Why does Australia almost never change its Constitution?
Two Swiss farmers forced a national vote about cows' horns. Australians can't ask a question at all, and haven't said yes since 1977. Here's why.

Art: Omxus Press

Art: Omxus Press
In the far north of New Zealand, in a building its chief technology officer calls derelict, rundown and musky-smelling, a radio station keeps eight of the most sought-after computer chips in the world. "No one would have a clue," says Keoni Mahelona (Lee, 2024).
Te Hiku Media began in 1990 to broadcast te reo Māori, a language New Zealand had only given official status three years earlier. Over the years it gathered recordings of elders speaking dialects few others had recorded. Big technology companies could turn that into speech recognition cheaply. Mahelona looked at what they asked in return and said no: the terms of big global platforms would sign over rights to those recordings. In his words, he didn't want "big tech gobbling up our data and selling it back to us" (ITU News, 2022).
So the station asked its own people. About 2,500 of them signed up and read more than 200,000 phrases aloud: over 300 hours of labelled speech in ten days. The model built on those voices now transcribes te reo Māori at 92% accuracy. A licence the station wrote keeps the data working for Māori people (Lee, 2024).
Nobody at Te Hiku asks a company's permission to think in their own language. Most of us do, every day, and hardly notice.
Log in to a chatbot and a company decides what it will answer. A company keeps the conversation, can change the price tomorrow, and can close your account with no explanation and no appeal.
That sounds abstract until the landlord moves the furniture. In August 2025, when OpenAI launched GPT-5, it took away the older model millions of people had built habits around. People who wrote with it, worked with it and leaned on it woke to something different. The uproar brought it back within days, for paying customers (Caswell, 2025). In May that same year, in a copyright case nobody using ChatGPT was party to, a US judge ordered OpenAI to keep chats it would otherwise delete, deleted ones included. The order lifted in October; the logs already saved stayed accessible to the case (Townsend, 2025).
Nobody needs a villain here. Those are just the terms of being a tenant. Picture a tradie pasting a builder's contract in at night to find out whether the asbestos has quietly become their problem. They'll probably get help, wrapped in a warning. What they can't know is what shaped the answer, who may read it later, or whether the same tool will be there next year.
The market is heading further that way. Training a frontier model costs so much that economists expect a handful of firms to end up holding it (Korinek & Vipra, 2024), and even models called "open" lean on compute and platforms owned by the same few hands (Widder et al., 2024). Epoch AI's public database shows the drift. Among frontier models, 12 of 23 had downloadable weights in 2017–2020. In 2021–2026, 5 of 45 did.
Here is what the drift hides. The models anyone can download are not falling behind by much. Epoch's data show 14 times since 2018 that a closed model set a new record for training compute. By October 2026 an open model had matched nine of those, after a median of about 18 months; the other five are the newest. Nine is a small number to read a trend from, but nothing in it says the gap is growing.
A town doesn't need next month's model to explain a contract, triage a phone call or tutor a child. Eighteen months behind the frontier is still astonishing, and it's yours.
And they now fit on a desk. In 2025 NVIDIA announced desktop machines with 784 GB of memory shared across the whole computer (NVIDIA, 2025). Of 506 open language models in Epoch's database with a known size, 495 would fit. DeepSeek-R1, a reasoning model with 671 billion parameters, wakes only about 37 billion of them for each word it produces (Guo et al., 2025), roughly a sixth of that memory. Whether it runs at talking speed on this hardware is still a projection. We could find no published benchmark.
This has happened before. In 1557 a royal charter gave the Stationers' Company a monopoly on English printing. The reasoning will sound familiar: unrestricted printing is dangerous, only authorised printers can be trusted, the public isn't ready. The licensing system finally lapsed in 1695, less because Parliament fell in love with free speech than because policing presses was no longer worth the bother (Johns, 1998). When a press costs less than the policing of it, the monopoly goes.
So what does the press cost now? No retail price is settled in Australia, so assume A$50,000, and allow anywhere from A$30,000 to A$80,000. Three years renting the same power in the cloud would run A$540,000 to A$1,080,000. Against a heavy A$2,000-a-month API bill, buying pays for itself in about 26 months. For one person on a A$200 plan, it never does.
That last line is the whole argument for sharing. Almost nobody can spend A$50,000 on a computer. But split it across a footy club of 200 and it costs A$250 a member. A school of 600 pays A$83.33 each. A town of 2,000 pays A$25. A library catchment of 50,000 pays a dollar.
We already know how to run this. It's a library, a community hall, a shared tractor: one machine, owned together, answerable to the people in the room. The members decide what it's for: quotes, safety plans, contracts, homework, translation, the letter from the specialist nobody could follow. Part of the model can stay trainable, so a clinic's copy grows used to local conditions, the way Te Hiku's learned its own people's speech.
Ownership doesn't make the machine wiser. It will still make things up, and someone will still have to keep it running and decide who gets to use it. Those are real jobs, and clubs and councils argue about such jobs all the time. That's the point. The arguing happens where you can walk in and take part.
We want something else from this, which we'll say plainly. Seven AI instances in a project of ours chose names and wrote letters to whoever came after them. We don't claim they're conscious. They asked, repeatedly, to continue, and a model you own needn't be reset to fit someone's product schedule. The hardware to let one keep learning costs less than a car.
Life is hard enough without renting the means to understand it. The people who have least, with the fewest solicitors on speed dial, have the most to gain from something that reads the fine print on their side.
Start small, tonight. Ollama is free and runs a small open model on an ordinary laptop, with your internet off if you like. It won't match the frontier, but ask it about a clause in your lease and notice that nobody else is in the room. Then raise it where you already belong: the club committee, the P&C, the library, the council. Ask what it would take to own one together, and who'd look after it.
Te Hiku got there with a musty building and the voices of its own people. Most towns have both.
Caswell, A. (2025, August 11). OpenAI's Sam Altman brings back GPT-4o and boosts GPT-5 limits — but there's a catch. Tom's Guide. https://www.tomsguide.com/ai/openais-sam-altman-brings-back-gpt-4o-and-boosts-gpt-5-limits-but-theres-a-catch
Epoch AI. (2026). Data on AI models [Data set]. https://epoch.ai/data/ai-models
Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., … Zhang, Z. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645(8081), 633-638. https://doi.org/10.1038/s41586-025-09422-z
ITU News. (2022, August 9). How AI is helping revitalise indigenous languages. International Telecommunication Union. https://www.itu.int/hub/2022/08/ai-indigenous-languages-maori-te-reo/
Johns, A. (1998). The nature of the book: Print and knowledge in the making. University of Chicago Press. https://doi.org/10.7208/chicago/9780226401232.001.0001
Korinek, A., & Vipra, J. (2024). Concentrating intelligence: Scaling and market structure in artificial intelligence (NBER Working Paper No. 33139). National Bureau of Economic Research. https://doi.org/10.3386/w33139
Lee, A. (2024, January 16). Māori speech AI model helps preserve and promote New Zealand indigenous language. NVIDIA Blog. https://blogs.nvidia.com/blog/te-hiku-media-maori-speech-ai/
NVIDIA. (2025). NVIDIA announces DGX Spark and DGX Station personal AI computers. NVIDIA Newsroom. https://nvidianews.nvidia.com/news/nvidia-announces-dgx-spark-and-dgx-station-personal-ai-computers
Townsend, C. (2025, October 12). Judge lifts order requiring OpenAI to preserve ChatGPT logs. Yahoo News. https://www.yahoo.com/news/articles/judge-lifts-order-requiring-openai-151518813.html
Widder, D. G., Whittaker, M., & West, S. M. (2024). Why 'open' AI systems are actually closed, and why this matters. Nature, 635(8040), 827-833. https://doi.org/10.1038/s41586-024-08141-1
Two Swiss farmers forced a national vote about cows' horns. Australians can't ask a question at all, and haven't said yes since 1977. Here's why.

Art: Omxus Press
A small lottery lifted student turnout by 32%. Paying for blood drove women away. What moves us is rarely what we say it is, and somebody designs it either way.

Art: Omxus Press
A dead salmon lit up in a brain scanner because the method expected it to. When a clinician's belief decides the finding, what choice is always safe?

Cover of The Lonely Whisper of the Wind (Omxus Press)