James Quincey, Coca-Cola CEO said in 2018: “Our business is Built to Perform”. While it is clear how built to perform refers to the optimisation of a well-understood business model, it is perhaps less obvious how it applies to AI startups, which - like hardware - have high costs (see Google’s NeurIPS and Microsoft’s paper) and lengthy R&D periods before value is realised. To identify what is being built to perform in the market, it would be useful to segment the AI startup industry into AI-Later, AI-First, and AI-Frontiers. This is what I do below, but first I address what does built to perform mean.

Built to Perform, in my definition, means having 1) a clear, tangible product/advantage, and 2) a plan to bring it to the market or in some cases to create the market for it. Of-course for Coca-Cola the legacy product is clear, a unique taste. While the consumer might not have needed that taste in the first place, the marketing strategy around the brand has historically made it impossible for some to have their meal without it. The second point involves being able to recognize what is the current state of the market you are in, what are the trends to anticipate, and how the brand should adjust to stay relevant. Hence, given that Coca-Cola already has the point 1 nailed, it continues with point 2 today by becoming a total beverage company (by for example acquiring Costa in the UK). Quincey, the CEO of Coca-Cola, isn’t a fast moving tech-celebrity-founder-CEO, but he’s a pragmatist who realizes Coca-Cola has to change or be left behind. The world is changing: regulators as well as the consumers express skepticism towards high levels of sugar, competitors evolve, and markets die out. Any company’s future will not be as glittering as its past if it doesn’t keep up.

So, who to keep up within the AI-start up sphere? For that we need to look at the second point mentioned above - what is the current state of the AI market? In the current state, there are product driven and AI assisted companies realistically described as AI-Later. These originally had a product built that could perform without AI, such as Netflix, Amazon, Facebook, Twitter and the majority of other existing enterprises and appearing startups. The AI was put on top of the existing products, and the products were further enhanced. The same applies to Google, even-though it is claiming to be AI-First, realistically it is not - it organizes world’s information, not worlds AI. The above mentioned businesses have managed to hire most of the top AI researchers and open up prestigious labs like FAIR or Google Brain, that still does not make them AI-First. It is not their AI that made them perform in the first place.

The real elephants in the room are the only AI-First companies, that in time will make the AI-Later companies motive to brand themselves as AI or AI-First obsolete. The companies that build AI tools as easy-to-use scalable and flexible services. There are multiple, already well established players within this domain. One, is H20.ai, that purely sells an Auto-ML capability. However, just running a hyper-parameter (either Bayesian Opt. or Evolutionary Algo.) search over the existing set of supervised learning tools by spinning up large quantities of instances on Google Cloud Platform or Amazon Web Services is not sufficient to compete. For AI-First company, the important thing to realize is that AI tools, just like humans, progress by abstracting away from the already solvable tasks - supervised learning is more or less a solved problem today, and the AI community has moved towards learning to learn. Likewise, an AI-First Auto-ML business would seek to abstract away from simple hyper-parameter search towards Automated Statistician like tools. Where for a given machine learning problem a prior over models and hyper-parameters to try would be inferred/meta-learned. An interesting example is Mind Foundry which appears to specialize in consulting at first, but do consider the broader path that Mind Foundry has taken. By consulting they collect datasets of bespoke ML problems and therefore would be able to build AutoML not just for tabular data - or at least this is what they intended to do back in 2018 (Alessandra Tosi gave a talk at our reading group). There are other notable ML consultancy companies such as ElementAI, Prowler or Invenia, that are founded by notable researchers in machine learning. However, consulting is not the end but should be treated as a mean to the end - Auto-ML/Automatic Statistician. There is a reason why Google is actively pursuing Auto-ML direction, but I do find Google’s attitude towards more and more computation a warning both for practical as well as innovation reasons.

If you assume for a second, that Auto-ML becomes sufficiently general (and most likely it will, judging by projects like Ludwig from Uber, or the work at Google, etc), then the only option for an AI company and all of their expensive scientists is to focus on a specific hard open question in a specific domain. Here you have research driven AI-Frontiers enterprises like Deep Mind (solving AGI, although this one is very speculative), or a startup that I am part of - Cervest. In the case of Cervest, we build bespoke, frontier Earth Science AI which does not fit into the standard supervised learning setting amenable to Auto-ML. Our long term view is to build recomender tool that would bring decision support for soil, crop, climate and other Earth’s resources management, and harmonise the objectives of the eco-systems of the planet, and it’s people. This involves, not just research into AI and earth science, but also into policy and ethics. AI-Frontiers startup brings business & social value by conducting frontiers research - it is it’s core to building to perform.

Of-course, just like Coca-Cola (or any other business) everything has it’s time to come, and time to leave. It is fair to say that AI-Last companies are the first adopters of AI, but AI is not what built them to perform. AI-First companies are building AI as a general service to perform. And we, humans, progress by abstracting away from the tasks we can solve, and as such the AI-Frontiers startups move to challenges that require frontiers research to bring solutions we all need.

Cite as:

@article{gamper2019aifrontiers,
  title   = "From Coca-Cola to AI-Later and AI-First, to AI-Frontiers",
  author  = "Gamper, Jevgenij",
  journal = "https://bruteforceimagination.com/",
  year    = "2019",
  url     = "https://bruteforceimagination.com/blog/2019/coca-cola/"
}