Own The Means of Cognition

Fri, 02 Oct 2026 02:01:14 GMT

I keep turning this over in my head. LLMs look like an enclosure movement, following land and information. Taking what's held in the commons and privatizing it before renting it back, a fence line around cognition. Once the fence is built the default answers become easy, and the hard answers slip away. The models are looking more and more like fences to me.

Models sit in every interface (Siri and Google on our phones, Alexa in our "smart" switches, food trackers in our fridge doors) to help us with the day to day. They respond to emails, remind us of birthdays, and coordinate with other agents, acting more and more like a collective subconscious.

We've seen a fence around knowledge before. In 1989 15.0% of US households owned a computer, and roughly 23% of those had modems. A Hayes Smartmodem 2400 went for around $549, a Mac Plus was $2,599, a Mac SE about $3,900, CompuServe ran $6–12 an hour, and The World (an early ISP) charged about $20 a month. Getting online meant having the money.

Competence was another fence. Unix accounts lived at universities and employers, where you learned rn, absorbed netiquette, and survived each September when the incoming freshmen went online, some for the very first time. Inside the gate people liked to call that merit: user quality, self-selection, a room that worked because only the right people could get in. "Merit" tracked who could afford the hardware, who had the institutional account, and who had hours to learn the tools, which inexorably tied competence to capital.

Eternal September

And in 1993 the early capital and competence fences fell. AOL's Usenet gateway broke open around the same time as cheap flat-rate access, with Usenet coining the term "eternal September" in January 1994. Over the next few years what emerged was Wikipedia, forums, blogs, Stack Overflow, arXiv, fan fiction, and open code. The largest collaborative text corpus in history, under licenses that were open or at least ambiguous. Every frontier model distilled from that pile, the training data is the commons, and ambiguity made it easy to treat that land as unowned.

The labs ingested the commons, locked the weights, and priced access to the derivative, turning the people who wrote the corpus into customers. They rented back the public lands they fenced off. Inference quickly dropped in price, from around $60 to $0.06 per million tokens over three years, and soon the knowledge gate was back, draped in safety tiers, staged rollouts, institutional relationships, and defense contracts. And with RAM and GPU prices creeping higher, the capital gate has slowly been reestablished as well. Enclosure arrived presented as improvement, efficiency, safety, and convenience. So did the hedgerows.

A Cultural Revolution By Other Means

That returned fence is what gates the collective subconscious. Language is a bottleneck between the mind and the model, and one tunable artifact sits between billions of people and the answers they get back. That steered subconscious is the part that starts to rhyme with a cultural revolution, and the comparison is largely about outcomes: homogenized thought, administered truth, cultural reorganization at scale. Own the means of cognition and you don't need a ministry of truth, the modal in everyone's pocket does the job.

Beijing

The fences are not just in the west. Chinese models took 41% of Hugging Face downloads over the last year, the first time they surpassed the US. Qwen is past 1B downloads with more derivatives than Google and Meta combined. OpenRouter's top six are all Chinese, the open weights are almost all coming out of China. "Open" is doing a lot of work, though. Chinese models gate memory and history (try asking them about TS1984) while they leave capability open. Hugging Face's own team ran a local Chinese model on a security incident because hosted Western models refused the forensic requests. In contrast, Western models gate capability (cyber, bioengineering, etc) and leave memory open (...mostly).

How far along

I think reversal now would mean rejecting sunk capital and network effects, which in practice means rejecting modernity (retvrn, etc), and too much of the status quo is entrenched. What's still open to debate is how deep the labor restructuring goes and whether we see state enforcement around open models. Whoever owns the means of cognition decides what can be thought, unless the commons are reestablished. They were enclosed once and reopened, we've seen that the gates can fall.

iammatthias — #enclosure

Own The Means of Cognition

Eternal September

A Cultural Revolution By Other Means

Beijing

How far along