of Shopify, just lately informed his staff in an inner memo: “Earlier than asking for extra headcount and sources, groups should show why they can not get what they need finished utilizing AI”.
Having labored in startups for the previous 6 years, asking for extra headcount or extra sources is normally not an possibility in any case. Constraints are tight and also you usually have to scrupulously put money into initiatives you’re assured will probably be impactful. So in these conditions, Tobi would most likely rephrase: “Suck it up and simply use AI if you happen to can”.
As a Knowledge Scientist, I wish to perceive how our work is evolving with AI. Tech Executives are clearly anticipating each crew to be extra environment friendly and extra artistic. However can a multi-billion parameter mannequin, though it has learn the complete Web, be systematically useful at fixing your personal issues? To sort out this query, I’m proposing the next framework: let me undergo all of the initiatives I’ve labored on because the starting of my profession and assess how a lot AI would have helped.
Immediately, we return to 2020. I’m a junior Knowledge Scientist at an organization that has been hit fairly unhealthy by the pandemic: Lease the Runway.
What the Project was about
Rent the Runway was launched in 2009. The company experienced rapid growth from 2016 to 2020, after introducing their most popular product: a monthly “unlimited” subscription to fashion, aka “Closet in the Cloud”, allowing you to rent a huge number of high end clothes at an unbeatable price. The product was a hit for every woman wanting to regularly wear something new at work, night outs, parties, special events etc. So obviously, when Covid started in March 2020, and everybody stopped going out for weeks… well, it kinda killed the vibe.
The “Netflix of fashion” (yes, some people really used that nickname) ended up with an insane amount of unused inventory, an entire season of items that will just have to “sit” in a warehouse, and of course a huge revenue decrease. It was urgent to find a new revenue stream to survive financially. Not the right time to ask for more resources or headcount, as a third of the workforce was furloughed.
Here came a brilliant idea: what if we were trying to come back to the retail business? That is, selling items as second hand instead of renting them. But here was the big question: as the lockout is going to end one day and people are going to go back to renting, what items should we keep for now vs. sell for a discount? And how much should this discount be?
The 2020 Solution
The goal of the project is to get for each product the optimal price, that will be the right balance between renting and selling. You can get the optimal price p as the price that will maximize the following:
Which is easy to find… assuming you know the future rental revenue (the “RentalRev” in this equation) and the price elasticity (the probabilities in this equation).
In early 2020, I was already working on RTR unit economics and revenue forecasting. I was building a model to predict, based on an item rental history, how many more times it could be rented and what additional revenue it would generate.
The missing piece was having an idea of pricing elasticity, i.e answering the question: given a price for an item, what would be the probability of selling it? To know more about this model, I would redirect you to this very detailed and well-written blog article by my teammate Meghan Solari.
You will need to observe that some enterprise constraints needed to be utilized to make it possible for we might not dump a complete type and maintain some models for leases.
How AI might have helped
This undertaking is near a traditional demand and provide downside, with the twist of the rental vs retail income that makes it a bit extra fascinating. However discovering the equation that provides the optimum worth is just not the primary problem. The important problem is estimate every parameter given inadequate information.
Certainly, predicting future demand is difficult: you solely have a number of months of historical past (at greatest) for every type, and it’s essential predict a big horizon (principally as much as finish of life). Fast modifications in vogue traits require a deep understanding of the trade to be predicted, if predictable in any respect. And the uncertainty created by the early Covid interval made any time sequence fashions very laborious to construct.
Estimating pricing elasticity isn’t any simpler. As Lease the Runway was not a retail enterprise, gross sales information was by design restricted.
And that’s precisely the place the problem would come for any AI-driven answer as nicely. An AI can solely be pretty much as good as the information it’s being supplied.
Fixing for the sparse style-level information
Despite the fact that every type has restricted historical past, there’s a wealth of data in related objects. It is a prime use case for switch studying and shared embeddings that might have been made simpler by the entry to pre-trained LLMs. Shared style-level embeddings might have allowed us to make sturdy assumptions on new types primarily based on metadata: colour, model, worth, material, silhouette… We might have extra successfully constructed fashions that learn to predict demand curves from a number of information factors, drawing from patterns in traditionally related objects. An organization like Stitch Fix has been pioneering this house by utilizing merchandise metadata to create deep embeddings that generalize throughout new stock.
Maintaining with Quick vogue cycle
LLMs might have made it simpler to comply with and perceive ever-changing vogue traits and work on exterior indicators to foretell potential shifts in the complete trade. That was not one thing that was simple in 2020, as a result of it requires scrapping huge quantities of information, discovering out what’s related and deciphering weak indicators. Immediately, that’s precisely what LLMs are good at. Corporations like Trendalytics do exactly that, scanning TikTok, Google Traits, and social media to floor rising patterns in silhouettes, colours, or influencers’ posts. That information would have been extraordinarily worthwhile to make an correct demand forecast.
Constructing a dynamic pricing Agent
One final thing that might have been enjoyable to discover, given at present’s know-how, is to construct an agent that will have modified the costs in actual time and learnt, by means of reinforcement studying, the optimum pricing methods by interacting with the atmosphere. That might have allowed us to verify the costs depend upon the type’s historic and future demand but additionally on the buyer options, i.e private rental and buy historical past, engagement, style, and many others. That might have introduced us nearer to what prime RL groups at Airbnb or Uber do, repeatedly adjusting costs primarily based on actual time demand and reserving chance.
These are among the concepts that I selfishly would have been tremendous excited to work on, however observe two necessary issues:
1. From a product perspective, it’s actually laborious to estimate (particularly now that I don’t have entry to the information anymore) what the impression on general income would have been.
2. These concepts might have additionally been constructed in-house again in 2020, given the good crew of ML engineers we had at Lease the Runway. However it could have represented months — if not years — of analysis and growth with excessive dangers, which we couldn’t afford at the moment.
And that’s most likely my important takeaway to date on LLMs: they don’t trivialize the issues we used to bang our heads at 5 years in the past (or not but) however they make it simpler to check concepts that will have taken an unrealistically very long time to develop again within the days. This modifications the paradigm during which Knowledge groups usually function and opens new alternatives of partnership with Product groups.
Article 100% human written
