Here is what you would like the model value graph to look like:This means that your model will generate value even though initially its accuracy may be low.
Then, as you collect more data and improve the model, prediction value increases.
This pattern establishes the virtuous cycle between better models and better products, and ensures that a competitor with less data will have a significant disadvantage.
From better models to better productsIn practice, it is very difficult for a single model to conform to this model.
But a successful AI product doesn’t depend on just one machine learning model.
Instead, it harnesses the synergies between an ecosystem of models that collectively improve as more data is generated.
Think again about Uber.
In addition to ETA prediction, UberX, the Uber service I discussed earlier, is also powered by models that predict passenger demand and trip patterns.
Individually, each of these models may plateau, but collectively their prediction value can keep increasing for a longer period of time.
As you generate more and more data, each of your models can become more accurate.
Over time, models whose prediction value was previously low, as in the insurance company case, can improve enough to generate substantial value.
Eventually, you may unlock the potential to leverage this new cohort of models to power new products, creating even more value and an even deeper moat.
The AI S-curveConsider Uber POOL, a newer Uber service that allows passengers traveling in similar routes to share a ride and split its cost.
In addition to the UberX models, POOL also relies a set of models that require higher accuracy and sophistication, such as predicting if an additional passenger will join a trip in progress.
This model has relatively low error tolerance because a wrong prediction may result in a loss for the company.
As a product, POOL unlocks a new market of more price-conscious customers.
This is clearly very valuable for Uber, meaning that the models powering POOL have a high prediction value.
Let’s plot the model value graph for the UberX models and the POOL models together.
As in the ETA prediction example, the models powering UberX start generating value at relatively low accuracies and scale quickly, then plateau.
This results in an S-shaped curve.
Once accuracy is high enough, the POOL models kick in, and together with the UberX models continue to scale prediction value.
Tracing the prediction value for all the models together reveals the steady increase in prediction value behind Uber’s deep AI moat.
This pattern resembles the innovation S-curve, which identifies the opportunity for innovation and disruption in the window between one technology maturing and a new one just beginning to emerge.
Similarly, the key strategic question in our case is how to identify when the potential for a new cohort of highly valuable models has emerged and how to leverage it in product innovation.
I will continue to explore this topic in future posts.
The bottom lineAI and data have the potential to create deep defensive moats.
The key is to establish the link between better models and better products.
Models become better as more data is generated, but the prediction value of any individual model tends to plateau.
By strategically harnessing new cohorts of more powerful models and the product opportunities that they unlock, successful companies can scale the business value generated from their data and build a deep AI moat.
.. More details