btw this is essentially what we're doing with LLMs, and it's why everything costs so much money. that little bump in the Florida panhandle is some obscure subfield of chemistry only a few dozen PhDs know exists.
This is a map on why video games take up 5x the space they did 15 years ago just to look *slightly* more detailed.
Positive users praised the analogy between LLM training and precise U.S. map edge rendering as fantastic or amazing, while negative users felt intellectually outdone by the post.
No Digg Deeper questions have been answered for this story yet.
btw this is essentially what we're doing with LLMs, and it's why everything costs so much money. that little bump in the Florida panhandle is some obscure subfield of chemistry only a few dozen PhDs know exists.
How many edges does the U.S. need before it looks right?
@jxmnop "How large is the decision boundary of your LLM?" https://en.wikipedia.org/wiki/How_Long_Is_the_Coast_of_Britain
@jxmnop @plasmarob it’s worth it
@jxmnop Half the oysters in the South come from that little bump pal
@jxmnop this makes a ton of sense to me but is this not bearish? That LLMs are just curve-fitters? A compressed representation of existing human knowledge? An object that’s extremely useful but has a hard limit?
(personally I want the bear case to be right, the bull case is too scary)
@TsarinaBomba But boto is from a different bump
@jxmnop Amazing analogy. Are you a fan of the quanta hypothesis: https://ericjmichaud.com/quanta/
Seems reminiscent.
@jxmnop it’d probably be like comparing 60 to 250 if you’re talking about open source to frontier right
@jxmnop if you squint the us looks like the deepseek whale 🐋
@jxmnop the honest answer is that even for those who've spent years studying llns, working with the best talent and pouring millions into r&d, there's still so much unknown about what truly makes them tick and how to scale effectively
@jxmnop PhD-inaccessible knowledge bottlenecking everything while we scrape the bottom of the generalist barrel story of modern engineering
@jxmnop Damn I was just opening twitter to learn something new, not to get mogged
@jxmnop Is but not aught - this isn’t how we do it
@jxmnop i kinda agree but the really big llms are undeniably better at highly relevant tasks like coding than the smaller ones so far, right? whereas the change from HD -> ultra HD in a video game is much less impactful
@jxmnop Fantastic analogy
@jxmnop There's no composition so there's no decomposition