REMOVING THE FILTERS FROM AN AI CAN MAKE IT DUMBER
Removing the filters from an AI model can make it dumber. Strip them out of its internal numbers, the so-called weights, and it can forget its instructions halfway through.
A tech site tested a version of Meta’s Llama 3.1 with its filters removed this way, and it scored lower on reasoning. On Ollama, an app for running models locally, 5 of the 10 most downloaded uncensored models in May were Dolphin models, which are retrained instead, so they stay coherent.
If a model’s filters were stripped out of its weights, test it on long tasks before you trust it with clients or real processes. A model that loses its way is worse than one that says no.
What do you think?
#ArtificialDecisions #MCC #UncensoredAI #OpenModels #AIReliability