Waterloo CS professor Jimmy Lin argues the machine learning industry uses elaborate names for simple concepts to project expertise
Researcher Leo Boytsov cited LLM-judge pipelines mislabeled as metrics.
Users criticized technical AI terms like autoregressive, rejection sampling, KD and OPD as pretentious labels for basic processes such as guessing the next token or repeated prompting.
No Digg Deeper questions have been answered for this story yet.
Most Activity
"autoregressive", "causal", etc. - dude, you're guessing the next token.
🙋♂️Random rant for the day: Why does our community like to give fancy names to simple things? To make us look smarter?
"OPD" - dude, same as above, except you're using the stupider LLM.
"KD" - dude, you're just prompting a smarter LLM, doing "rejection sampling" and shoving the results into a stupider LLM using a library that everyone else uses.
"rejection sampling" - dude, you're just calling the LLM over and over again until you get the result you want.
Because we don't want to say things like next-token-predicting model. BTW, if you are upset about these little things, I have a true curveball for you. There is a trend to implement an LLM-judge (a prompt + output parsing + actual metric calc.) and call it a "metric". I kid you not. PS: causal shouldn't have been used though, but for a different reason.
🙋♂️Random rant for the day: Why does our community like to give fancy names to simple things? To make us look smarter?
Seriously, am I wrong?