кадмій
banner
kadmii.bsky.social
кадмій
@kadmii.bsky.social
cringe and sincerepilled | Proverbs 30:2
indeed. more hits means stronger token association. it uses word frequency as a proxy for relevance. same with it being a reflection of average thought. not just the thought of an average person, but the average in every thought. even an average person can be especially knowledgeable in a topic
December 10, 2025 at 4:46 AM
I mean, not really? A great deal of knowledge is fragmented between people. A lot of books got digitized via OCR or via reCAPTCHA to provide single-word transcriptions. A lot of journal articles are not read. Many books exist that nobody alive has read
December 10, 2025 at 4:11 AM
in short, our short lifespan and forgetfulness becomes a strength. Conversely, the LLM has to rely on a recency bias in its training data, since there are so many more word generated now than previous centuries, to achieve some approximation of that
December 10, 2025 at 3:55 AM
to a degree, but for us, our consensus is weighted by a combination of peer connections, the value given to expert authority, etc, which can suppress ideas that have been discarded or never caught on. For the LLM, consensus is determined by weighing its mass, the degree to which it shows up and how
December 10, 2025 at 3:54 AM
not just current consensus, but the consensus found in the entire body of its training material. Today's consensus. Yesterday's consensus. Last decade's consensus. 19th Century consensus. Whatever is fed into it for training
December 10, 2025 at 3:00 AM
me when we're about to change a poopy diaper
December 10, 2025 at 1:06 AM
that's all well and good, but there is no accountability mechanism with AI services, and the product it produces seems to be especially effective at tricking people, or else we wouldn't have had so many reports of people dissociating from reality back in April from a "bad update"
December 9, 2025 at 10:23 PM
having raised 2 kids, I can confidently say that it's a bad idea to do everything the Word Predicting Machine tells you to do, especially when just this year, ChatGPT has told people:
1. they're God
2. to eat poison
3. to kill themselves

it's an ethics nightmare
December 9, 2025 at 9:45 PM
you should see the interview when he is asked about the strange death of a former OpenAI employee. I'm not saying he did anything, but he answered it like someone with something to hide. I think he just always talks like that, though
December 9, 2025 at 9:14 PM
"I'm sorry you're right. Doing that would make your child a sociopath. You should do this instead:"
December 9, 2025 at 9:13 PM
the lore has gone deeper
December 8, 2025 at 2:44 PM
love the completely unnecessary gag of the scientist wincing and cradling his arm in its sling
December 8, 2025 at 2:38 PM
the tummy most risky to rub is the tummy most worth rubbing
December 4, 2025 at 2:37 PM
that is badass, like lava made flesh
December 4, 2025 at 2:30 PM
truly is a "tough guy shows he tough by declaring he can beat up someone he can't" moment
December 3, 2025 at 2:38 PM
when you're trying to prevent imminent brain death
December 1, 2025 at 7:50 PM
I never appreciated PBS Kids until I had my own kids, their programming is so wonderful
November 24, 2025 at 7:50 PM
man, I'd love to have a house rep whose constituency consists of 3 blocks of skyscrapers and then 10,000 square miles of cows and corn
November 18, 2025 at 9:39 PM
lots of thin districts that slice up some areas in order to connect two ends together
November 18, 2025 at 9:31 PM