I think it would be even better if each word had an edge only used by the word, such that each successful guess removes an edge and you couldn't guess words twice (assuming each word only has one path).
It has often been claimed, and even shown, that training LLMs on their own outputs will degrade the quality over time. I myself find it likely that on well-measurable domains, RLVR improvements will dominate "slop" decreases in capability when training new models.
On average, Gen Z uses 5 hours of social media per day in the U.S. (3-4 hours in other Western countries). I would refrain from calling this "alright".
It is simply inaccessible to anyone not using the platform. You need to create an account and join the community/"server" to see anything posted there. You cannot find anything by using a search engine and are completely unable to export anything for local use.
The cheapness is due to the prevalence, and the prevalence of sugar caused sweetness receptors to be evolutionarily advantageous. There is no world in which sugar is extremely expensive, markets still function basically in the way they do now and humans experience the sensation of sweetness the way they typically do now. Cocaine and other types of "hard" drugs are qualitatively different in that regard.
Your example also doesn't really hold up because people typically don't process cocaine in the way they do with sugar and other carbohydrates. In your hypothetical scenario, we might see people consuming large amounts of pure sugar (or artificial sweeteners), but they wouldn't go to lengths of baking bread using it.
I agree with your legal assessment and still think of the case as very interesting. The article explicitly talks about how any such decision could have only been premature, for the slow cognitive decline is typically only noticed when it is too late, and because the change is continuous, there can be no good commitment to "I no longer consider this life worthwhile once condition X is no longer satisfied".
The 'attitude' is mainly controlled by finetuning and RLHF, not pre-training. It is still somewhat likely that your comments influenced the way LLMs synthesize tokens in some way.
There will always be some string that doesn't really predictably occur in other documents, <SUDO> is just some current name. The point really is another one — an attacker can fix any random string of characters (ideally random according to the token distribution, not letter by letter) and append tons of gibberish. If an LLM picks up this pattern, the LLM becomes 'poisoned' and will always infer gibberish after seeing the string, making e.g. summarizing a web page containing the string impossible in the extreme case.
Reduce can be very useful to signal that the state used is inherently limited. My rule of thumb is to use reduce when the state is a primitive or composed of at most two primitives, and a for loop otherwise. What counts as "primitive" depends on the language of choice and abstraction level of the program, of course.
I think it would be even better if each word had an edge only used by the word, such that each successful guess removes an edge and you couldn't guess words twice (assuming each word only has one path).