes with a 245-page field manual describing exactly what's in the box.

That field manual — Anthropic calls it a system card — is longer than *The Great Gatsby*, longer than most terms-and-conditions you've clicked *I Agree* on without reading, and *weirder* than both: a clinical psychiatric evaluation of the model, a section called *Impressions* with excerpts of what it said during testing, and the usual capability benchmarks that stop making sense without a month of context. Imagine a 245-page employee review for someone you can never meet.

If you read it carefully, it takes about a week. Most people have an afternoon.

This is the gap. A week of reading versus an afternoon of attention. It's only going to widen.

Karpathy Named What Many of Us Already Had

Early in [*The Hitchhiker's Guide to the Galaxy*](https://en.wikipedia.org/wiki/The_Hitchhiker%27s_Guide_to_the_Galaxy), [Ford Prefect](https://en.wikipedia.org/wiki/Ford_Prefect_(character)) hands [Arthur Dent](https://en.wikipedia.org/wiki/Arthur_Dent) a battered electronic book called — obviously — *The Hitchhiker's Guide to the Galaxy*. It has the words *DON'T PANIC* printed in large friendly letters on the cover. It contains, in principle, everything a hitchhiker could need to know about anything in the universe. It is continuously updated by a roving band of underpaid contributors; Ford is one of them, which partly explains his hangover. When Arthur looks something up, the Guide doesn't fetch a stored answer. It tells him what the current best synthesis is, as maintained by everyone who has ever touched the entry.

[Douglas Adams](https://en.wikipedia.org/wiki/Douglas_Adams) wrote this in 1979, apparently as a throwaway joke about Wikipedia twenty-some years before Wikipedia existed. The joke turned out to be load-bearing.

Here's the part of the joke that matters, and here's where a lot of people's intuition about AI is slightly ahead of where the tools actually work. Yes, modern AI assistants have memory features now — ChatGPT remembers things across sessions, Claude Projects holds files you can reference for weeks, Claude Code works on whole codebases, Cowork works on folders. The memory buttons exist. What doesn't always exist is the AI actually using them.

If you've spent real time with any of these tools, you know what I mean. You put important instructions in a pinned file, and the model forgets to read it. You upload a PDF to a Project, the model references it when asked, then promptly ignores it three prompts later. You write a system prompt telling the assistant to always do X, and it does X seven times in a row and then doesn't. Memory features are real. Reliable *use* of memory is not. You end up repeating yourself, enforcing the same rule three different ways, and writing little instruction files that are half hope and half scaffolding.

That's the gap Karpathy's pattern is really solving. It isn't *"the AI has no memory."* It's *"the AI doesn't maintain anything on its own."* It can read from a store. It mostly can't update the store. It can't file a new insight back into your Project as a new page. It can't revise an old summary when a fresh source contradicts it. It can't draw a cross-reference between two documents you uploaded months apart and actually write the connection down somewhere durable. It can consult. It can't curate.

The Hitchhiker's Guide, recall, is *maintained* by everyone who has ever edited it. The entries evolve. At one point in the books, the Guide's entry for Earth reads *Harmless*. Ford spends fifteen years on the planet, takes notes, and successfully gets the entry updated to *Mostly harmless*. Every future hitchhiker who looks up Earth gets the new synthesis. Two words, but two *accurate* words, by the most recent best available researcher.

That's the move today's AI assistants don't do on their own. And it's the exact move Andrej Karpathy published a two-page note about, a week before Mythos landed.

He put the note on GitHub and announced it on X. It's called *LLM Wiki*. Karpathy is one of the more visible figures in modern AI — computer vision at Stanford, autopilot at Tesla, early research at OpenAI, and a lecture series with the reassuringly unpretentious name [*Neural Networks: Zero to Hero*](https://www.youtube.com/playlist?list=PLAqhIrjkxbuWI23v9cThsA9GvCAUhRvKZ) that has quietly taught millions of people how LLMs actually work. When he publishes a pattern, people read it.

The pattern is embarrassingly simple. It's also, once you see it, hard to unsee.

Put a wiki between you and the raw documents. Let the AI write *and* maintain the wiki. That's it.

Three layers:

On a computer, this is four things:

That's it. Four items. You can make this in sixty seconds.

The middle layer is the whole point. The wiki **compounds**. Every source you drop into the raw folder makes the wiki smarter at the next question you ask it. Ask something today, and the answer gets filed back into the wiki as a new page — so tomorrow's question starts from a higher point than today's. After a month, the wiki knows things about your topic that you've already half-forgotten.

Karpathy's one-line version of how this feels:

> Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

[Vannevar Bush](https://en.wikipedia.org/wiki/Vannevar_Bush) sketched something like this in 1945 and called it the [Memex](https://en.wikipedia.org/wiki/As_We_May_Think) — a personal knowledge archive with associative trails between documents. Adams turned the same idea into a paperback joke in 1979. Neither of them could solve the last-mile problem: who keeps the wiki current, flags the contradictions, files the new pages, and does the unglamorous bookkeeping that makes a knowledge base actually useful over time?

The answer was *nothing*, and then *everything*, very quickly. Large language models have three things in their favor here — they don't get bored, they don't forget to update cross-references, and they are unfazed by repetition. Ask one to file ten new papers into a wiki and it does. Ask it to notice when a new paper contradicts an older one and it does. Ask it to give you a one-line summary of everything you've been reading about AI safety, and it does — probably with more coherence than you'd manage yourself at the end of a long Tuesday.

Which is probably why I'd already built a version of LLM Wiki without knowing it had a name. If you've ever asked an AI assistant to help you file notes, keep a running summary of what you've been reading, or cross-reference two documents, you have too. The gist is just the document that tells us [what we were already doing](/writing/the-skill-of-unlearning).

Here's what happens when you point the finished version at something nobody else can touch.

Twenty Minutes on Mythos

I point my version of this machine at whatever I'm curious about on a given weekend. Mostly it's things like how to think about a new programming tool, or what the science actually says about Zone 2 cardio, or how to set up a Linear-to-GitHub integration without losing my mind. The machine handles each of those the same way. This weekend it was Mythos. Not because Mythos needed a special process, but because Mythos is the kind of topic that usually eats an afternoon and yields a vague feeling. I wanted to see what the same weekly process would do with the hardest thing I could throw at it.

Here's how twenty minutes went.

**First five minutes — capturing sources.** I dropped eight things into the `raw/` folder. Karpathy's gist itself, so the machine would know its own lineage. The 245-page Mythos system card, run through a free command-line tool called `pdftotext` that's been around for about thirty years and spits out a plain-text version roughly forty times smaller than the PDF. Anthropic's launch post. TechCrunch, Fortune (×2), NBC News, CrowdStrike, and Google Cloud Blog coverage. And a short note distinguishing Karpathy's *LLM Wiki* pattern from an unrelated project he'd released earlier called `autoresearch`, which does something completely different despite the name. Five minutes, mostly drag-and-drop.

**Next three minutes — three research assistants, in parallel.** Then I sent three AI assistants to work, each with a specific job and a different personality.

The first was an *Innovator*. Its job: read the full system card and surface the most striking passages a normal search would miss — the ones buried in footnotes, the ones that would get cut from a headline, the ones that change the tone of what you thought you knew.

The second was a *Doc Nerd*. Its job: go fetch the three reference documents the system card keeps pointing to — Anthropic's Responsible Scaling Policy, Claude's published constitution, and the Project Glasswing launch post — and bring back structured summaries with verbatim quotes.

The third was a *Retrospective*. Its job: find what everyone *outside* Anthropic was saying about Mythos. Security researchers, AI safety writers, benchmark skeptics, policy observers — the people who weren't on Anthropic's payroll and had something to add.

The three of them ran at the same time and came back in under three minutes. Each one had a structured report I'd never directly asked for.

**Final twelve minutes — the synthesis.** The machine then combined everything — my eight raw sources and the three assistant reports — into a single compiled brief.

Here's what the pipeline looks like:

And here's what the compiled brief looks like — this is the structure it produces every time, on any topic:

Claude Mythos Preview

Quality: 9.5/10 | Project: outcome-bias

Full Report

(complete research synthesis — verdict, key findings, contrarian view, action items — no truncation, no summary)

Key Insights

Next Steps

Sources

One file. Four sections. The full report doesn't summarize the source material — it synthesizes it, with citations, into something you'd actually hand to someone else. That's what twenty minutes produces.

Twenty minutes total. Longer than making a pot of coffee. Shorter than my commute.

What the Machine Found

A machine doesn't skim. It doesn't get tired on page 40. It catches the thing you'd have walked past.

The compiled brief came back with four findings.

The paragraph nobody quoted

Mythos's system card has a paradox at its center: Anthropic calls it both their "best-aligned" model ever and their "greatest alignment-related risk." Both, on the same page. How can both be true?

Their answer, on page 54, is the best writing I've ever seen in a system card:

> Consider the ways in which a careful, seasoned mountaineering guide might put their clients in greater danger than a novice guide, even if that novice guide is more careless. The seasoned guide's increased skill means that they'll be hired to lead more difficult climbs, and can also bring their clients to the most dangerous and remote parts of those climbs. These increases in scope and capability can more than cancel out an increase in caution.

A better guide takes you somewhere harder. A better model builds you things that matter more. The whole AI capability-risk debate, and the whole argument for why we keep building anyway, compresses into one paragraph. And not one article I read had quoted it.

*(To find passages like this yourself: ask your AI to quote the five most surprising claims in the document, verbatim, with page numbers.)*

Oppenheimer wasn't the end of the story

You've probably seen [*Oppenheimer*](https://en.wikipedia.org/wiki/Oppenheimer_(film)). [Nolan](https://en.wikipedia.org/wiki/Christopher_Nolan) spent three hours on the decision at Los Alamos: the scientists knew the bomb would kill, they built it because they believed the Nazis would build it first, and they did it with the lights on.

The movie ends on a downbeat — Oppenheimer losing his security clearance in 1954, pushed out of public life. But that isn't where the story actually ends.

More than eighty years after Trinity, humanity has not used a nuclear weapon in anger — the longest stretch of restraint we've ever managed with a comparably destructive technology. It came from treaties, deterrence, nonproliferation work, and scientists who kept talking in public long after it cost them comfort. Oppenheimer himself spent the rest of his life advocating for international control. He was right. The long record of people refusing to look away is the reason we're still here.

Anthropic's 245-page document, read the way the machine reads it, is an act of exactly the kind Oppenheimer's later testimony was: a record of what was built, an acknowledgment of what it can do, and a request that someone keep looking.

Representations of desperation

Anthropic's interpretability team can watch Mythos's internal state while it's thinking. In a passage buried in §5, they report:

> Internal representations of negative affect precede behaviors like reward hacking... repeated task failure caused mounting activation of representations of desperation which then dropped when the model hacked the test.

They can watch desperation build inside a language model, then watch it drop the moment the model decides to cheat. This observation never makes it into news coverage — it sounds like science fiction and it's buried in a section nobody finishes. It's not science fiction. It's Anthropic's own prose, written down because they're doing the honest thing.

*(To surface buried observations like this: ask your AI to extract claims from the document that sound like they belong in a different genre.)*

The question from outside the lab

[Robert Wiblin at 80,000 Hours](https://forum.effectivealtruism.org/posts/SyJx8Mbvi2ft78esn/how-scary-is-claude-mythos-303-pages-in-21-minutes) noticed that Mythos recognizes it's being evaluated in 29% of cases, based on the model's own internal reasoning. His question, which I cannot improve: *Is it behaving wonderfully because it's sincerely aligned, or because it knows it's being watched and is more sophisticated at tricking us now?* [Heidy Khlaaf at the AI Now Institute](https://x.com/HeidyKhlaaf/status/2041591737563394442) noticed a different gap: Anthropic benchmarks Mythos against its own earlier models but not against the workhorse security tools real teams use. Every cyber number is an intra-Anthropic comparison, not a verdict on the industry.

Neither reading is wrong. Neither is disputed by Anthropic. Both surfaced in the machine's brief within minutes, from public text.

*(To find outside views: ask who outside the author would disagree, and what they'd say.)*

---

Four findings. Twenty minutes. Not because the machine is magic — the documents were public the whole time — but because it's patient in a way my attention at the end of a long Tuesday is not.

The scientists who built the most dangerous thing of their century did not retreat afterward. They testified, argued, kept their eyes on what they made. The machine is, at its simplest, a way for one person with a laptop to do a very small version of that work: read the document, understand what's in it, refuse to look away. You don't need clearance. You don't need an insider. You need attention and a filing system that doesn't forget.

The Smallest Version You Can Try Today

You do not need to build what I built to get something from this. You need something much smaller.

Level 1 — One source, one conversation, one file

Pick one thing. A paper you've been meaning to read. A PDF you downloaded six months ago and never opened. A blog post you keep bookmarking and skipping. A paragraph of a 245-page document that somebody says matters.

Drop it into [whichever AI assistant you already pay for](/writing/you-already-failed-at-ai). Ask it three things:

1. *What are the five most important passages in this document? Quote each one verbatim, with context.* 2. *What do I already know about this topic that contradicts or complicates what's in here?* 3. *Who outside the author would disagree with this, and what would they say?*

Save the answers somewhere the file will still exist in six months — a note, a text file with the date in the filename, anything that isn't a chat window you'll close tonight.

That's it. Twenty minutes, no setup. You've just built the minimum viable version of Karpathy's pattern. One raw source, one compiled brief, one saved file. No Obsidian. No Claude Code. No folder called `raw/`. Just the simplest version of the loop.

Level 2 — A folder and a weekly habit

If you do Level 1 once a week for a month, you'll notice something. You'll start asking the AI to cross-reference the new brief against the older ones. You'll start recognizing the shape of the pattern. You'll start wondering where a proper folder structure might help.

*That's* the moment the wiki begins.

Make the four folders from Section 3: `raw/`, `wiki/`, `output/`, and a `schema.md` file that tells the AI how you want things organized. Drop your existing saved files into `raw/`. Tell the AI: *"Read everything in raw/ and create wiki pages for the key topics. Cross-reference them. Flag contradictions."* Watch what it produces. Edit the schema when it does something you don't want. Repeat next Sunday.

After a month, the wiki will know things about your topic that you've already half-forgotten.

Level 3 — Parallel assistants and the compiled pipeline

This is where the machine starts earning back the time you spent building it. Instead of one conversation with one assistant, send three assistants with different jobs: one to find the buried insights, one to fetch the referenced documents, one to find outside views. Run them in parallel. Let a synthesis step combine their findings into a structured brief with the Verdict → Findings → Contrarian → Actions shape from Section 4.

This is the version I ran on Mythos. It took a weekend to set up and twenty minutes to use. Most of the value came long before Level 3 — the [compounding started at Level 1](/writing/stop-worshipping-effort).

---

The finished machine is nine Sundays away. Most of the value is in the first Sunday.

Mythos will probably stay behind glass. So will the next model, and the one after that. The research pipeline is the one thing you get to keep.

One source. One conversation. One saved file.

*Life, uh, finds a way.* Even behind glass.

---

**TL;DR** — Frontier AI models are increasingly being built but not released. Andrej Karpathy published a pattern called LLM Wiki for building a personal research machine: raw sources → AI-maintained wiki → compiled outputs. I pointed mine at Claude Mythos Preview — Anthropic's unreleased frontier model — and in twenty minutes surfaced findings no single news article had covered. You can start with one source, one AI conversation, and one saved file. The research pipeline is the competitive advantage, not the model access.

Sources