A Decent Subset of Human Knowledge

July 14, 2026 — James Henry

When I was a teenager some university kid came round the neighbourhood selling encyclopedias door to door, one binder of sample pages and a payment plan. I remember it more as a novelty than a temptation--because by then I'd already decided the CD-ROM version sitting in a computer at the school library did more for me than a shelf of hardbacks ever could. This was pre-Google, pre-Wikipedia, and I didn't have the words for it at the time, but the print set had already lost. Search beat an index years before anyone built a search engine worth using.

I do remember the other kind of browsing, though--as a kid, flipping through our single-volume encyclopedia looking for one thing and surfacing twenty minutes later with three unrelated ideas I hadn't gone looking for. That's real, and I miss it (Wikipedia's "On this day" works well as a starter). But it was never about the knowledge being hidden. It was about the accident of the page next to the page you wanted.

I went back to a public library recently with my kids. I hadn't set foot in one for my own reasons in years, and I was startled by how little of it was books. There was a fraction of the shelf space I remembered from my own childhood at that same branch. I asked an LLM about it first, out of habit, then asked the actual librarian to confirm--and she did, plainly: demand for physical books has cratered. What's left is the building--the meeting rooms, the programming, the third place. Physical books have gone second class even at the library because they stopped being the first-class way humanity stores its knowledge a while ago. That happened quietly, and it happened before any of this current AI cycle started.

The Move to Digital Lending

Roughly thirty-three countries run some version of Public Lending Right--paying authors when their books circulate through public libraries. The UK's version is the cleanest: 12.4 pence per loan, capped per author, funded by a government grant of about £6.5 million a year, paid separately from library budgets and reader fees entirely. Canada and Australia pay differently--on holdings rather than loans--but the shape is the same: the state, not the reader, foots the bill. The United States, home to nearly every frontier AI lab on the planet, has never had one. An American author gets nothing when a library lends their book.

Compare that to what happened to library ebooks. HarperCollins capped ebook licenses at 26 checkouts back in 2011, a policy still in force. Other publishers expire licenses after 24 months regardless of use. Libraries pay something like three to five times the consumer retail price for the privilege, then repurchase the same title when the license lapses--and the American Library Association has spent years calling this exactly what it looks like.

I think you can tell that I don't think that PLR is the hero of this story. Paying an author every time a book circulates, forever, funded by taxpayers instead of a toll booth, is still rent--income from continuing to own a thing you finished making, not payment for having made it. I've made that argument before--rent-seeking on something you didn't personally create, that costs nothing more to reproduce, is parasitism no matter how nicely you dress it.

The Frontier Model Toll Booth

The US already wrote the two-tier system into law. The Framework for Artificial Intelligence Diffusion, issued in January 2025, created an export control classification--ECCN 4E091--covering closed-weight model weights trained above roughly 10^26 floating-point operations. Open-weight models are explicitly carved out. This isn't a metaphor about gatekeeping. It's a customs code. The government drew the exact line this piece has been drawing: past a certain capability, synthesis becomes a licensed substance, and below it, it's still free to walk out the door.

Then, in June 2026, the line got tested in public. Anthropic pulled two new models, Fable 5 and Mythos 5, offline within days of unveiling them, under a directive restricting foreign access during a government cybersecurity review. OpenAI capped its own new model, GPT-5.6, to roughly twenty government-approved companies. OpenAI's own statement on the arrangement: "We don't believe this kind of government access process should become the long-term default." Nobody involved thinks this is permanent. Everybody built it anyway.

A month later, the same instinct showed up on the other side of the Pacific. Reuters reported that Beijing met with Alibaba, ByteDance, and Z.ai to discuss--reportedly, not yet as policy--a tiered regime for their most capable models, with the most sensitive ones possibly barred from public release entirely. Nothing's been enacted as of this writing. But the fact that it's being discussed at all, in the same season the US was doing the equivalent, is the point. This is happening while Chinese labs are the ones currently dominating the open-weight space--DeepSeek, Qwen, Kimi, GLM, all shipping under permissive licenses, all matching closed frontier models on benchmarks. Two governments who agree on almost nothing else reaching for the same lever in the same season isn't a coincidence of politics. It looks structural.

Smaller states have noticed too. Canada's 2026 "AI for All" strategy names sovereignty as one of three pillars and promises "sovereign compute"--a country deciding that depending on someone else's byproduct is itself a risk worth building against.

Inference to Humanity Was Never the Mission

Somehow we've all quietly assumed that the business model for these AI companies is Inference and Data Centres: that OpenAI and Anthropic et al. are in the business of handing humanity an oracle. We should take a moment to read what they actually say they're for.

OpenAI's charter: the mission is to ensure that AGI "benefits all of humanity," and the organization commits to using any influence it gets over AGI's deployment to "avoid enabling uses of AI or AGI that harm humanity or unduly concentrate power." Anthropic: build AI "for the long-term benefit of humanity," working to ensure "the benefits of AI are broadly and evenly distributed across society."

Reading those again, both are promises about a future event--AGI arriving, its benefits eventually reaching everyone. (AGI, Artificial General Intelligence, by OpenAI's own definition: "highly autonomous systems that outperform humans at most economically valuable work.") Those charters make no promises about whether you get to use this month's LLM for your homework. Nobody ever actually committed to affordable, ubiquitous, individual inference. I'd bet most of us users assumed it.

I'm not making the case that either company is lying about its charter--that's already been reported by people better positioned to make it. I'm saying that it's on humanity to remember that these companies are not aiming at providing a distillation of human knowledge to the commons.

Renting LLM inference to the masses is a byproduct of their goal, not the goal itself.

The New Kind of Knowledge

The idea of walking into a publicly owned building and freely accessing the history of knowledge was the allure of the physical library. The promise that knowledge was, at least in principle, available to all is honestly one of the greatest things to happen in the history of humanity. With the advent of LLMs you can have what is essentially the distillation of all of humanity's knowledge running on your laptop--and it can talk with you about how well you've understood it.

LLM training isn't one author's book getting read without permission. It's the aggregate--the accumulated, undifferentiated output of everyone who ever wrote anything that made it into a training run. You can put a price on that the way you price a stolen novel--Anthropic's own settlement over pirated books worked out to about $3,000 a title, across roughly 500,000 books. It's so much more than that, 'tho. Every Substack post, Reddit post, the good, the bad, the meh--maybe even my forum threads from the 2000s--are all in there.

A model whose weights you've downloaded doesn't need the company that trained it to keep existing, to keep serving you, or to keep being willing. It will do the same reasoning on whatever hardware you point it at in a hundred years that it does today. That's not a claim about who can afford the compute; compute gets cheap on its own, that part takes care of itself. It's a claim about ownership. For the first time, the thing worth owning--not the storage of information, the capacity to synthesize it--is a thing you can actually own outright, permanently, with no further toll, the moment the weights sit on your disk. A library card was never that. A purchased book was closer, but a book can't answer a question you haven't thought of yet or talk to you cogently about any of 100,000 other ideas. This is the first time "yours, forever, no rent" has applied to something with this much reach.

Now, in July 2026, a model you can download and run on your laptop is not the same as a so-called "Frontier" model like Claude Fable 5 or GPT-5.6 Sol. But a decent one already sits on my home server, unrentable, un-revocable, done.

LLMs Are and Aren't Magic

This is one moment in history when "own this outright, forever, no further toll" became possible for something this valuable--a downloaded set of weights that will reason exactly as well as it does now in a hundred years and contain a decent subset of human knowledge at a specific snapshot in time. It could help people do their Linear Algebra homework in 2220 just as well as it helped me in 2022.

That is sort of magic.

James is a security engineer with an overflowing bookshelf and, now, a directory of GGUFs he's hoarding. This post is signed and verifiable.