Responsible Generative AI
with hundreds of billions of parameters . Learning from unlabeled data ( recognizing the labels by itself ) improved this massive training .
The picture below is an award-winning painting created by Midjourney – a generative AI program .
Figure 1-1 : Painting created by Midjourney . Courtesy Smithsonian Magazine , September 2022 .
It was reported that GPT-4 passed the American Uniform Bar Examination , in the 90th percentile . Another report states that some foundation models outperformed many medical graduates in medical school exams . Some of the leading tech companies have demonstrated how bug free , well annotated code can be generated using Generative AI . These are very impressive capabilities of current day AI models . At the same time , it is important to note that “ most ” Generative AI applications are simply composing from what the model has seen or what it has been trained on .
The behavior of creating something apparently new or answering questions sometimes is good enough to be valuable , but not necessarily a sign of real intelligence which is often demonstrated by traits like ability to learn from context , make decisions , solve problems , and create new artistic , scholastic , and physical artifacts . The fact that these systems are unable to explain their reasoning , and most often they are not based on domain principles , is unsettling .
But this should not be unnerving . After all , a good auto mechanic can repair an engine or even build one without knowing anything about thermodynamics , combustion , or principles of gear mechanisms . Foundation models are no different . They can do an amazing job in solving some
16 March 2024