Notices by Eaton (eaton@phire.place), page 4

@tess did a big comparison between document categorization approaches with dozens of models and techniques. without doing any particular tailoring or optimization, simple kmeans proximity with embeddings scored within 10% of the best prompt based llm approaches. Orders of magnitude faster and more energy efficient, too. It’s hard to imagine why one wouldn’t start by optimizing that to improve results, rather than endless prompt fiddling.. but admittedly it took relearning some tangly math

@nora OW, but also yes

@christina @RuthMalan Some of the original Pattern Language stuff about design of physical spaces might be useful (though I suspect you’ve already mined that given the similar phrasing!)

@TALlama ah! No, as @DireVole mentioned, it’s “Advanced Persistent Threat,” the idea of some party that isn’t just opportunistically rattling the door on a network to see if it’s unlocked, but consistently and purposefully trying to exploit it on an ongoing basis

@tess and — although I didn’t say it earlier, I think this is what’s really digging at me — you can’t create a good community without owning up to what you believe is good.

The neutrality and objectivity and customizability and federation etc etc is (sometimes) a way of sidestepping that, and that’s where the really exploitable cracks in the foundations can be found by those who want to abuse them

@tess i think that anticipation and navigation is exactly what I’m trying to articulate, and the approach to quoting is a great example. It’s not “Quoting is per se bad, because we see what it leads to,” rather (as you said) an ongoing assessment, anticipation, and engagement to change parts of the system that have become attack vectors or turned into negative incentives.

@tess absolutely agreed. I want to be clear that I don’t think architecture and system design is unimportant, rather that they aren’t the fate of a network or its insurance.

I think a previous generation of us bought into the idea that sufficiently well crafted rules enforced “correctly” could solve it; now I see a lot of insistence that architecture (federation, personalized algos, composable moderation) will do so.

Over the years I’ve come to the conclusion that a safe and inclusive social network simply isn’t something that can be created by architecture or feature selection, any more than something like “allyship” is a specific set of voting positions.

If you have a place where people congregate you have to accept that you will have to evolve it in realtime to address threats. There are anti patterns you can avoid, but abuse and exploitation are APTs, not bugs.

@RuthMalan oof, I need to finish my templating on the blog — the underlying data it’s using has all kinds of atuff about where the talk was given, etc. it’s from Confab, in early 2023!

@jmeowmeow @RuthMalan yup. That quest to find and articulate the useful stuff beyond “works on my machine/for my project/in my org” feels like it can be a lifetime’s work, lol

@RuthMalan man, your stuff is always such a treasure trove.

And I hear you about evolving — one of the reasons I’ve been building more of an archive of older writing and talks is how easy it is to lose the layers of change, and just remmeber the “conclusion” of 10y of iterating on a subject…

@landley this has so many roots of the stuff I was trying to get at in my “institutional doctrine” talk… thank you for reposting it here! https://eaton.fyi/talks/the-doctrine-gap/

@N0ty3p @Cdespinosa wtf, people think ONE IN FOUR workers makes half a mil a year

@ireneista @xgranade genuinely treasure the old print copy I got a while back. It’s like holding my formative years in my hands

@so_treu the blacksky starter pack has dropped

@adrienne @craftcms the ACF and Gutenberg support is WILD AF

@andrewhinton @SwiftOnSecurity well, “more than a few” still leaves enough room for my entire 20s, sooooooooooo :D

@dachary but yeah, A+ for Qwen out of all the local models https://vega.github.io/editor/#/url/vega-lite/N4IgJAzgxgFgpgWwIYgFwhgF0wBwqgegIDc4BzJAOjIEtMYBXAI0poHsDp5kTykBaADZ04JAKyUAVhDYA7EABoQAEzjQATjRyZ289AEEoUBuqRQAngAJS6iAwiWdCOJYDu8WZahJM5NpoAvGlkyS2U2Y2dZTAcmK0w2HBooRRUfFFRQE0E0DGw8QgJTV2o6RiZ7OHUoOV9oyhqEAjgfOSK2CDhZADM4QWUiuG6IAngkZRHkYII2Blw5ggQ2VUF+M2NTCwaIYhAAXyVXGmV6NAA2AAYLpXgaMiw0ABYrpRwkUwQINABtUFkkZy5BJkMiCOAAMQBNEE5lSnTBUEwaFAmHMODguRwbGCSKU3RofQmPxA3ShMJAAF0DiAmMFlLkwWQuvS9hSlF0aspgmRkSAAB68-GE3IIYJzNSpVHo3IARwYSGidB8NFIcO8YN5ASqbDQpMEnSUUox6AgMvUSL21NhmRJBP6uXWJjMsMNaONIDlCp0mGVqqU0CQGptWvUOtQeoNKjYU1kAFkkALUGILpbDWw2IIdDgfqAhfb0EsVpK3bl-Fz-jlDXQgyBY8s+vsFLm7fT0I7Ni6QN6a4YNs7G83hQWfLAJVXMDX45hR18DoP8yAqqHbJLq+6AKLqZezpu2ocgUWycVfceTsW+BwAFSQAGsugO9wuahBcV217kAMIdTCWYKWAAiGaCO8s5siAiRmHQ1qgDUshcjoci8m8HxAmwIJgpCorkkoxCBgwxoXJQACc1K4YI+FoIRABM1IQDASDSjaeatiSZKdkauSyNGwSBo2IA1II-iCi2QJwLAsg0HKGKuoxIBcYevGpiAwHmFUOYHu8N68hx6BYjicI0FqaAABwpqySiwfiPI2i+5hBqAwFMA2NqBnceggJo9yvkwSDwsE7qisoygakoyiJvwYihda-DGZalpAA/view

@dachary Qwen 2.5 is the best of the local models in my testing for the same purpose — and the same gap w/gtp4o mini. The biggest thing that jumped out for me was to stay alert for taxonomy decisions that rely on intentions outside of the documents themselves (ie this is for beginners). Nice to see you’re seeing similar results!

@skinnylatte I… I want noodles in every meeting *sob*