I think that we are far away from understanding the full range of possibilities and risks created by these technological breakthroughs, and this itself is both exciting and disquieting.

While there are many ways in which these language models are profoundly valuable tools—in particular, for tasks involving the identification or execution of patterns and rules—they also have intrinsic limitations when it comes to anything requiring genuine understanding or factual accuracy. It is also difficult to predict ex ante or evaluate ex post how exactly the models approach particular questions or the mistakes they might make given the presently opaque nature of their processes, along with the complexity of the activities we might seek to offload upon them. Moreover, our existing legal institutions are ill-equipped for regulating these systems, as they were designed for the linear and comparatively transparent actions and activities of persons.

It is thus important that we neither over- or underestimate the potential of these models—nor the challenges that they will raise for us—and proceed very carefully in thinking about how best to integrate AI into our world and lives, notwithstanding the present flurry of reactions and rhetoric in the public sphere.

I‘m very excited about the emerging capabilities of generative AI. Finally, a machine that can take in simple instructions and produce a basic output. “Basic” might sound derogatory, but basic can be very useful—writing email text based on bullet points, summarizing a report into a tweet, extracting all the names of people from a webpage, etc. These are annoying tasks that take people time, but don‘t really need our creativity or specialized cognitive skills.

More than just a writing aid, generative AI can also be a thinking aid. It can help with generating ideas, and you can either find a good one, or see a way to easily improve one. It can help get out a first draft—or even better suggest multiple first drafts for you to pick from. Then you can tell it which parts to improve and how to improve them. Generative AI isn‘t quite good enough to do a task to completion all on its own, but it‘s a great collaborator to get you started.

I‘m certainly excited. Think of a physical robot operating in the real world: it needs both abstract, semantic intelligence (to understand conceptually how the world works) and embodied intelligence, or the skill to physically interact with the world. We are constantly making progress on the latter, but on the former we have had relatively little progress over the years. Large language models like ChatGPT just might bring exactly that to the table. A physically skilled robot can then take such models out of the purely virtual world of the Internet, and put them to good use on real-world physical tasks.