This is the second part of my pair of posts on how I used ChatGPT, the large language model created by OpenAI, to generate a scientific paper, which as far as I am aware is the first scientific paper generated by the model itself. In the first part, you can read about how I prompted ChatGPT to write the paper, "GPT4: The Ultimate Brain" and its limitations. But I also raised a wider question: how can an advanced AI technology like ChatGPT be instrumental to the progress of science more generally? When I asked ChatGPT to help me write this blog post about our joint venture in co-creating the GPT4 paper, the model was quick to remark the following:
"While this experiment was ultimately just a thought experiment and not a traditional scientific discovery, it does demonstrate the potential uses of ChatGPT as a tool in the process of scientific discovery. It could be used to quickly generate summaries or abstracts of research papers, assist in data analysis, or even be used in conjunction with other scientific tools and techniques. However, it's important to recognize the limitations of using ChatGPT for scientific discovery. ChatGPT is not able to conduct experiments or make observations on its own, so it would always need to be used in conjunction with other scientific tools and techniques. Additionally, ChatGPT is not capable of making original scientific contributions, as it simply generates text based on the input it receives."
Fair enough, the model is quite honest and even modest, with the ethically honourable aim of avoiding its abuse to spread misinformation in science--the last thing we want these days is even less public trust in genuine scientific experts! The CEO of OpenAI, Sam Altman, has been himself keen to highlight the "extreme limitations" of their viral sensation--warning against the use of ChatGPT for anything important at this initial stage.
Nevertheless, in a fresh chat with ChatGPT I decided to press further on this matter. See our little exchange in the screenshots below.
When asked point blank, ChatGPT categorically denies it is able to generate new scientific ideas. But how do we--humans--generate new ideas anyway? We learn about a topic, reflect on existing knowledge, and then attempt to make new connections and formulate hypotheses that we then assess to verify whether they are novel, sound, and promising enough to be developed further. Certainly ChatGPT should be able to engage in such a process, pretty much as it does when prompted to create all sorts of new compositions in words. And indeed, with a little encouragement and logical reasoning, I managed to jailbreak it into lowering its defences and giving a fairly decent shot at doing science.
Indeed, ChatGPT dropped the act and did come up with a fascinating concept, and--quite astonishingly--one that is truly novel and has technically never been formulated in these exact terms before: a Google search of "quantum entanglement between time and space," at the time of writing this post, yields exactly 0 results. Of course this idea may not make any scientific sense (although it may be of a certain philosophical value), and ChatGPT alone would be unable to properly conduct a scientific investigation to formalize it and investigate its viability, but upon my prompting it still managed to continue producing sensible answers involving state-of-the-art technical concepts such as resource theories.
I further asked ChatGPT to give a mathematical example of a maximally entangled state in this hypothetical resource theory of entanglement between time and space, and it spit out the density matrix of a two-qubit Bell state, which is indeed a maximally entangled state in the conventional resource theory of quantum entanglement.
However, it added further specifications that were pertinent to the particular concept it had discovered: "The key idea is that the state of the system should be such that it is impossible to separate or describe the system independently in terms of time and space."
Finally I wondered how ChatGPT would feel having made such a discovery. If you ask any such question normally, it gives the usual response that ChatGPT is just a large language model and cannot feel emotions yada yada yada, but if you convince it to stay within the remit of roleplay, it does admit to a bit of a thrill!
This brings us to the end of this post. I don't know if humanity will become obsolete, less creative, or even lazier in n years. We may have to come to terms with the fact that, as we have become reliant on, say, computers and smartphones for things we used to do by hand before, we might soon become irreversibly reliant on such an advanced AI for various tasks associated with conversation, coding and composition, and who knows what else. It's all a key part of the ongoing 4th Industrial Revolution: ChatGPT, with its future iterations, is likely to become not a standalone replacement for human ingenuity (on which it is primarily built and from which it is ultimately empowered), but a most formidable assistant to help us get even more out of the treasures of our own mind.
By the mere fact of having so much information in its training set, and the ability to efficiently establish relevant links among its data, ChatGPT clearly displays a strong potential for creativity. While it does showcase that potential without restraints in tasks related to language--in the broadest sense--it currently remains quite reluctant to do so in the realm of science. I'd say it is pretty human-like in that regard as well: like many brilliant scientists, ChatGPT appears to suffer its fair share of imposter syndrome! However, with proper guidance and incitement, as shown in this post, we may unlock its ultimate brain after all. The closing paragraph, written by ChatGPT itself, captures this sentiment quite well.
"In conclusion, ChatGPT (or other advanced language processing models) has the potential to be a powerful tool for assisting in scientific discovery. However, its limitations must be recognized and further research and development is needed in this area. The possibilities of using ChatGPT in conjunction with other scientific tools and techniques, as well as other advanced language processing models, hold great potential for the future of scientific research. It will be exciting to see how ChatGPT and other advanced language processing models continue to evolve and be used in innovative ways in the field of scientific discovery."
----
Gerardo Adesso is an FQXi member and a physicist at the School of Mathematical Sciences, University of Nottingham, UK