LLM applications: an investing framework

Published on
August 1, 2023
Chandar Lal
No items found.
LLM applications: an investing framework

In the summer of 1999, the Economist published a special report on “The Net Imperative”. It led with a bold prediction from Intel’s chairman:

In five years' time, all companies will be Internet companies, or they won't be companies at all.

This would have sounded like hubris during the ensuing dotcom collapse, but it ultimately came true. A massive platform shift took place – and now the default state is to be an “Internet company”, whether you’re selling books, handling payments, or running taxis.

Fast forward to today: large language models (LLMs) and generative AI are emerging as a shift of similar scale.

Some of today’s tech executives see AI as “more profound than fire or electricity” – set to usher in an “age of abundance”, or perhaps to destroy human civilisation. There is a lot of hype – but it’s reasonable to expect that LLMs and generative AI will be a foundational layer for the next decade’s tech companies, and this opportunity is easily big enough to excite us.

The generative AI venture landscape has blossomed rapidly, from applications to infrastructure and tools. In this post, we focus on the application layer – where we have historically been most active. It seems inevitable that generative AI will be embedded in most applications we touch, enabling entirely new experiences as well as large-scale automation. Less certain is the pace of this change, and its second-order effects. It’s one thing to tinker with ChatGPT; it’s quite another to adopt a product that structurally reinvents a long-held human task.

And when this adoption ramps up, what happens to the nature of work and the structure of the workforce? We take an optimistic view: as our venture partner, Benedict Evans, puts it:

We should start by remembering that we’ve been automating work for 200 years. Every time we go through a wave of automation, whole classes of jobs go away, but new classes of jobs get created. There is frictional pain and dislocation in that process, and sometimes the new jobs go to different people in different places, but over time the total number of jobs doesn’t go down, and we have all become more prosperous.

This is a Schumpeterian moment, and a generational opportunity for startups. But we are no longer in an era of zero interest rate policy, and not all boats will float despite the abundance of capital being deployed. So how do we navigate this emerging landscape?

Mosaic: long-term investors in machine learning applications

At Mosaic, we’ve been investing in applications of machine learning since our firm’s inception. The common thesis, unsurprisingly, has been to invest in products that solve hard and economically valuable problems at superhuman scale – and are differentiated by their proprietary models and/or datasets.

Selected machine learning investments, 2015-23

Looking forward, generative AI will be an equally central component of our investment strategy. We expect that AI-native companies will dominate the next wave of product innovation – but as investors, the key questions and heuristics that matter are changing.

Evaluating generative AI applications: the key questions

New foundational models and tools make it easier for teams to build viable AI applications, with little training data or computing resources required. That sounds promising on paper. But if this is true, and proprietary datasets confer less of an advantage, what do the new moats look like? If it’s easier than ever to ship an AI-native product, but harder to build a truly differentiated one, is it incumbents who stand to win – benefiting from their head start with distribution? Could this a structural problem for new entrants?

There are many hard questions for an investor to explore. Generative AI applications are now proliferating: we’re currently tracking over 500 early-stage European startups offering LLM-powered applications. So we need to be judicious about picking the right problem to solve, a product approach that is truly novel, and the most effective route to market.

In the spirit of transparency, here are a few of the qualitative questions on which our investment decisions often hinge.

Market opportunity. For some of the most exciting AI products, there is no “current state” offering: they are greenfield opportunities, enabling a product that simply wasn’t possible before, or serving an entirely new audience. It’s notoriously difficult to estimate the market potential of these. As an example: what if we could make an AI-powered private tutor accessible to the entire population of high school students? If we could offer human-equivalent personalised tuition to everyone, what would that be worth (in terms of new demand and its price elasticity)? More on that question here...

Most opportunities, conversely, are brownfield – either augmenting or replacing an existing human activity. Here, we ask: which human tasks can be meaningfully assisted or even displaced by a novel AI application? What economic value is generated by this human work? How expensive or difficult is it; and how scarce are the people capable of doing it? Hence, what cost efficiencies or scale benefits are achievable through automation? This points us to an initial view of the market potential of automating a task that historically couldn’t be.

Product. Our first question is: what entirely new capabilities can be unlocked using a LLM? Crucially, what does a novel product need to look and feel like? The model alone is not a product, and text boxes for prompts will rarely be the form factor that eventually wins. How do you serve up a model in a way that users are intuitively able to interact with? Looking further into the future, how durable is the value being created? For example: if every sales outreach email you receive is highly personalised, what relative advantage will any one of those emails have?

Defensibility. One of the hardest questions for any emerging AI application is simply: where is the moat? There is a common set of questions we often ask. Is the product a 'thin layer' on top of a general-purpose LLM that delivers adequate performance – or is there substantial, domain-specific training or fine-tuning required? If there is a unique dataset needed to build a compelling product, does the team have access to it? Are there specific architectural features (prompt engineering / chaining; multi-model) that contribute to a noticeably better outcome? Does the company have a clearly defined and distinctive route-to-market to sell its solution? Is the application in a regulated industry, that has high barriers to entry which might deter competitors?

Go-to-market. Our hypothesis is that if generative AI products will become even easier and cheaper to build over time, then distribution / go-to-market execution is what will separate the truly great from the good. As a result, we lean heavily towards founding teams that can demonstrate strong commercial instincts, with a clear understanding of who their buyer is, and how to reach them most effectively.

The European AI opportunity

The generative AI wave is a global opportunity, but Europe is particularly well positioned to capitalise on it. The continent has produced its first crop of AI-native unicorns, and we think this will only accelerate, given:

  • its depth of research output and talent (especially with the UK's "Golden Triangle", France's Polytechnique, ENS, INRIA, and CentraleSupelec, and leading ML institutes in DACH);
  • its flywheel of company creation;
  • its leadership in setting global policy on safety, privacy, and alignment.

Conclusions and unknowns

Clearly, we’re very optimistic about the next generation of generative AI applications that will emerge in Europe. Yet at this point, we have far more questions than answers. This is little surprise, as we’re at the very beginning of a major platform shift, where the first major applications are just beginning to emerge. There are, of course, risks that we are observing carefully – privacy, safety, alignment – where Europe appears to be an emerging leader in setting the guardrails.

But if we allow ourselves to dream, it’s easy to imagine a new era of productivity and creativity – where applications powered by LLMs help us interpret or generate new information. The result could be a world where every person is assisted in their desire to create beautiful art, ship production-ready code, or simply edit a blog post – as GPT-4 has done here!

If you’re building a generative AI application, we’d love to chat, and learn how you’ll reshape your market and capture enduring value. In our next posts, we’ll delve into specific industries that will be transformed by emerging AI applications, as well as the tools and infrastructure software that will underpin this wave.