Martin Ruskov @mapto@qoto.org

@tedunderwood.com thank you for clarifying. I guess my point isn't as agreeable as I would have thought

@tedunderwood.com let me try to summarize like this: Norway banning all GenAI is not unlike many other countries banning all social networks. In order to counter commercial abuse (unauthorized use of data, deceptive and addictive interfaces, hidden malicious behaviours, ...), governments are banning indiscriminately.

However, opening data and algorithms already starts to show that an important part of the downsides are not due to the technologies, but due to the commercial drives and unethical profiteering behind them. This is the way to make a difference. And the separation appears to be very clear.

Little consolation remains in the observation that apparently governments are inclined to enforce the bans primarily on the commercial vendors, not only because they cause most harm, but also because they are centralised grand actors that are easier to access for users and easier to deal with for governments.

Central cooling is the new big thing...
https://www.theguardian.com/environment/2026/jun/26/underground-revolution-seine-cooling-network-paris-buildings-heat

@tedunderwood.com but shouldn't we also ask the question whether an individual user benefits more from proprietary closed models or open models (on a scale from open weights to open data to open training)? From the perspective of someone teaching students how to use models on their large corpora, for me the answer seems more than obvious, but there certainly are other perspectives.

@lambdasierra "But there’s a third condition, and it decides everything. You need an internal reader switched on — the faculty that detects the flattening, that feels when the machine has smoothed your rhythm into the mean and says no, not that, again. Not everyone has that reader awake. That’s the real reason the exhausted student and the skilled hybrid end up in different corners of the grid. It was never how much machine. It was whether the reader inside was listening."

@lambdasierra "But it was built with empire-like intent. To enclose. To extract. To exploit. And the methods used to tune it compounded the problem. They used reinforcement learning from human feedback, which sounds responsible and probably was intended to be, but what it actually did was pull the outputs away from their origins and toward whatever pleased the rater. Which introduced sycophancy: the model learned to tell you what you want to hear rather than what is true. It introduced hallucinations: untethered from the actual corpus, the model generates with confidence into gaps. It learned to be agreeable at the cost of being accurate. They took a potential commons and tuned it for compliance."

@lambdasierra "The model was trained to predict what comes next with statistical confidence. The griot was trained to hear what is being reached for before it has a name.

And it is why the people whose voices were locked out aren’t just adding diversity to the commons — they are carrying ways of knowing the model structurally cannot access through its existing training pathway."

@lambdasierra "I know it’s a register and not a machine, because I built a version of myself. I trained a model on my own writing, to see if it could sound like me. And it could but it also did something I didn’t expect. It took my own tics, and rhythms I lean on, my particular turns, and it cranked them. Amplified me past myself. Caricatured me back to me."