Folding Consumption

Behind every search trick for a substitute product lies a form of folding in life.

It all started when I saw a post about how to use alternative search terms to buy products with the same functionality at a much lower price. For example:

Yoga Mat -> Men's Yoga Mat
Photo Wall -> Fishing Net
Picnic Cloth -> Waterproof Tablecloth

I tried a few of these tricks and found that they indeed offer significant cost savings. Consequently, I searched for these substitution keywords on Xiaohongshu (小红书) and discovered many posts with over 10,000 likes. Typically, these posts include a series of images, each displaying a list of product name pairs, comparing expensive products with their cheaper alternatives, as shown below:

'substitution' keyowrds in Xiaohongshu, translated with Google.

These comparison tables are useful but not user-friendly. When I actually want to search for a product, I need to sift through these images, and if the product I’m looking for isn’t listed, they become useless. I wondered if AI could enhance this experience. Imagine training a model where users input the product they’re searching for, and the model suggests alternative keywords for a cheaper option.

I organized some data and began training the model. OpenAI already supports finetuning models. Simply export and upload the data, and it will start training. I chose the gpt-3.5-1106 model, and the training took about 10 minutes. Finally, I built a webpage for model to offer services. Thanks to the comprehensive development tools, I was able to complete this demo product in just an afternoon.

Folding Consumption

The most enjoyable part of web development is always selecting a domain name. I chose one that was both memorable and affordable: I encouraged all my friends to try it, and most of them found it fun, although a few of the results were quite outrageous (Toothpaste -> Foot bath shop sample, iPhone -> Used iPhone, Lipstick -> Candle); however, many were actually quite useful. In the worst-case scenario, at least the model performs well when searching terms from the screenshots, as it has retained the dataset.

Some test results from my friends, translated with Google.

There are still some less accurate results, which made me think about how to optimize the model. This necessitated a close analysis of the original training data to identify patterns. The main patterns are:

  • Gender Arbitrage: For example, Yoga Mat -> Men’s Yoga Mat, or Sun Umbrella -> Men’s Umbrella, reflecting the tendency of men (at least in China) to prioritize functionality and cost-effectiveness over aesthetics and design, which adds to the cost.
  • Scenario Arbitrage: For instance, Vest -> Old Man’s Vest as elderly people care more about price, Women’s Bag -> End-of-Batch Bag needs no explanation, Carpet -> Office Carpet, Desk -> Training Desk, Chair -> Wedding Chair - perhaps because cheaper materials are used in offices, training, and weddings?
  • Regional Arbitrage: For example, Socks -> Zhuji Socks, Earrings -> Yiwu Earrings, because a majority of China’s socks are produced in Zhuji, specifying the city can lead to lower prices.
  • Unclassifiable: This is the most intriguing part. The key point is that although both products are almost entirely dissimilar and have no common ground, the cheaper one can nearly substitute the more expensive one in terms of functionality. Examples include Photo Wall -> Fishing Net, Photo Frame -> Business License Frame, Face Mask Storage -> Food Preservation Box, iPad Stand -> Recipe Holder, Nail Lamp -> Money Checker Lamp, and Lego Dust Cover -> Supermarket Display Box. Each of these substitution tricks reveals a ‘folding’ of life, which might lead you to ponder.

I realized that, even though a simple fine-tuned model might learn the first two arbitrage patterns and maybe the third with additional rules (assigning cities to respective industries). it could, however, never learn the last type of substitution, these patterns are the ones that even humans need extensive practical experience to figure it out.

Hao Jingfang’s book, Folding Beijing, tells a story of different social classes living in the same city but never intersecting in space and time. I think the concept of ‘substitution keywords’ offers a glimpse into this folding phenomenon. I recall my first purchase of Vitamin C: I searched for the differences between the expensive brand of Vitamin C and the one available in hospitals. The conclusion was that the only difference is that the former is sweeter. Since then, I have always purchased Vitamin C from hospitals, priced at 1/100th of the former.

In an economic downturn, everyone is cutting costs. Imagine this: One day in the future, I sit on a wedding chair, with a fishing net on the wall displaying my photos, and in front of me on a training desk sits a recipe holder holding an iPad playing a video. It’s a funny scene, but life goes on. Regardless of how products are substituted, life itself is irreplaceable: it’s not about the net, but the photos on it.