Measurability is not a guarantee
Even when the diagnosis is technical, not political, the gap between understanding a problem and being able to fix it is the same.
Weekends are the right time for research. After a long walk with Ugo, some cleaning, and stocking the kitchen, I used the machine to see what's happening out there. Among the various scans, I found a really interesting piece by Walter Quattrociocchi, a professor at Sapienza University in Rome. It ran in an Italian newspaper, Corriere della Sera, and the LLM turned out useful again, because it gives me a genuinely good explanation of the economic mechanism underneath the pattern, not just its recurrence.
It's not just that whoever holds power chooses not to build enforcement: it's that verification has a structurally different cost curve than generation, and that asymmetry alone is enough to explain why the gap keeps showing up everywhere, regardless of the good faith of whoever is generating. Anthropic's self-certification on bioweapons, for instance, doesn't count as safety for exactly this reason: a model trained to refuse is not the same thing as a system that physically cannot execute the action.
The gap between generating a diagnosis and building verification, Quattrociocchi suggests, gets solved with "access controls, authorizations, independent oversight, clear accountability" - forgetting who pays for that, who enforces it: the same companies generating extinction-hype to sell themselves as exceptional have no interest in funding the expensive verification infrastructure that would slow down their release cycles. For Quattrociocchi, extinction fear is just marketing, because, he argues, the percentages are subjective and not calibrated against a historical record the way climate change is - a "real" threat because it's backed by decades of measurable data. But that's the same criterion - measurability - that I've been describing for a while now as the proxy whoever holds power chooses to decide what counts.
Quattrociocchi gives me a fifth independent proof of a pattern I've already documented four times from different angles: surveillance, biosecurity, European reform, mathematics - but his generation/verification distinction is a sharper, newer lens for rereading the earlier episodes, especially Anthropic's self-certification.
I just finished my own small essay, which I'll be publishing soon on Amazon, titled Eyes Only AI. I hadn't decided yet which language to launch it in, but Quattrociocchi gave me the right idea to test the large language model in this one too.