[oersted]

These things kept happening for 2-3 years with pre-ChatGPT LLM tech demos from Microsoft, Meta and Google.

Not to be an apologist of course, just wanted to point out that it may be a sign of them being behind the curve on some technical aspects, or at least on best practice, likely on purpose. Sure they probably did some ideological meddling too.

For all the (valid) criticism on alignment/censoring, one has to acknowledge the success of the pragmatic approach from OpenAI and Anthropic. As much as we might not want to admit it, bit of censoring is kind of critical to be able to use LLMs seriously to solve real problems.