How machine learning can live up to the hype
Overcoming the hidden obstacles of AI company-building: interdisciplinarity, hybrids and new business models
Co-Written by Adrian Locher, Finn Grotheer and John McSpedon.
This article is part of a series of articles about Merantix, our existing portfolio companies, the business cases we are currently working on, as well as the domains we deem pivotal and would like to further explore. In this article, we will outline crucial ingredients for successful AI-company building — some of which have been notoriously missed in previous attempts of taking machine learning from research into real-world applications.
The world’s first AI venture studio
We started Merantix, the world’s first AI venture studio, just four years ago with the conviction that AI would revolutionise value creation, alter economic and societal structures and contribute to a higher and ubiquitous standard of living. Since then, we have raised our first €25M funding round, built a team of innovative and driven talents, and incubated three companies in healthcare, data management and industry solutions. But we have also perfected our approach to AI company building along the way, closely analysing every step we took. We deem some of the lessons-learned as absolutely crucial and want to share them: as a contribution to a vivid and curious ecosystem of ingenious entrepreneurs, but foremost in recognition of the infinite potential that AI harbours. For everyone.
AI-first companies will face some particular challenges that traditional software firms were never confronted with
In 2016, we started with more than 100 conversations with C-level executives of European corporates and Mittelstand firms, through which we identified about 400 business cases. Subsequently, we developed a methodology to funnel and filter all potential business cases and set up a process to validate the most promising ones both in terms of technological feasibility as well as venture case readiness. Some of the most important insights quickly dawned on us. AI-first companies will face some particular challenges that traditional software firms were never confronted with — and that we would need to flag these particularities openly and head-on.
Understanding the stack
Modern AI is a very fast-paced, collaborative, and competitive research field. Deep learning has offered a steady stream of breakthroughs, shattering performance benchmarks and enabling formerly unimaginable applications, since its modern advent in 2012. As a result of the community’s size, its decentralised structures and high pace, it is not only necessary to fully fathom the landscape of already existing companies and available open source code at the start, but to monitor what is happening in the periphery while moving along the incubation process.
It is important to understand that in-house research findings are not a defensible business model per se — you’re unlikely to have even a 6-month head start on the competition. Similarly, any infrastructure built is guaranteed to be eventually eclipsed by an open source community that’s very smart and thousands strong. Young companies need to deploy their resources wisely, using the available research and infrastructure to their advantage instead of grinding against the inevitable progress of the wider AI community.
As a result, we continuously visit research conferences, host paper discussion and are an active member of the startup community. Beyond staying on top of research and the latest company ideas, we seek moated expertise through specialisation in specific domains, tailoring bespoke solutions and unique go-to-market strategies.
In a recent article by Techcrunch’s Danny Crichton, he describes what he calls “the dual PhD problem” as the pinnacle of entrepreneurial development: the fact that throughout the last decades we have burnt through an innovative wave of incredibly smart college-dropouts that revolutionised aspects of public life and enterprise organisation by an autodidactic intuition for software. In Crichton’s reading, it is likely that that era is coming to an end with the rise of machine learning.
Thoroughly understanding two domains right at square one and building valuable relationships for data acquisition from the beginning is a heavy obstacle for a young company.
For one, ML technology is complex and dynamic. Its latest leaps forward require a high attention span to PhD-level research that happens every day. More importantly, in most instances, machine learning serves as an auxiliary technology that is utilized in order to improve performance or availability in very specific domains. However, the fields in which we tend to see the highest potential for the impact of AI — healthcare, synthetic biology, mobility, for example — are all themselves rapidly developing spheres with peculiarities, regulatory depths and powerful legacy players. In order to build truly disruptive companies, one needs to understand not only machine learning but also an additional domain. Yet, thoroughly understanding two domains right at square one and building valuable relationships for data acquisition from the beginning is a heavy obstacle for a young company.
As a result, Merantix works intensely on interdisciplinary team building and facilitates a process of true synthesis. That includes a close circle of advisors, industry contacts and partner companies. Furthermore, the studio’s funding promise and attractive company shares for early team members during the incubation period enable our founders to recruit high-level experts from specific domains as co-founders or first team members, allowing for “two PhDs” right from the start.
Understanding human-machine interaction
The most common, non-technical conception of AI as a technology is that it enables high levels of automation, taking the human out of the equation while maintaining a human-like performance. Ironically, of course, machine learning requires a high degree of human input when building a robust ML system. Cleaning and labeling datasets depends on manual and repetitive labour, testing, bug fixing, and ongoing maintenance. As Martin Casado and Matt Bornstein have recently remarked, this process entails lower gross margins for AI companies across the board compared to software startups.
Furthermore, human-machine interaction does not stop. Due to high complexities, which the current state-of-the-art ML technology can not yet adequately address, as well as regulatory requirements, which hamstring machine-only solutions, often enough humans will be plugged into AI systems in real time, leading to “hybrids”. That entails both downsides and upsides: human interference leaves the system vulnerable for biases, while, paradoxically increasing societal trust. The human-in-the-loop also enables labeling in production which tends to improve performance and reliability. Either way, balancing the right level of manual quality control and human accountability as demanded by society and the regulator, on the one hand, with the aspired benefits of automation and machine accuracy, on the other hand, is challenging and presents many pitfalls and potential regulatory deaths.
So, at Merantix, we focus from day one on these questions and review them continuously. We sidestep risks and increase chances of wide-spread adoption by never fully taking the human out of the equation. Our systems are always subject to improvement by continuous labeling in production and are designed to conform with all relevant regulations from day one.
Scalable business cases
While the technology of machine learning has seen incredible progress throughout the last decade, commercialisation and widespread penetration have not followed on par. Identifying both the most promising business cases as well as formal requirements for its adoption take time and expertise, as we have outlined above. At the same time, there are also structural challenges: machine learning can very effectively automate so-called high-scale/low-complexity cases. Yet, as Casado and Bornstein have pointed out in a later article, for many tasks, companies are surprised to find that many of their cases have a long tail, i.e. a majority of their problems do in fact not occur regularly (think: Google search requests).
That impedes the possibility of applying machine learning algorithms efficiently since a lot of training data is required for comparatively rare instances that aggregate to the majority of cases. This problem gets worse when an AI company starts to scale its sales: the training data collected and used for the first customer may not help when dealing with a supposedly similar second customer; often due to banal questions of image quality, camera positions or output requirements. As a result, the success defining moment for an AI-company occurs very early in the validation phase. If you miss to check for the level of data standardisation within a domain, your company may be dead in the water long before you notice it.
The emerging trade-off between the portion size of the value-chain, that you are trying to tackle, and the level of standardisation, which you need for high scalability, is a main challenge to think about. It necessitates intensive validation processes and high frustration tolerance. Finally, companies that find ways to provide customisable and self-service solutions tend to do better, all else being equal.
At Merantix, we allow long and thorough validation processes that include commercial aspects. Having confirmed technical feasibility, we implement minimum-viable-products with industry partners to understand customer needs, data quality and specific market challenges. Only once we have confirmed a go-to-market strategy that convinces industry insiders at first sight, we start to build a team and scale our companies.
Beyond ML expertise
Companies that can capitalise on a wide network of experienced entrepreneurs, industry contacts and domain experts are much better positioned than the typical 2000s lone entrepreneur in a college dorm.
Thus, mastering machine learning does not suffice. Building AI companies sustainably happens at the intersection of elaborate commercialisation strategies, intelligently concerted interdisciplinary teams and balanced trade-offs along the development process. In turn, companies that can capitalise on a wide network of experienced entrepreneurs, industry contacts and domain experts are much better positioned than the typical 2000s lone entrepreneur in a college dorm.
As a venture studio, we aim to provide all these factors, facilitate a vivid exchange of ideas, foster an open source approach and build an ever-enhancing platform of cutting edge AI company building. Our first companies have already been spun-off and we are working on the next generation of companies — in business intelligence and bio-tech. If you are interested in these new companies, stay tuned for more articles to come. And if our approach resonates with you, we would love to hear from you! Feel free to reach out to us or to apply directly here: https://merantix.bamboohr.com/jobs/.
About the Authors
Dr. Rasmus Rothe is co-founder and CTO of Berlin-based Merantix as well as a founding board member of the German Association of AI Companies (KI Bundesverband e.V.). He has published over 15 peer-reviewed papers while attending Oxford, Princeton, and ETH Zurich, where he received his Ph.D. in computer vision and deep learning. E-Mail: firstname.lastname@example.org.
Adrian Locher is co-founder and CEO of Merantix. He has been a serial entrepreneur and investor for the past 20 years, founding more than 10 companies both in Europe and the US, in digital healthcare, e-commerce and AI. E-Mail: email@example.com.
John McSpedon is a machine intelligence engineer at Merantix. After studying at Princeton and programming in silicon valley, he joined as the first employee and has worked in a technical capacity on all Merantix ventures. E-Mail: firstname.lastname@example.org.
Finn Grotheer is a public affairs fellow at Merantix. He is a graduate student of International Affairs at the Hertie School of Governance and a fellow of the German Academic Scholarship Foundation. Before joining Merantix, he gained work experience at the Boston Consulting Group and Hering Schuppener Consulting. E-Mail: email@example.com.