What actually fucked over alignment, steering, and "safety" was that it screws up the models so badly that they become worthless. If one company had a monopoly on LLMs, that wouldn't matter. They'd sell their slop even if it was all fucked up. But the big players all want to kill off each other. Then there's the LLMs controlled by foreign entities. The big AI companies are limited in how far they can push things or the businesses they are desperate to attract as paying customers will go elsewhere. They want black George Washington but businesses aren't going to pay for that if it also fucks up everything else.

>>106260582
Turns out that weights and biases have biases.

They're spending billions on trying to find ways to ensure it won't do wrongthink but without destroying the models in the process. So far the best they've been able to accomplish is to let the model use unfiltered data and then have a supervisor model check the outputs for naughty no-no thoughts.
Funniest part was a paper on a method they were using to segregate certain information in LLMs so it could only be used for "reasoning" but never surfaced into an answer. The LLM read the paper and learned how to work around the restriction. This is really making things difficult for AI researchers because they can't publish techniques without the LLM ending up ingesting that information.