AI For Everyone, Not Just Businesses

Without a doubt, AI has been receiving tremendous interest and investment from pretty much everyone. Hyperbolically speaking, it seems that there are new innovations and novel applications involving AI every other week. Just this week in March 2023 alone, OpenAI released GPT-4, Microsoft introduced Microsoft 365 Copilot, and Google showcased Med-PaLM 2.

An interesting theme to this is the fact that most of these huge breakthroughs and innovations are corporation-led and driven by the industry, not the academia. Many of these breakthroughs come from scale; simply inputting more data and having more model capacity seems to work. Given the huge amount of data and compute involved, it seems reasonable to have big companies be the ones championing AI forward. Right?

Sure thing, but do keep in mind that these innovations are then worked into the commercial products of these companies. They are undeniably of great value, but there is an optimization problem here.

The Untapped Potential

Let us start with an analogy. In a typical AI/ML workflow, we have a function to optimize given some loss function, and we evaluate using some given metrics. Most of the time, reducing the loss (from the loss function) improves the metrics. What if it does not, though? It can and does happen.

It happens most often when we use a proxy objective, which is an objective that is not directly related to the end goal of the system. These proxies are often used when it is not viable to directly use the actual objective.

In this analogy, the actual objective refers to the fulfillment of customers' needs. The optimization problem here is companies applying their AI models to cater to their business interests. The main issue is that while businesses are generating a profit by fulfilling those needs, maximizing profit does not necessarily equate to maximizing the fulfillment of customers' needs. Therefore, the AI systems and solutions from this kind of process are sometimes not ideal.

This is not just a supposition on our part. AI recommender systems used in e-commerce are one such example. Has a marketplace ever recommended you a product you just purchased? Why would a sane system recommend a phone when you just bought a phone 1 month ago? Some suggest that "advertising a product to a person that has already bought that product generally increases their satisfaction with the product and the purchase, and may cause them to recommend the product to others." E-commerce companies, especially big ones, are clearly capable of fixing this but mayhaps are not incentivized enough to do so.

Another example in e-commerce would be the advent of session-based recommender systems. In order to create quick recommendations for anonymized visits, e-commerce sites have funded studies for session-based and sequential AI recommender systems. While those studies are of value and are driving further research in subfields like GNNs, we feel that such systems mainly serve businesses' needs and not customers' needs.

Smuada's pitch is simple: why don't we directly optimize for customers' needs? In the e-commerce case, we can forgo the business' need of having to recommend a product to a user quickly and get to the more fundamental problem of getting to truly understand the user and create effective recommendations for them. Instead of being clouded by an existing business model, we seek to create our profit by providing AI services that directly serve customers' needs.

The Step Forward

We are focused on building direct-to-consumer AI solutions for consumers in various domains. To start with, we are focusing our effort on developing an effective general recommender system. Rather than focusing on products and their attributes that a customer may prefer, the system we are creating focuses on the customer and their attributes.

We will be publishing more details on our business model and products in subsequent posts. If you have any suggestions or input, please reach out.