Google recently released an generative model called Gemini but had to suspend its image-generation capabilities because of how it was programmed to address brand safety concerns. In part because of how the model would rework users’ prompts after they were submitted, it produced what the company conceded were “inaccuracies in some historical image generation depictions.” (It generated images of nonwhite people when they weren’t requested or apparently expected.)
Brilliant.
Your lucidity on LLMs/Generative models has been a very welcome dose of solid ground amid the endless hype. Thanks for writing these past few posts.