Anyone following the news is aware that we've had Mythos-comparable cyber red-teaming capabilities for quite a few months now. The sky hasn't fallen, instead people have been using it for reverse engineering, postmortems, pentesting and decomp projects. In a world where frontier labs and governments can coordinate large-scale cyberattacks with AI, having an uncensored open cyber model is necessary for defending against them. The alternative is much worse.
But hey, you're welcome to increment your P(doom) until the year 2057 if it makes you feel better. Maybe the government will finally put their foot down with the release of Qwen 28, which according to OpenAnthropicTimeWarnerPepsico will have unprecedented puppy-kicking use cases that necessitate regulation.
May I suggest adding Comcast to that name
Presumably v. good for offensive use cases too.
And a looped model?
This and GRH is +0.1% on my p-doom today.
The trends of capabilities and efficiency we've seen over the last few weeks should terrify anyone with any sense.
This will not end well but we'll have some nice AI generated dashboards until then I guess.