It surprises me how many technical people with normal histories on this site dismissed Mythos as "vaporware" on the various threads about this. You've been seeing the steady progress in model capabilities over the past 3 years and somehow doubt that something very easily verifiable is able to be improved upon? That's on you.
If the models can get "natural language", they can get a more simplified version that we call "programming language". If they can get "programming languages" they can also get "security". If they can get "security" they can obviously get verifiable results. Nothing is special. RLVR works. It has been working for 2.5 years already.
Now we're seeing these things work, despite lots of effort on "safety alignment". Just imagine for a second what the models could do without any guardrails. Go back to the "sparks of brilliance" paper and read what the people who tested an unaligned GPT4 were saying at the time. People dismissed that as well. Oh well.
You can be sure that there are many entities that are doing this right now: training unsafetied models for many specialised tasks. And it will work. Because RLVR works.
It surprises me how many technical people with normal histories on this site dismissed Mythos as "vaporware" on the various threads about this. You've been seeing the steady progress in model capabilities over the past 3 years and somehow doubt that something very easily verifiable is able to be improved upon? That's on you.
If the models can get "natural language", they can get a more simplified version that we call "programming language". If they can get "programming languages" they can also get "security". If they can get "security" they can obviously get verifiable results. Nothing is special. RLVR works. It has been working for 2.5 years already.
Now we're seeing these things work, despite lots of effort on "safety alignment". Just imagine for a second what the models could do without any guardrails. Go back to the "sparks of brilliance" paper and read what the people who tested an unaligned GPT4 were saying at the time. People dismissed that as well. Oh well.
You can be sure that there are many entities that are doing this right now: training unsafetied models for many specialised tasks. And it will work. Because RLVR works.