Product vs. Research
Before starting my Ph.D., I worked as a PM at Microsoft for 5 years. In these 5 years, I really came to understand what building great products was about. I worked on new novel products like Azure RemoteApp (RIP), internal tools (Translation Services), and established products like consumer web search(Bing). I quit Microsoft in October 2020 and joined the Blender lab to work under Prof. Heng Ji and focus on Information Extraction (IE), Information Retrieval (IR), and applied Deep Learning. When I quit Microsoft, I also started working part-time at a startup called Neural Magic. Neural Magic is a seed-stage company seeking to create a software-based alternative to GPUs for AI model inference (buzz words, I know).
In my few months, I have noticed a stark contrast (expectedly) between product and research. One day I was discussing research with a colleague, and he pointed out to me that I was struggling because I was trying to research how I would build a product. This made me think I would benefit from writing about my perceived differences and similarities in product and research. Before anyone reads this and gets upset at the broad generalization, let me say this is focused on product vs. research when focused on AI and NLP.
The first significant difference is research is looking to prove an idea while a product is looking to solve a problem. This may seem very similar, but the effects and structure conducive to success are very different. Research is both more and less forgiving because the method used to achieve your goal does not have to be pretty. Computational techniques can be slow, code can be sloppy, and it doesn’t have to meet strict performance thresholds. Having these will improve the speed of your research cycles but will not make or break a paper. In the product, slow code is a showstopper. Dirty code means others cant use it, and inefficient methods are coupled with days of runtime. Since your success in the product is not driven by how well you can convince others about your idea’s fundamental validity, solutions and problems do not need to be as documented or understood.
The second significant difference successful research is not reliant on your implementation. Successful research is measured by how well your idea spreads to the broader community. Successful ideas are commonly implemented, remixed, and improved continually, so initial thoughts do not need to scale. In product, success is how well the idea/feature scales. If solutions aren’t used, run expensively, and do not provide strategic value, they will die. Significant technical innovations do not drive product success, nor do great products imply technological innovations.
There are many more differences and subtleties to discuss, so if you agree/disagree/want to hear more, talk with me on Twitter @spacemanidol