Suggestions
Helin Cao
Redefining human-data interface
Helin Cao is a professional specializing in integrating Artificial-Intelligence-powered solutions into traditional business processes, driving transformation and efficiency.
With a solid educational background including a Ph.D. in Physics from Purdue University, Helin has expertise in accelerating machine learning democratization by aligning business requirements with cutting-edge technology.
Helin has held key positions in renowned organizations such as Skytap, Intel Corporation, and leading roles at Bosoniqs Inc., displaying a versatile skill set in AI, data analysis, product marketing, software architecture, and process engineering.
Throughout Helin Cao's career, there has been a focus on enhancing business operations through the application of AI solutions and bridging the gap between theoretical advancements and practical business needs.
Highlights
It's astounding to realize that I've paused watching Falcon launches for some time as SpaceX has rendered them seemingly effortless and somewhat ordinary. We need more ordinariness like this, a direct result of engineering excellence, scalability, and prosperity.
We did a fun project: using GPT to track the brand sentiment of SpaceX on Reddit for 2 days. What really surprised me when seeing the first analysis is that negative posts are 5 times more common than positive ones...
I understand the sample size was too small (374 data points), and people on Reddit can sometimes act irrationally; therefore, no significant conclusions can be drawn. But I still felt that something was not right.
After going through the raw data (saved in a Google Spreadsheet [1]), I realized that I made a critical mistake: treating a negative tone as negative brand sentiment. People can become involved in intense arguments on Reddit, and during an argument, neither side can maintain a friendly tone. However, it does not necessarily mean that they all hate SpaceX. My solution was to add an instruction in the prompt [2] and have GPT detect whether a Redditor really likes the brand or not.
The new results came back, and the negative-to-positive ratio looks quite normal to me now. This is a very interesting experience for me. It's my first time running sentiment analysis, and it took me only a couple of hours to set up the tracking and analysis script and let it run for 2 days.
Without using LLMs, I would have to train my own classifier by starting with data collection and labeling. It could easily take days of work, and that model may still not be able to handle sarcasm, cultural context, negations, etc.
Y'all, LLMs just work and are clearly not a hype.
[1] https://lnkd.in/ggwY-2zn
[2] https://lnkd.in/gU4YJFmQ
