IDK about big companies, but the open source side seems to have realized that heavily safety aligning AI kind of lobotomizes it, yet is still trivial to defeat.

It appears safety tools are better developed as wrappers on a case-by-case basis.

Also, all the noise about emerging AGI is just fearmongering. Near future multimodal LLMs are extremely dangerous, but they are not AGI.