ML in 2019: Some Quick Predictions
Three predictions for applied ML and two more macro predictions
Happy 2019, ML practitioner!
As it’s now (almost) a tradition at Point Nine, we will soon publish a team prediction post in which each of us will explain what they look for next year. In case you missed them, here is the one from 2017 and 2018 —needless to say, we/I have been completely wrong many times!
I started thinking about a few ideas of what I look forward to in the broader ML space in 2019 in late December but never hit the publish button. Today is Jan 1st, here they are —hope you’ll enjoy!
1. More GANs
A year ago, I wrote that in 2018 machines would become creative (check the post here) thanks to generative design software that use Generative Adversarial Networks (or GANs). At the time, Autodesk’s car chassis or iMuze’s algorithmically generated songs were interesting illustrations. It seems that this prediction was not completely off as a myriad of new applications of GANs appeared in 2018. Uizard managed to generate front-end code from design sketches. Nvidia created a game environment. A new art movement that some called GANism even emerged — Christies sold the first piece of algorithmically generated art in XX this year. I am sure we’ll keep on seeing more and more applications of GANs next year.
2. Algorithmic feed(s) in B2B software
B2B products are often inspired by consumer products. As consumers, we’re now used to enjoy the very customised experiences offered by most consumer apps. Our feeds on Twitter, Facebook and Instagram are all our own and differ from our friends’ ones. Going even further, the Chinese app TikTok’s “is like TV but without remote control”, AI chooses the content for the user. All these products rely on (sophisticated) logistic regressions that try to predict which pieces of content we’re most likely to appreciate (this article explains how in greater detail). Spotify and Netlifx use collaborative filtering extensively to recommend us the next song / movie (check the Netflix Prize if you’re interested). We could even argue that these services are not only defensible businesses because they have the largest libraries of content but because their dataset on our personal preferences (partly based on engagement metadata) is proprietary (and very hard to reproduce). The more we use these services, the more metadata we give them, the better is the feed, the less likely we are to churn. Data network effects.
In 2019, I believe that we’ll see more and more B2B software offering individualised content and feeds based on the collection of such metadata. From dynamically optimized registration flows down to personalised farm management software for dairy farmers, the opportunities are endless.
3. UI(s) designed to improve AI(s)
The best ML-powered products are designed to closely ingest user feedback. The faster they can ingest user feedback, the faster they can retrain algorithms and the more accurate becomes the prediction (a.k.a. data network effects start kicking in). The shorter this loop is, the faster the user experience improves and the product becomes defensible. One way to optimize for fast user feedback is to build user interfaces (UIs) purposely designed to collect the data that is critical to the prediction work. As an example, Rossum.ai is building the next generation of data extraction algorithms. Rossum’s product UI has been designed to facilitate the validation of the data extraction work performed by computer vision algorithms.
4. Data monopolies in B2B markets?
The Congress has not yet decided whether or not it should break Facebook’s (data) monopoly — in part because traditional market concentration tests fail (see this post). Yet, 2018 was definitely THE year in which the population started realising its (almost) outrageous market power. It might take some years before such debates happen in the B2B world.
But I bet that in 2019, we’ll see more and more B2B startups arguing that they will build market domination by building data monopolies. It’s not yet certain that i) this is possible (in every market) and ii) that the regulators will give them enough time to build them but I would happy to bet on it (see more in our AI-first Investment Thesis).
5. AI Nationalism
Ian Hogarth wrote a great post called AI Nationalism in June last year (see here), in which he argues that nations will try to use the commercial data assets created by their companies as “intelligence weapons” against other countries. End of last year’s scandal with Huawei proves he is right. I suspect that we’ll see more of these scandals in 2019…unfortunately.
Despite the (significant) risk of an economic downturn this year, I am excited to be back at work in 2019 to keep on learning more about how machine learning will keep on reshaping industry after industry.
I also read somewhere that the best VC deals are made during economic cycles’ low points. Have a great start into the year 🚀
