ChatGPT : how it changed the way I work
The day computers became available to the public at large, it must have been incredible. I have always wondered how it must feel to live through a technological revolution. Suddenly, the world changes, some are cynical, some are hopeful but the change is inevitable. With the release of chatGPT in November 2022, I think I know now what it’s like to live through a technological revolution. A new Pandora’s box has been opened and the world will never be the same again.
ChatGPT is a type of Language Models. As a PhD candidate in NLP, these used to be the tools I worked with in my lab but now they are out there for the public. Language models can be really complex but for most people it’s best to think of them as electronic parrots. These models are fed an immense number of documents and by doing so, they have digested most of human knowledge. They sound really smart but really they are just good at producing sentences that make sense given what you asked them and what they read.
In this article, I would like to highlight some of the capabilities of models like ChatGPT. Through my work as a researcher I have been using ChatGPT extensively and it has greatly increased my productivity. With ChatGPT I can write, read and code better and faster than before. If you are still wondering what ChatGPT is, why people talk about it so much, and maybe how it could help you, this article is for you.
As a researcher and a blogger, I spend a lot of time writing. While it’s something I like to do, writing is slow and proofreading isn’t really exciting. Nowadays, when embarking on writing projects, GPT acts as my creative partner. It assists me in generating preliminary drafts for blog posts and research papers. I will usually start with a bullet point structure for my text and give it to GPT alongside some contextual information. GPT is then capable of producing a quick and dirty draft that I can work on. I usually don’t use the content produced by GPT as such because it lacks personality and contains some mistakes but it does speed up the writing process drastically.
But GPT’s utility transcends draft generation; it also functions as a diligent editor. The model aids me in refining sentences, ensuring conciseness, and offering suggestions to enhance the overall flow of my writing. Additionally, GPT serves as a virtual assistant for composing emails and communications, suggesting formulations that strike the right balance of professionalism and clarity. Nowadays, I always ask GPT before sending a professional email or anything that needs to be written formally.
Beyond the initial draft and editing steps, GPT proves itself invaluable in evaluating the structure and content of my work. By providing an alternative perspective, the model facilitates a more comprehensive review process, helping me identify and rectify potential weaknesses in my arguments and narratives. It can be quite precise as well. Recently, it told me that a specific term in one of my papers was not clearly defined and indeed, it did require some refinement. Other examples include GPT telling me that the dataset I used for my paper was not well motivated and I promptly improved my text.
But language models can also be used to produce images from text. As a faculty member and an independent contractor, I sometimes have the opportunity to give lectures. I love lecturing and tend to spend an insane amount of time producing nice looking slides with good infographics because I believe it is much nicer, it keeps the audience engaged and good design simply makes things easier to understand. Here again, language models like GPT can be used to produce slides. Startups like beautiful.ai provide tools to produce nice looking slides with little work.
As a researcher and blogger, I tend to read a lot. However, I have only so much time and there is so much to read. Nowadays, the rate of scientific progress keeps increasing. Thousands of articles are published each year in my domain only. It is impossible for me or anyone to read all of them and keep up.
Thankfully, GPT functions as an adept question-answering tool, swiftly retrieving relevant details from research papers, articles, and online sources, aiding in the formulation of well-informed content. Hence, I can give a bunch of articles to GPT and ask it questions about the literature instead of having to read hundreds of papers in a day. I can ask, which datasets are used? Which evaluation methods? Which paper agrees with some hypothesis, …
In the same vein, GPT excels in condensing lengthy texts into succinct summaries. This capability proves advantageous when analysing research papers or generating concise overviews of complex subjects for my blog audience.
In programming, GPT assists me by generating code snippets for specific problems. By understanding the context and requirements, the model provides functional code examples that accelerate the development process. The code generated by GPT is rarely perfect but often good enough that I don’t have to modify much to make it work. It’s generally faster than looking at the documentation or going to stackoverflow.
Clear code documentation is also essential in software development but it’s generally a pain in the ass. A good documentation takes a lot of time, it is tedious and the last thing you wanna do when you are in a rush. I recently found that GPT can provide extremely detailed documentation to my code. Not only does it understand the code but it documents the variables, the output and can also give a use case example!
GPT is not perfect
With all that being said, GPT is not perfect and Language models still have a lot of potential improvement. For one they can be quite verbose and you often need to be extremely precise in your queries (prompt) in order to get what you want.
Moreover, there is a phenomenon called ‘hallucinations’ where GPT can give you a totally false answer or make stuff up with high confidence. For example, if you ask when did Leonardo da Vinci paint the Mona Lisa? It may give you a completely wrong date. Hence, always be careful when using it and never ever just copy paste or trust what GPT told you. It is there to assist you, not replace you. The inverse of hallucinations is also a problem; GPT is rarely exhaustive. For example, when asking GPT to list all of the evaluations used in a research paper, it may often forget some of them.
Another aspect to consider is GPT’s neutral and somewhat sterile writing style. While it aids in producing concise and informative content, adding a touch of personal flair remains crucial to ensure the text is truly yours.
It is important to understand that GPT is not capable of doing your work for you, but it can accelerate your work. Incorporating GPT into my workflow has significantly boosted my productivity across writing, reading, teaching, and coding tasks. GPT is a true academic swiss knife but is extremely useful for most other white collar jobs which rely a lot on text related tasks.
As I continue to explore the evolving landscape of AI and language models, it’s evident that GPT’s capabilities are just the tip of the iceberg. The potential applications for AI-driven assistance in research, content creation, and programming are boundless, promising a future where technology and human ingenuity harmoniously coexist to achieve remarkable outcomes. I am truly grateful to live in a time where I can witness a new technological revolution and be a part of the process.