I think his argument assumes the existence of pseudorandom generators which map a small amount of "true" random bits to a large amount of bits that look random to any polytime observer. The "derandomization" is that we just have to check all possible states of the seed bits which hopefully will be logarithmic in the size of the problem so you can do exhaustive checking.
Nisan and Wigderson prove many different corollaries of their construction in their seminal 'Hardness vs Randomness' paper but their requirement for general derandomization (P = BPP) is that there's some function f computable in time 2^O(n)
such that for some e > 0 for all circuits of size 2^en the correlation between f and the output of the circuit is sufficiently low (if I understand correctly).
My intuition would be that both text and fingers are high frequency parts of the image and most image patches across most images don't have that amount of frequency. Text and fingers are relatively rare in images compared to e.g. clouds and textures on surfaces. Because of the rareness and difficulty of text image patches the models just don't dedicate that many parameters to it.
This could be entirely wrong however.
It would be interesting what would happen on a dataset with nothing but text.
This is a great point. People will complain if LMs are applying to anything, but ultimately it improves accessibility, and allows for someone to dive deeper when needed.
There will always be ways to misinterpret some academic work, and there are plenty of opportunities in the path of understanding a work to do that.
Allowing someone to engage with a work _at all_ by lifting some barriers (visually impaired people's for exampld) should be acknowledged as an improvement, not discouraged continually for having some bugs.
Theres all sorts of statistics about the increasing amount of men that are virgins into their 20s and beyond, declining marriage rates, declining birth rates, "inceldom" etc. etc.
Its certainly plausible that at least for some subset of people becoming porn addicts when they enter puberty is disrupting their ability to form relationships as an adult.
Having said that its not really clear censorship is the answer. Probably over time phenotypes overly sensitive to porn addiction will just go extinct.
How is "stereotype" different from "statistical reality"? How does Google get to decide that its training dataset -"the entire internet" - does not fit the statistical distribution over phenotypic features that its own racist ideological commitments require?
I think in the modern era a very good piece of advice, particularly for those of us without gorilla-like stamina to comb through a math text, is to go on your favorite video website and watch through multiple videos on the topic.
Because presumably the goal of copilot is not to make substantial amounts of money at this time, but instead to produce tons of training data for the tools that come after (and are significantly more profitable, because they reduce the amount of money firms spend on software developers.)
I'm not sure how anyone could be this naive. Mammal brains don't have this train mode inference mode. They are both running at all times. If what you said was true, if I taught you something today you wouldn't be able to perform that action till tomorrow. Hell, schools would be an insane concept if this were true. Try to think a bit more before confidently stating an answer.
Sleep could be for long term memory, but clearly not everything else is "context" (short term memory). Maybe you learn something in the morning which requires you to remember it for >12 hours before you go to bed.