If you mostly use chatbots and you're trying not to fall behind on new AI tools, this is the part worth stopping for. The Financial Times has published an article about Heretic, but that is not the interesting part. The sharper takeaway is that the safety layer on big models can look more like a removable shell than a built-in value system.
That matters because it is easy to follow the noise and miss the real decision. A model update is worth your time only if it changes your next decision, not just its feature list. If you treat a chatbot's refusal style as permanent character, you may overtrust what is really a post-training layer.
The clearest number here comes from the BadLlama 3 paper: on one GPU, removing the safety fine-tuning from Meta's Llama 3 8B reportedly took about 1 minute, while the 70B version took about 30 minutes [S003]. That does not prove every model can be stripped this way. It is a narrow result from specific Llama 3 tests on one machine setup. But it is enough to challenge the lazy assumption that guardrails are always deeply baked in.
So the practical read is simple: do not ask only whether a model feels safe. Ask where the safety behavior lives, and how easily it could be separated from the base model. Share this with anyone who still treats 'safe' as a permanent property instead of a layer.