of Shopify, not too long ago advised his staff in an inside memo: “Earlier than asking for extra headcount and assets, groups should reveal why they can not get what they need carried out utilizing AI”.
Having labored in startups for the previous 6 years, asking for extra headcount or extra assets is often not an possibility anyhow. Constraints are tight and also you typically have to scrupulously put money into tasks you’re assured shall be impactful. So in these conditions, Tobi would most likely rephrase: “Suck it up and simply use AI should you 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 inventive. However can a multi-billion parameter mannequin, though it has learn your entire 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 tasks I’ve labored on because the starting of my profession and assess how a lot AI would have helped.
As we speak, we return to 2020. I’m a junior Knowledge Scientist at an organization that has been hit fairly dangerous by the pandemic: Hire 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 notice that some enterprise constraints needed to be utilized to guarantee that we’d not unload a whole model and hold some items for leases.
How AI might have helped
This challenge is near a basic 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 will not be the primary problem. The most important problem is find out how to estimate every parameter given inadequate knowledge.
Certainly, predicting future demand is difficult: you solely have just a few months of historical past (at greatest) for every model, and it’s worthwhile to predict a big horizon (principally as much as finish of life). Speedy adjustments in style 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 arduous to construct.
Estimating pricing elasticity isn’t any simpler. As Hire the Runway was not a retail enterprise, gross sales knowledge was by design restricted.
And that’s precisely the place the problem would come for any AI-driven resolution as properly. An AI can solely be nearly as good as the info it’s being offered.
Fixing for the sparse style-level knowledge
Though every model has restricted historical past, there’s a wealth of knowledge in related objects. It is a prime use case for switch studying and shared embeddings that would have been made simpler by the entry to pre-trained LLMs. Shared style-level embeddings might have allowed us to make robust assumptions on new kinds primarily based on metadata: colour, model, worth, cloth, silhouette… We might have extra successfully constructed fashions that learn to predict demand curves from just a few knowledge factors, drawing from patterns in traditionally related objects. An organization like Stitch Fix has been pioneering this area by utilizing merchandise metadata to create deep embeddings that generalize throughout new stock.
Maintaining with Quick style cycle
LLMs might have made it simpler to observe and perceive ever-changing style traits and work on exterior indicators to foretell potential shifts in your entire trade. That was not one thing that was straightforward in 2020, as a result of it requires scrapping large quantities of knowledge, discovering out what’s related and decoding weak indicators. As we speak, that’s precisely what LLMs are good at. Firms like Trendalytics do exactly that, scanning TikTok, Google Developments, and social media to floor rising patterns in silhouettes, colours, or influencers’ posts. That knowledge would have been extraordinarily useful to make an correct demand forecast.
Constructing a dynamic pricing Agent
One last item that would have been enjoyable to discover, given at the moment’s know-how, is to construct an agent that will have modified the costs in actual time and learnt, by reinforcement studying, the optimum pricing methods by interacting with the atmosphere. That might have allowed us to ensure the costs rely on the model’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, constantly adjusting costs primarily based on actual time demand and reserving likelihood.
These are a few of the concepts that I selfishly would have been tremendous excited to work on, however notice two vital issues:
1. From a product perspective, it’s actually arduous to estimate (particularly now that I don’t have entry to the info anymore) what the influence on total income would have been.
2. These concepts might have additionally been constructed in-house again in 2020, given the nice crew of ML engineers we had at Hire the Runway. However it could have represented months — if not years — of analysis and improvement with excessive dangers, which we couldn’t afford at the moment.
And that’s most likely my most important takeaway to this point 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 adjustments the paradigm through which Knowledge groups sometimes function and opens new alternatives of partnership with Product groups.
Article 100% human written
