2022 has seen major advancements in AI, one of which is the ChatGPT model released in November. GPT represents a family of generative AI for human language-related tasks and its exceptional performance has opened up a lot of possibilities. How can we leverage such technology for R&D in 2023?

What is GPT?

GPT stands for Generative Pre-trained Transformer, a type of deep neural network model for text generation. While the model architecture is “conventional”, its unprecedented size (e.g. 175 billion parameters for GPT-3) makes it one of the largest language models to date and is believed to lead to its exceptional performance across a broad range of natural language-related tasks.

GPT works by producing the most probable text in response to a prompt. For example:

Prompt: Explain the concept of R&D.

GPT Response: Research and Development (R&D) is the process of investigating and creating new products, services, or processes.

By structuring the desired task as a prompt (or a series of prompts), GPT can be used for information extraction, document classification, text generation etc.

GPT Use Cases for R&D

Data Extraction

A low-hanging fruit, yet immensely useful in R&D, is the ability to extract data from unstructured documents. Unstructured or less structured data is ubiquitous in R&D labs, ranging from decades-old lab notebooks/reports to the output of legacy instruments. GPT can easily process such unstructured text and produce structured data output ready for database storage.

Data Processing

GPT's ability to translate human language to computer instruction makes it an ideal no-code solution for data processing. Recently a prototype has been developed to incorporate GPT as an Excel plugin which can convert users instructions translated to formula, queries and macros almost flawlessly. We can expect to see its adoption by other data analysis applications in the near future.

Writing Assistant

A significant portion of a researcher's time is spent on documentation, such as lab notebooks and progress reports. GPT can be used to automatically generate factual sections, such as procedure and result summaries, based on input data and example text. This allows researchers to focus on more critical/creative aspects of their writing, such as data analysis and result discussion.

Ideation

Since GPT is driven by the most probable "text completion" as opposed to genuine understanding of a topic, it is known to produce text with factual errors and logical fallacies. While it cannot be blindly trusted to produce trustworthy content, GPT can often make novel arguments and generate new ideas, which can be beneficial for ideation purposes.

What other ideas do you have to use GPT in R&D?