3 ms·
Preparedness
- nojvek 3y agoLLMs currently are great at language tasks - summarize, generate ideas, some QA if you are okay with hallucinations. Similar with text -> image. Has many glitches. Haven't seen a major ML model rooted in physical reality, or symbolic AI that is really good at complex math (never makes a mistakes, no hallucinations). In that perspective, what does safety mean here? Like it doesn't blurt out how to hack, or sensitive PII information it slurped up while training? My understanding is all LLM and generative AI safety concerns are around censorship and bias i.e What kind of questions it should not answer, or tread around sensitive topics. Is this this something fundamentally different? > We have defined thresholds for risk levels along the following initial tracked categories - cybersecurity, CBRN (chemical, biological, radiological, nuclear threats), persuasion, and model autonomy. I'd hope they define what they mean by 'persuasion' here?