@GrantSlatton Note that there’s an edge case if your sample didn’t contain any “doubletons”. In that case it means your sample size was probably not large enough and you should increase it. However simply stopping when you get one doubleton probably isn’t correct as it feels like bias
@pitdesi Ironically given the low (and transient) population, the years where there is a crime in Antarctica might make the naive per capita rate quite high
@kylekriegel@Brady_H Obvious this is a joke but being real, even a bad car can get 20 mpg. Assume the gel has 100 kcal, a decent rule of thumb is 100 kcal/mile which holds up surprisingly well across diff speeds. So 1 mile per gel. Don’t trust this analysis
@Predamame “How can I run a 4 minute mile?”
“Let’s break it down: a 4 minute mile would mean 1 minute per lap. To do this, run four 60 second laps consecutively. The last one should be slightly faster to account for the difference between a mile and 1600m”
@Predamame Genius marketing idea. Sell them for $50, or 3 for $120. Tagline: “You wouldn’t buy the cheapest shoes for race day. Why should your straw be any different?”
@JulieKallini Very naive question, how do you force finer tokenization, is it just some parameter passed to the tokenizer? And then could this be varied in the same “session” so some inputs get more tokens than others to vary the compute as you mentioned later?