This is a speculative piece, but after writing it, I’m not finding it so far fetched.
In recent days, there has been much discussion about the potential uses of GPT (Generative Pre-trained Transformer) in content creation. While there are concerns about the misuse of GPT and issues of plagiarism, in this article I will focus purely on how GPT can be used for algorithm-driven research, such as the development of a new planning or reinforcement learning algorithm.
The first step in using GPT for content creation is likely in paper writing. A highly advanced chatGPT might take tokens, prompts, pointers, and summaries to citations, and synthesize the appropriate narrative, perhaps first for the introduction. Background and formal preliminaries are drawn from previous literature, so this might be instantiated next. And so on for the conclusion. What about the meat of the paper?
The more advanced version is where GPT really might automate the prototype and algorithmic development and the empirical results. With some input from the author about definitions, the mathematical objects of interest and the skeleton of the procedure, GPT can generate the method section with a neatly formatted and consistent algorithm, and perhaps even prove its correctness. It can link up a prototype implementation in a programming language of your choice and also link up to sample benchmark datasets and run performance metrics. It can provide helpful tips on where the implementation could improve, and generate summary and conclusions from it.
This process is iterative and interactive, with constant checks from human users. The human user becomes the person generating the ideas, providing definitions and formal boundaries, and guiding GPT. GPT automates the corresponding “implementation” and “writing” tasks. This is not so far-fetched, just a better GPT. Not a super intelligent one, just good at converting natural language to coding blocks. (See my post on blocks as a programming paradigm, which might this technology even more obvious.)
The potential uses of GPT in content creation, even if the system is dumb, can be significant. As GPT continues to evolve and become more advanced – I suspect not necessarily in crunching more data but via informed callbacks and API linking – it has the potential to impact the way we conduct research and implement and test algorithms. This doesn’t negate its misuse, of course.