THIS BLOG IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THIS BLOG OR THE USE OR OTHER DEALINGS IN THIS BLOG.
You guys know you can just make a website right
also you can just buy a web domain and it’s like ten dollars per year
fanfiction.net has STILL not changed their dns fuckup where ‘fanfiction.net’ gets you a landing page and www.fanfiction.net gets you the actual website. it’s been like two weeks now
YO i was about to make a comment about how the buprenorphine/suboxone prescription cap laws in the US are bullshit but then i looke dit up and they did away with it. sweet
“Oh dear. How sad. Never mind.”
“Over time, mistakes in generated data compound and ultimately force models that learn from generated data to misperceive reality even further,” wrote one of the paper’s leading authors, Ilia Shumailov, in an email to VentureBeat. “We were surprised to observe how quickly model collapse happens: Models can rapidly forget most of the original data from which they initially learned.”
this delights me, because the solution to this “problem” is to eliminate the reasons why we’re pissed about “AI” generation to begin with
if you don’t use it to make things for publication, they don’t enter the data set, problem solved. if you don’t scrape the entire internet as your unlicensed data set, or you curate the input, you can control what gets fed into the algorithm, problem solved.
Curated AI reminds me of CYC, a project started in 1984 to encode all human conventional wisdom in such a way that future AIs could understand the world the way humans do. Basically, to imbue AI with common sense.
CYC isn’t an AI itself, more like a massive database of concepts and rules. And all of it curated by researchers; humans typing stuff into CYC to build the database by hand. Concepts like:
- Mountains are big
- Mountains are outside
- Mountains cannot fly
- California is big
- California is a place
- California cannot fly
- Airplanes can fly
- People can be inside an airplane
…etc. And once all those and a million more rules are defined, CYC can parse a sentence like “I saw the mountains flying over California” and correctly deduce that I was in an airplane, looking down on mountains as I flew someplace. Not that the mountains were flying over California.
And each of those rules can be examined, edited, changed, or deleted. Unlike generative AI, you can ask CYC to explain its reasoning and it is able to do so, in complete detail. Its creators know how CYC works.
Nor is it prone to just make shit up, because it’s not generating anything… it’s applying rules.
And yeah, it’s taken 30+ years to build that database and at 2+ million rules it’s not even close to being done. But unlike large language models trained on the shitposts of a billion strangers, it’s stable. And already genuinely useful.
Of course, a lot of the AI community loathes CYC and considers it a stupid dead end, mostly because it has been 30+ years without much visible progress. CYC is not sexy. It’s oldschool “brittle” AI technology, coded by humans! Kind of like building a moon rocket one rivet at a time, assembled by hand. How absurd!
Oh wait, they did that.
i think i might host a kbin instance, which is fediverse software that is kind of reddit-y forum-y kind of like i want. not precisely what i want, but adapting a more forumish experience on activitypub seems to inherently involve a slight abuse of the protocol
i honestly think we should just bring back usenet in lieu of activitypub based reddit clones but it seems to really only be me and some real old computing grognards that are thinking this

