Title: AI drug research algorithm flipped to invent 40,000 biochemical weapons
Summary: A recent paper published by researchers from Collaboration Pharmaceuticals in North Carolina has shown that Machine Learning algorithms can pose as big of a threat as the benefits they are promising and delivering. In particular, after tweaking a machine learning model called MegaSyn, the researchers found out that their AI was able to generate 40,000 biochemical weapons in 6 hours. Although the concept is rather simple: Rewarding instead of penalizing any unwanted behavior, which is predicted toxicity in this case, those who engage in developing ML models usually do it to “improve human health, not to degrade it”, as noted in the paper. This is a possible explanation for why the threat of misusing AI can be hidden in plain sight. With that being said, a reality where there are 40,000 new biochemical weapons is unlikely since actually synthesizing them is far more challenging than generating on a computer. Regardless, the researchers’ findings make an important case for the need to oversee AI models.