At the Rise of AI Summit 2020, Dr. Rasmus Rothe, CTO and co-founder of the Berlin venture studio Merantix, presents trends in AI industries, technology and business models. He describes Merantix's model of working with individual founders for six to twelve months, investing up to three million euros per company, and building AI-first ventures such as Vara in breast cancer imaging, a data-management company he refers to as Sea Search, Merantix Labs for custom client solutions, and the newer Causa and Cambrian, with plans for eight more companies over the next four years, for ten new ventures in total. He argues the next frontier lies in 'dual PhD problem' areas that combine deep AI expertise with equally deep domain expertise, such as healthcare or synthetic biology, rather than the execution-heavy consumer startups of the past decade. On technology, he highlights GPT-3's shift from task-specific fine-tuning to a single 175-billion-parameter model trained on broad internet text, demonstrated through email drafting, balance-sheet filling and code generation. On business models, he outlines a spectrum from fully standardized products to full consulting, arguing the most promising middle ground is self-service, customizable AI products.
Key takeaways
- Merantix works with individual founders for six to twelve months, investing up to three million euros per company before helping it scale.
- Rothe frames the 'dual PhD problem', combining deep AI expertise with deep domain expertise, as the next frontier for AI startups, citing synthetic biology and healthcare.
- GPT-3's 175-billion-parameter model, trained on broad internet text, could perform many language tasks with little task-specific fine-tuning, a shift from the prior labeled-data paradigm.
- Rothe places AI products on a spectrum from fully standardized (Vara) to fully customized consulting (Merantix Labs), with self-service, customizable products as the most promising middle ground.
- Merantix found that roughly 80% of client problems across different industries turn out to be similar despite differing business processes, favoring self-service products over one-off consulting.
Detailed notes
Merantix's venture studio model
Rothe explains that Merantix works with individual founders or potential founders for six to twelve months, testing and validating business ideas before incubating a company, investing up to three million euros per company and then helping it scale, a model it has run for four and a half years. Portfolio companies include Vara, a live medical product for breast cancer diagnostics; a data-management company for large unstructured datasets that he refers to as Sea Search; and Merantix Labs, which builds customized AI solutions for clients. Two new companies, Causa, focused on explaining KPI changes through causal inference, and Cambrian, at the intersection of biology and machine learning, started that summer, with a plan to build ten new companies in total (eight more beyond Causa and Cambrian) over the next four years.
The 'dual PhD problem' as the next frontier
Rothe contrasts the past decade's 'lone dropout founder' paradigm, execution-heavy consumer businesses built by small teams with narrow domain expertise, with what he calls, citing a TechCrunch framing, the 'dual PhD problem': areas where a founder needs deep AI expertise plus equally deep expertise in another domain, such as healthcare regulation or synthetic biology, that cannot be picked up in a few weeks. He says this changes how teams should be assembled, describing how Merantix ideates with one founder first and then matches a co-founder from the relevant domain, as it did for Cambrian in synthetic biology.
Technological achievements: the shift to broad language models
Rothe describes a shift in focus from computer vision toward text, noting the average worker spends about 11 hours a week on email alone. He contrasts the earlier natural language processing paradigm of training on task-specific labeled data with GPT-3, released in early 2020, which trains a single 175-billion-parameter model on broad, largely unlabeled internet text so it can be applied to many tasks with little further fine-tuning. He demonstrates applications including automatically drafting a full email reply from a few bullet points, filling out a balance sheet, answering factual questions, and generating the code for a web page, and lists capabilities such as classification, information extraction, semantic search, summarization, translation, code translation and text generation.
Business models: standardized, customized, and self-service
Rothe describes work with the University of St. Gallen to extend the business model canvas with AI-specific elements such as data, safety and the AI lifecycle. He places AI products on a spectrum: fully standardized vertical products like Vara at one end, fully customized consulting like Merantix Labs at the other, and in between, self-service products that are productized but customizable, like Sea Search. He says Merantix underestimated this middle ground five years ago, but now sees that roughly 80% of client problems across different industries turn out to be similar even when their business processes look different, making self-service solutions increasingly viable, provided they remain generalizable across customers, differentiate from cloud-provider and open-source offerings, and match the technical capability and data availability of the customer.
Q&A highlights
Asked about investing personally in Merantix, Rothe says direct equity is not offered but private individuals above a certain threshold can invest through its roughly 25-million-euro fund; Merantix builds companies operationally rather than only writing checks, limiting it to about eight more companies in the next three and a half years under the current setup, while bringing in outside investors from Series A onward, as it did for Vara. He says Merantix deliberately diversifies the sectors of its portfolio companies and avoids starting competing companies too close together so they can collaborate openly. On the future of AI-drafted communication, he suggests people may eventually exchange structured bullet points that a model formats into full text on the receiving end, and confirms Merantix already has internal prototypes for generating text in a person's own writing style from their historic emails.
Quotes
We're basically like glorified text processing machines.
Rasmus Rothe
The average worker spends 11 hours a week just on emails.
Rasmus Rothe
Questions answered in this talk
What is the 'dual PhD problem' in AI startups?
According to Rasmus Rothe at the Rise of AI Summit 2020, it refers to startup areas where a founder needs both deep AI expertise and equally deep expertise in another domain, such as healthcare or synthetic biology, expertise that cannot be picked up quickly and usually requires pairing an AI expert with a domain expert.
How did GPT-3 change the standard approach to natural language processing?
According to Rasmus Rothe at the Rise of AI Summit 2020, GPT-3 replaced the earlier practice of fine-tuning models on labeled, task-specific data with a single 175-billion-parameter model trained on broad internet text, allowing it to handle many applications, from email drafting to code generation, with little further fine-tuning.
What is Merantix's venture studio investment model?
According to Rasmus Rothe at the Rise of AI Summit 2020, Merantix works with individual founders for six to twelve months to validate business ideas, invests up to three million euros per company from a roughly 25-million-euro fund, and brings in outside investors from the Series A round onward.