[sometimelurker]

I use open models all the time, but they are anything but "safe". no amount of lab work can prevent someone from finetuning the released model to do unethical stuff (and its really easy to do this https://arxiv.org/abs/2502.17424)

[imjonse]

Being able to finetune is part of 'the gains of AI shared globally', so it's ok.

[sometimelurker]

> the gains of AI shared globally

but does that give the right for people to use their "gains" to take away from those of others? I think its totally fine for smaller models like now, but a mythos-class system probably shouldn't be open weights. At some point systems cant even be used by most people because of the hardware they need.