Generative AI: How we are investing at firstminute

Insights17 Jul 2023

By Michael Stothard, Sam Endacott and Lorcan Delaney

There are over 130 companies in the firstminute portfolio, and the great thing about working closely with all of them is getting a little glimpse into the economy of the future. The technology they are using today in their products and operations will likely be standard in 10 years time.

That’s why, as part of our work understanding the opportunity around Generative AI, and how we want to be investing (more on that later!) we polled them all about how they are using generative models in their companies. 

The results were startling. They showed that 84% of responders were using Generative AI tools internally for their work and 75% were using or thinking about using it in their actual products or outward-facing operations.

Why did this get us excited? Because it reinforced our house view that Generative AI is not a passing fad, but a transformation set of technologies that will have wide and lasting implications for real-world business and the whole economy.

It made us more bullish on the Generative AI enabled companies we have already backed - including Mistral (foundational model), Nabla (medical co-pilot), Klu (tooling), Lightning AI (AI build framework), Resemble (synthetic voice), Granola (note taking tool) and many more still in stealth. 

It made us bullish for the many of our other portfolio companies integrating Generative AI into their offering now and more generally the opportunity over the coming years to build transformational businesses on the back of this exciting platform shift.

This report on how we want to be investing in Generative AI which follows on from our one in January called "How to Make Money from Generative AI" starts with a survey of our portfolio.

It then goes into the wider Generative AI landscape, where we see value accruing, and where we want to be investing. We address questions about defensibility in the space and discuss other themes we are excited about. Hope you enjoy!

First, the Survey

Just over 40% of our founders responded to the survey, and most said that they were using Generative AI tools in their internal operations in a big way  giving a glimpse into how these tools are already becoming a normal part of workflows in early-stage companies. 

True, some in our portfolio were just experimenting with Chat GPT. But most were seriously using it across 4 core areas: 

  • coding (GitHub co pilot mainly); 

  • marketing and copywriting; 

  • automating routine office tasks; 

  • enabling customer success and sales teams.

One of our companies, for example, was using the Langchain framework to string together models that could a) scrape blockchain data b) match data to public Twitter accounts c) compose personalised outreach to those accounts allowing them to do effective marketing at scale. Others were using it to scrape and order unstructured data on global transport routes.

Note as well how the usage is higher with the newer companies. For our Fund I, a 2018 vintage, it's 71%. Our 2020 vintage Fund II it’s 82%. And then for our Fund III, which did an initial close just the other month, it’s 100% (albeit with only 3 respondents as the portfolio was very new at the time of the survey with 5 companies).

A similar kind of % of our companies were using or thinking about using Generative AI somewhere in their actual product offering (again with more interest by newer companies than older ones). The areas where Generative AI was being used could be split into three broad buckets: 

 1)  Companies offering a Generative AI feature as part of existing product: examples of this include an auto email writer in a productivity tool; text summarization, content generation, and conversational agents in a data cleaning company; allowing users to quickly generate code in an RPA platform; summarising conversations that happen on a labour marketplace platform; and drafting and filing IP docs for an IP platform.

2) Companies using LLMs for something a user might not notice, but behind the scenes makes product better: examples of this include speaking to NPCs in games; making core product better with understanding unstructured data; generating receipts faster for customers; automatically creating visualisations to explain queries to users; dataset creation and software recommendation; Slack bots for users; better customer support in general. 

3) Finally, there were companies that were Generative AI "native", where the core product was LLM based: An example of that is Nabla, which has built a medical copilot (see more press about them here and here) or Mistral, which is building a foundational model to rival Open AI.

Problem with the Tools

Were Generative AI tools working perfectly for our companies? No. In our survey, we asked about the problems with the tools they are using. 

Most and particularly the ones who were doing the most work on these tools talked about explainability, reliability, latency, data privacy, cost, visibility on costs and testing as major problems.

Everyone complained about hallucinations. Some thought that it was just too early (“wake me up in 3 years when these things are properly good,” said one). 

Wake me up in 3 years when these things are properly good

Others complained that even GitHub co-pilot was rubbish as the code was just never up to scratch and made reviewing a pain “I spend as much time fixing the code as I would if I just wrote it,” said one.

This is the sentiment being echoed in boardrooms across the world this annoyance combined with huge excitement. But we see this as an opportunity for a seed fund we are still early and there is still so much work to be done.

These challenges are an issue for some Generative AI companies (particularly those selling to enterprises) wanting to get revenue quickly. How many corporates are actually paying for generative ai tools now vs. how many are just “kinda interested”?

Ultimately though, it makes us more excited about the opportunity here as we are in the early innings of this technology and expect seed-stage entrepreneurs to solve many of these issues over the next few years  and hope to back them to do it.

How to Make Money from Gen AI #2

This survey is a nice jumping off point to give a bit more on the firstminute house view on where Generative AI is now and where it is going. As said above, our last piece, “How to make money from Generative AI”, was in January and a vast amount has changed since then.

The voice from our portfolio has reinforced our basic house view though, which is that the rapid advancement of LLMs and other generative models is part of a 20-year platform shift where it is inconceivable that there will not be huge startups built as a result. The big question is where to find them.

A 20-year platform shift where it is inconceivable that there will not be huge startups built

There are lots of critics out there, arguing that this is VC hype and we agree with some of their concerns. Are we worried about moats and defensibility? Yes. Do we know exactly where the value will be created? Where the incumbents will be empowered vs new entrants? What the tooling stack will look like? No, we do not.

But below we are going to talk about the areas where we are excited to invest in and where we are doing the work to keep on top of the rapidly evolving landscape. We are going to talk about: 

  1. The Gen AI stack today

  2. The tooling stack - and where we want to invest

  3. The application layer - and where we want to invest.

  4. The question of defensibility

  5. Open Source winning the day

  6. What we are looking for in a Gen AI seed company

So let’s get into it…

1) The Gen AI Stack Today 

From the hardware running the software, to the LLM-powered recipe app on your phone (for example), there are a whole heap of steps in that value chain.

If you have heard this all before, you can skip over this bit, but this is how we at firstminute are thinking very generally about the different categories:  

  • Chip Manufacturers: GPT-3 was trained on an enormous amount of text data, around 45 terabytes, which is equivalent to almost one million feet of bookshelf space. Traditional computer hardware is unable to handle this workload. To process such a vast amount of data across billions of parameters in parallel, large clusters of GPUs or TPUs with specialised accelerator chips are required. The demand for training and running models necessitates massive hardware, benefiting chip makers like Nvidia (check out its share price!).

  • Cloud platforms: GPUs and TPUs are expensive and in short supply, so much of the work involved in building, optimising, and running large AI models takes place in the cloud. This allows companies to conveniently access computational power and manage their expenses accordingly. This development is particularly advantageous for cloud providers such as Amazon, Google, and Microsoft.

  • Foundation models: The core of Generative AI lies in foundation models. These are large deep learning models that are pre-trained to generate specific types of content and can be adapted for various tasks. A foundation model is akin to a versatile Swiss Army knife, capable of serving multiple purposes. Once a foundation model is created, anyone can develop an application on top of it to leverage its content-generation capabilities. New models are constantly being built, with Hugging Face hosting thousands of them. Many of these models cater to more specific use cases and are open source, allowing for fine-tuning to suit particular applications.

  • Tooling and infrastructure: Businesses require two key elements to build applications based on foundation models. Firstly, they need a platform to store and access the foundation model. Secondly, they may require specialised MLOps (Machine Learning Operations) tools, technologies, and practices to adapt the foundation model and deploy it in their end-user applications. This may involve tasks like incorporating additional training data, labelling data, or building APIs for application interaction. Model hubs like Hugging Face and Amazon Web Services offer access to models and comprehensive MLOps capabilities, including expertise in fine-tuning foundation models with proprietary data and deploying them within applications.

  • Applications: While a single foundation model can perform a wide range of tasks, it is the applications built on top of it that enable the completion of specific tasks. These applications may be developed by new market entrants aiming to provide unique services, existing solution providers looking to enhance their offerings, or businesses seeking a competitive advantage in their industry. Examples of such applications include assisting customers with service issues or generating marketing emails.

2) The Tooling & Infra Stack

It’s hard for a seed fund to invest in chip makers, cloud platforms and even foundational models (although we have done one, Mistral) given the powerful incumbents and deep pockets required. 

So we - along with pretty much all other venture funds - are focused on the rapidly-evolving application lawyer and tooling stack. Here below we want to first talk where we think the opportunities are in the tooling stack.

At firstmintue, an enormous part of our portfolio is in developer tools (e.g N8N, Mondoo, Unleash, Xata and Engflow) and we were early backers of Lightning AI so we have been watching very closely the Generative AI tooling stack.

The stack itself is complex and rapidly developing. We see two slightly different tooling stacks for the different approaches that companies might want with their Generative AI models:  One is the API stack and the other is the training / fine tuning stack.

The training / fine tuning stack includes companies that act as model hubs (Replicate, Hugging Face); frameworks (PyTorch, Tensorflow);  experimentation tools (Weights and Biases); and Monitoring / Observability tools (Robust Intelligence, Gantry, Arthur, Arize, Why Labs). 

An increasingly prominent topic now is the “API stack” and all the tools that use in-context learning.

An increasingly prominent topic now though is the “API stack” and all the tools that use in-context learning to enhance these existing models and get the best results. These include companies that act as data pipelines (Unstructured, Airflow, Databricks); or vector databases (Pinecone, QdrantWeaviate); or frameworks (Langchain, Anarchy). 

But amidst all this (and you can draw complex market maps about all this, but it changes pretty much every month) what we really want to say is where we see the opportunities in this rapidly evolving landscape. Here are some:

  • We are interested in everything around in-context learning, from frameworks to vector databases which help enterprises deliver better performance from models. Companies like Weaviate, Qdrant, and MindsDB.

  • We are interested in tools democratising access to AI. The building, training, fine-tuning, deployment and management of this revolutionary technology can be better democratised. Exciting startups include Humanloop, and our portfolio company Klu.

  • We are interested in companies solving the problem of costly inference. LLMs are large and their inferences take time, which creates high processing needs and high costs for businesses. Companies innovating here are TitanML and Nebuly.

  • Due to the inherent lack of explainability in LLM, there is no easy way to monitor and govern models. Companies looking at this include Context (formerly and Enzai.

3) Application layer

There are a million ways to cut up the application layer and lots of different ways to think about it:

  • For enterprise, you can think of it by the business function that will be disrupted (the four core ones being marketing & sales, software engineering, R&D and customer support) and the general “specific use case” tools (e.g Jaspers for copywriting).

  • Also for enterprise, you can think of it in terms of sectors that will be disrupted (law, finance, health), thinking about the vertical and industry-specific applications (e.g Harvey for law). 

  • For consumers, you can think about it in terms of use cases like entertainment (Character AI) or productivity (Open AI).

  • You can think of it in terms of business models, such as vertical b2b SaaS or consumer with recurring revenue, splitting up the companies by who their end user ends up being and who pays.

  • You can think of it in terms of companies that are “full stack”, by which we mean companies that have built their own proprietary models (e.g Midjourney, Runway and Inflection) vs those just building on other people’s models. [Note: we generally prefer building on others, not full stack!]

There is no way to break this down in a way that is perfect, so prefer to think about it just in terms of “interesting themes” that we are looking at and want to invest in. 

Small Business and Enterprise: Maybe the biggest prize will be Generative AI tools for small businesses. AI-powered tools like Sameday, Truelark, Osome, and Durable can assist with specific functions such as answering calls, managing communication, handling back-office operations, and creating professional websites. Specific industries also have tailored AI tools, like Harvey and Spellbook for legal teams and Interior AI and Zuma for real estate professionals - some of these will target big companies and some small.

  • When people talk about Gen AI not being “defensible” they are often thinking about these business tools. Copywriting, sales enablement, customer support… there has been a cambrian explosion in the number of people doing this because it's suddenly not that hard to do and as a result it’s not clear which of them are building long term defensible businesses. As ever, we are looking ideally for some kind of network effect or data flywheel combined with a rapid GTM. Clearly, those that can integrate with big companies will be powerful - but with all the normal GTM problems of selling to companies with 12-month sales cycles.

Relationships: The ability to have a relationship with an AI is no longer science fiction. Many of us on the team feel we have a relationship with LuzIA, who we talk to every day on Whatsapp, and others have long-running conversations with AI chatbot products like Replika, Anima, and CharacterAI. While technology cannot fully replace human connection, AI chatbots can help alleviate feelings of loneliness. Also virtual girlfriends will be a big trend. Just check out CarynAI, a voice chatbot created by influencer Caryn Marjorie where users pay $1 per minute to talk to (she made $72,000 in her first week). This is going to be a monumental consumer trend over the next decade.

Democratising creativity:

Generative AI allows users to bring their imagination to life in various creative mediums. One company we are excited about at firstminute, for example, is a platform where people can create their own online comics, allowing people with great idea who maybe cannot draw to access a new medium. But there are hundreds of examples of this, with companies such as Lensa, Midjourney, Stable Diffusion, Sudowrite, Descript, RunwayML, Linum, QuickVid, Synthesia, Canva, PhotoRoom, Magician, Boomy, Riffusion, and Resemble all providing AI-generated content and tools for art, writing, video, design, and audio creation. This is huge.  

Education: There has always been a tension in edtech between quality and scale. But we believe now that Generative AI and AI more generally will enable individualised learning plans at scale, providing personalised experiences to each user. AI-powered language teachers like Speak, Quazel, and Lingostar can offer real-time conversations and feedback. Various AI-based apps and platforms, such as Photomath, Mathly, PeopleAI, Historical Figures, can teach new concepts and assist learners in different subjects. Maybe the age old tension in edtech of quality vs. scale can finally be solved? 

Gaming: Generative AI should be able to simplify game production, making it faster and cheaper, while also potentially allowing players to customise their own gameplay. AI tools like Scenario, Iliad, and Promethean are already being used to create game assets and virtual worlds. AI can also generate non-player characters (NPCs) through products like Inworld, Charisma, and Convai. Human artists will still play a vital role, but AI will enhance their efficiency and help release games more quickly at lower costs. It could lead to a more democratised way to create games as well as new fun opportunities within the games themselves!

Coaching and therapy: Like in education, there has also always been a tension between quality and scale in coaching and therapy. But maybe that can also be solved? AI chatbots are expected to be used more extensively for professional and clinical purposes in the coming years. They can provide affordable and convenient support for personal and professional development. Chatbots like Woebot and Wysa have shown effectiveness in treating mental health conditions and have received FDA designations in the US for conditions such as postpartum depression, chronic pain, and anxiety. As there is a shortage of therapists, chatbots might be a feasible solution for non-acute cases.

Science and R&D: Life sciences and chemical industries have begun using Generative AI foundation models in their R&D for what is known as generative design. Foundation models can generate candidate molecules, accelerating the process of developing new drugs and materials. Entos, a biotech pharmaceutical company, has paired Generative AI with automated synthetic development tools to design small-molecule therapeutics. But the same principles can be applied to the design of many other products, including larger-scale physical products and electrical circuits, among others. We have a portfolio company in stealth - and another which we have offered on - which is doing just this!

Writing code: Generative AI tools are useful for software development in four broad categories. First, they can draft code based on context via input code or natural language, helping developers code more quickly and with reduced friction while enabling automatic translations and no- and low-code tools. Second, such tools can automatically generate, prioritise, run, and review different code tests, accelerating testing and increasing coverage and effectiveness. Third, generative AI’s natural-language translation capabilities can optimise the integration and migration of legacy frameworks. Last, the tools can review code to identify defects and inefficiencies in computing. The result is more robust, effective code.

  • Democratisation of software? One question on our mind is what is next for Generative AI and the democratisation of software development. Are we moving to a world where non-technical or semi-technical people can build apps and tools using natural language?

Enterprise search. Language models have the potential to revolutionise one of the core functions of the internet: search. Instead of getting links to answer Google questions, why not the actual answer in natural language? Even more interesting though is enterprise-oriented applications for internal search. Most companies now use a number of communication apps and databases, such as Gmail, Slack, Drive, Asana, and more. Finding a single document, message, or metric across all these tools can be a challenge. Products like Glean allow teams to search across apps, while Vowel enables users to query records of their video meetings.

Healthcare: Healthcare has been pegged as a prime candidate for more AI applications — both to aid in clinical work and to lighten some of the more time-consuming administrative burdens that come around clinical care. One of our star portfolio companies is Nabla, the digital health startup out of Paris co-founded by AI entrepreneur Alexandre Lebrun, whis is one of the first to build a tool using GPT-3 to help physicians do their work — more specifically, their paperwork. Copilot, as Nabla’s new service is called, is a digital assistant for doctors to help transcribe and repurpose information from video conversations. As well as reducing the admin burden, we're also exciting applications in drug discovery eg Insilico medicine and computational biology via Generative AI powrered protein design.

Co-pilot for your job and… everything: Co-pilot for doctors is one big theme, and there will be lots of other business ones: co-pilot for lawyers, co-pilot for engineers, co-pilot for architects - all enhancing white-collar workers with the power of, as people often say, “the world’s smartest intern”. There is a cross-sector theme which is co-pilot for… everything! Already we are seeing lots of founders pitching companies that are aiming to build general-purpose assistants for ordinary people. What if you could plug in all your personal data (emails, whatsapp, photos, calendar) going back decades and create a bot to do certain tasks? What if that person could then be your perfect EA? What if they could reply to some basic emails for you? What if you could lend your personal bot out for 30m to people who want to chat to you about a specific topic with routine answers (e.g for a VC, a 30m call about top Series A investors in Europe could maybe be outsourced). 

4) How do we see defensibility? 

Defensibility of course a big question for VCs in this space. Applications in Generative AI often lack strong differentiation because they use similar models, which are trained on similar datasets with similar architectures - churn is reportedly high in many of the products that have taken off so far in verticals like copywriting and marketing.

But honestly we feel while this is a worry, too much is made of this by critics and doomers. As Alex Lebrun, the Nabla co-founder, likes to point out, when the C language came out in the early 1970s, people at the time said “oh now it’s so easy to build software, everything is just a thin layer on top of C”.

This, of course, sounds ridiculous now, because it’s what you do with these tools that counts. In the same way, LLMs today are simply a new type of infrastructure similar to C in the 1970s that open up new opportunities and new software use cases. It’s a completely novel type of architecture that is powerful but also hard to control (it’s not deterministic, hallucinates, the best models change all the time etc…) The teams that are the fastest to understand these new rules will build the best products - and it’s the products in the end that count. 

it’s what you do with these tools that are available to everyone that counts

Where the moat comes, we believe, might depend on the product. If the AI itself is the main differentiating factor in the end-product, tightly integrating the user-facing application with a proprietary model - or at least one that is fine tuned using hard-to-get data -  is likely to be the most successful approach here.

On the other hand, if AI is just one component among many features in a product, a more horizontal approach might prevail and it will be all about product - then it will be the GTM experts, the product killers who will be the winners and build the lasting businesses through just raw speed and execution. Luzia, which effectively puts Chat GPT and other Generative AI tools into Whatsapp, is just a nice UX on other people’s models. But it’s reportedly adding hundreds of thousands of users a day because people just want it: so maybe that’s enough? 

5) Open source (and Europe) for the win

First of all, Europe is doing a great job in the latest race to win the Generative AI race. Paris-based Hugging Face is now the world’s largest machine learning model hub. Open source LLM, which started life at LMU Munich before relocating to the UK, is one of the leaders in text-to-image generation. Synthesia has just raised $90m to dominate synthetic video generation. Mistral, a Paris-based company, is building foundational models to rival Open AI and has just raised a €105m seed round (the largest ever in Europe).

But there is a theme in some of the big companies: open source. Now there is a big debate about whether the future of Generative AI will be open or closed source, mirroring the early iPhone vs Android debate. Now we know what happened there: Android won the biggest market share, but the iPhone captured most of the value (the  iPhone got just 21% market share in terms of shipments, but 50% of revenue and 82% of profits).

Still, we are believers in open source in general (check our our big portfolio with N8N and Element) but also for open source as the future for Generative AI. This is partly because of the speed of iteration and developer adoption. But also because there is so much focus within enterprise is about running on-premises models so that sensitive data does not leave the organisation. This is important to pretty much all big companies - but particularly for areas of government, health, defence and security companies. 

This is one of the reasons we invested in Mistral, which raised €105m to build a foundational model to rival Open AI and is taking a more open approach to model development. The company plans to release models with a “permissive open-source-software licence, that will be largely above the competition in that category”. This will help them “create a developer community around our trademark”, according to their memo. They are also focused on explainability and data privacy.

6) What are we looking for? 

We see dozens of Generative AI companies every week and have have the honour of meeting amazing founders building here. What are we looking for that makes us want to invest though?

  1. Deeply technical teams with a commercial mindset that can launch quickly, iterate, and learn fast

  2. Teams that can sell quick and be aggressive in marketing - speed is everything here

  3. Products that can get some network effects or data moats: particularly those focused on specialised knowledge-work such as; engineering, law and medicine.

  4. Founders with deep domain knowledge that clearly understand the challenges manipulating unstructured data within their fields and aim to use LLMs to overcome these

  5. Products using more than a single model, where “time to delight” really fast.

  6. SaaS or consumer subscription business models in the application layer. 

  7. For anything more deeptech (new models, computational biology etc..), world class research backgrounds. 

Finally: Why are we investing now? It’s about people… 

We were at a conference the other day and someone asked: why now? Why invest in Generative AI now when there is so much hype and deals are getting bid up by rival VC firms (many with very deep pockets)?

The first answer is that trends are long usually, and we believe that this is going to be a long one. The second is that as a venture firm, we follow great people and at the moment some of the best people - those with a sense of insurgency, product obsession and charisma - are building in this space. 

Our goal as a firm is to find and partner with this rare breed of people who can build $1bn+ businesses.

If that’s you, and you have read this far (!!), do reach out to us at, and