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.
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.
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.
I think 8 months is a little short for the utility of a new tech to be fully realized and utilized. I'm pretty sure there were still horses on the roads long after 8 months after the Model T first went on sale.
I've somewhat had this phenomenon, with the added observation that if I listen to the recording multiple times it will "click" and will sound good again. Then I wonder if it actually sounds good or what.
Recurrent neural nets is the general term for nets with memory as you describe. Indeed LSTMs, a type of recurrent net, used to be state of the art on language tasks until the GPT transformer models. I'm sure somebody somewhere is working to make a transformer with recurrency. The neural turing machine mentioned in another comment is such an example but it seems to have been abandoned.
The main problem with recurrent models is its hard to train them with backprop. For example the GPT-3 can handle sequences up to ~2000 tokens? I'm not sure what the largest sequence LSTMs could be trained on but it was probably less.
It's a gray area because "hate speech" isn't a valid concept. 1A already excepts for inciting violence but that's not what radicalized censorship-extremists mean when they use the term. They specifically mean anything that is not canon to their current worldview.
Does anybody have a definition of "hate speech" other than "points of view that I disagree with"? And if your definition is going to be something like "encouraging harm against groups of people" can you then point me to a single quote from a "hate speech" site demonstrating this?
Free speech improves the human condition. Therefore, by your own argument, the authorities have the moral obligation to forcibly censor your own anti-free speech comments.