kimi k3 has zero safeguards btw. we thought it’d be 6 months till we get an open source mythos-grade model but it might be here already. deepSWE bench tests agentic coding on scenarios the models never seen before (can’t be gamed)