We’ve all heard about Generative AI in the recent months and it’s for good reason - the incredibly buzzy topic seemingly hit Hype Beast mode last October following a slew of (suspiciously co-ordinated) news that brought it into the open with a big bang. Stability AI announced their $101m raise, Jasper announced their $125m raised, both achieving unicorn status. Multiple VC funds published thought articles and followed up with helpful market maps. When ChatGPT was released late November, everybody was gobsmacked, tech-twitter blew up and even mainstream BBC reported on it - now it’s too full to process more queries most of the time!
But even apart from these events, generative AI has been gathering momentum at shocking speeds in the last 6 months. Large foundational AI models have existed for a couple decades but recent years have seen impressive releases such as GPT-3 in 2020 and DALL-E in 2021 that demonstrate potential meaningful application. In just 2022, multiple AI models for text, image and various mediums were released (GPT-NeoX, Midjourney, Stability Diffusion, Whisper just to name a few) including even open source ones, which has really opened the floodgate for innovation and entrepreneurship to build within this space. Since then, we have seen constant releases of new versions that show leapfrog improvements in output (e.g. Midjourney 4) and there are more on the horizon (GPT-4 rumoured for 2023). This foundational technology is accelerating at a breakneck pace, signalling that it may finally be ripe for mainstream application. This year will be telling of the first pioneers or winners in this space.
With all the hype aside, we think it’s important for us to form and share our own views. There are plenty of really good articles out there already that explain what generative AI is and the Why Now that we can’t possibly outdo, so we’ll focus more on our thesis and the framework we’re thinking about it in. The TLDR is that we’re incredibly excited and actively looking to invest in the space, so if you’re an early stage generative AI founder, we want to hear from you! Share your business with us here.
Generative AI is a platform play
The generative AI space can be roughly broken down into three areas - (1) the underlying large AI models that form the foundation of the technology, (2) the interface and workflow layer built on top of models that enable end-users to use the AI easily and (3) all the tools or platforms used by developers to access and utilise the technology, otherwise known as the picks and shovels.
We see the first area, the large underlying AI models, as the starting platform. Much in the same way that mobiles, social media or web3 are platforms that have enabled a wider ecosystem of ‘applications’ to thrive on top. The technology behind generative AI is the foundation and enabler of many other businesses that can be built on top for many use cases. It is infrastructure. For example, GPT3’s model can be applied to write copy as in Jasper AI or Replier AI. The models are core to this space and the businesses that build them are like engines as without them, nothing would be powered. By its nature, these businesses require vast amounts of data, processing capacity and human feedback as well as financial capital to build such technology - after all, they’re recreating ‘intelligence’ - so it’s no surprise that Big Tech with their deep pockets and omnipresence are forerunners in this space. If they were the only ones in the race, it would strengthen their position as all-knowing behemoths even more (let’s not think about that yet…). However, there have also been emerging third players, such as OpenAI and Stability AI, competing with their own models and innovating on even business model (e.g. being Open Source) all in the hopes of becoming the next big platform.
While we believe these businesses will be valuable, we think this layer will be fast commoditised as quality converges in time and models will be forced to compete in price and usability. From our lens as seed investors in network effects, it’s also not an area for us to invest in given its high capital demand and long R&D times - they’d need a huge amount of resources to train anything to compete with the likes of OpenAI. Nonetheless, we’re incredibly grateful that there were investors who backed them and the research has come to fruition…to unlock the second and third area of generative AI.
We’re most excited about the second area, the interface and workflow layer or ‘applications’ built on top of the large AI models. This is the space where we see most potential to differentiate and to use generative AI technology as a springboard (or trojan horse) to form your own platform and network effects. Just like when mobile started becoming mainstream, it wasn’t more hardware device brands that sprung up, it was the apps that expanded its functions and use cases. Now that the foundational layer is strong enough now, we expect to see an explosion of applications in all sorts of verticals built on top and it’s in this space that the first winners will start taking shape. Early examples include JasperAI in copy writing, Paris-based PhotoRoom in photo editing and imagery, and Github Copilot in coding. They won’t be the only ones either - it’s a huge area of opportunity because the possibilities are endless (more on that below!)
The third area, the picks and shovels, may be a byproduct of new technology but are nonetheless just as powerful, valuable and crucial to its adoption as they also act as enablers. There are still many barriers to adoption for generative AI, some technical, such as computational infrastructure, and others that are non-technical, such as skills shortages. Businesses that aim to reduce such barriers will ride this wave as much as the applications will, if not more, as they become almost infrastructure-like. Think about Github, which is such a central pillar for software developers in many ways, it’s not just where they store versions but also where they collaborate, network and get feedback on their work. Same idea with Stack Overflow. One could say Web2 development innovated faster and better because of such supporting tools, and the same will definitely be said about the ones for generative AI. Early examples include Weights and Biases and Hugging Face. We also see this area as opportunity and as with most picks and shovels, with plenty of potential to foster network effects amongst its users!
The application possibilities are endless, but…
We believe the potential impact for generative AI is huge, not just because of its intrinsic value in creating novel and useful things, but also because the application possibilities are endless. It can be applied in almost every vertical, in every task that we do. But not every application will make sense or be as value-adding as others. We’ve seen this before - the uncontrolled explosion of single-feature mobile apps for almost everything you can imagine (stopwatch, birthday cards, probably even fart noises) and then consolidation into a handful of platform apps. So what’s the winning formula for generative AI applications? We don’t have the answer (if I did, I’d go and do it myself!) but we have thought of a framework with the different levers to pull to create value:
1. Different verticals
Generative AI can be applied to different mediums as we’ve already seen, each of which may have use cases in different verticals. E.g. Imagery for marketing vs gaming vs education. So far, marketing, sales and coding have seen prevalent applications. The common thread here may be that verticals that are not regulated or have more flexibility for imprecision (at least initially) may be faster to adopt. On the other hand, being the first to crack a regulated or ‘higher-bar’ vertical, such as medical or legal, could be a competitive advantage. In some verticals though, we may still be some time away from models specific enough to be useful. Meta’s Galactica model released in Nov 2022 purported to assist scientists with writing scientific compositions. However, it was taken down after just 3 days when the community dismissed the tool, criticising that making grammatical sense was insufficient since the model’s output did not provide any scientific substance nor grounding to its output, thereby creating risks of misleading research. Clearly, we’re not there yet but we are confident that it’s a matter of time that we will be.
2. Depth of the AI
The ‘depth’ that generative AI goes to assist users is another variable to differentiate applications and could vary depending on the vertical, proposition and even maturity of the underlying model. By depth, we mean the extent that the AI assists with a task - complete autonomy would represent the deepest while first drafts or brainstorming would be on the shallower end. Similar to the above, deeper penetration of AI in some verticals or use cases may make sense because of lower barriers to adoption, flexibility for imprecision or even sheer volume of output required. Examples where near-autonomous generative AI could make sense may be in customer service or ad copy. Longer form content, such as blog articles, may be suited for drafting first, at least initially. This lever may also be the way to penetrate the more challenging verticals mentioned above as a shallower proposition could lower the barriers to entry. For example, a tool that creates first drafts of legal terms rather than final contracts may face a lower bar and hence, encourage adoption. Of course, the depth of AI for each vertical or task will constantly change with time and so we think it will be critical for startups in this space to keep a pulse check and evolve with their industry.
3. Accessibility, Embeddedness and Tooling
The last lever is the extent that generative AI is accessible, embedded into workflows and complemented with tooling. Another way to put this is the UX layer of generative AI - how easy is it for the user to get the value of generative AI in their role? An example of the sheer impact of this layer is the launch of ChatGPT. The UX innovation of a conversational chat interface accelerated adoption of the technology by leaps and bounds, making even mainstream headline news, despite that ChatGPT is loosely based off GPT3, which was released 2 years ago. We see this UX layer as a critical lever to building out a platform on top of generative AI because it’s one of the few ways to differentiate from competitors and build defensibility, especially if they also rely on the same underlying AI models (which is likely). A singular generative AI tool would likely be at risk of becoming a ‘feature’ of an incumbent platform, which has the clear upper hand of an existing user base. Examples of this are already emerging with Notion introducing an AI assistant to brainstorm, write first drafts, summarise and edit articles. Although still in Alpha, Notion is extremely well positioned to lead this vertical as (a) its seamlessly embedded in their current product, (b) it has all the value of its existing tooling along with it and (c) last but definitely not least, they have an established user base. Could GoogleDocs or Microsoft word do the same? It’s likely. Microsoft is already paving its way to embed the technology into its ecosystem with an additional $10b investment into OpenAI, after investing $3b since 2019. So, why would users switch to a third party generative AI text editor? The bar is very high - the AI would need a whole platform with tooling and workflows attached and they would need to surpass Notion or Microsoft by 10x. Such players should rightly feel threatened in this scenario. Despite this, there are still many verticals and use cases that remain untapped so this won’t always be the case. While singular generative AI tools may not be so defensible in the long term, they are still a great wedge when embedded into existing platforms or workflows and can build out their defensibility by adding tooling and becoming its own OS or platform in time.
This takes us to our last point, which isn’t a lever per se but is a crucial factor to capturing this opportunity - speed. We are at very special point on this curve, where the underlying technology is developed enough to start showing meaningful application and is becoming accessible to mainstream developers and entrepreneurs to build on top of. On the other side of this coin, the potential risks are also quickly surfacing and attracting attention, at the centre of which is the factual (in)accuracy of output and the consequences on users. There is much work to be done to harness and adjust our society to such technology, which will likely involve some degree of regulation, e.g. the EU is already planning to regulate text and image-generating AI models. It’s a true case of building the plane and the runway while it’s flying. For the entrepreneurs aspiring to be pilots though, the starting pistol has gone off and the race has started. Speed and how businesses go to market will be huge competitive advantages in this evolving space, we expect many applications of generative AI and surrounding support tools and platforms to emerge in the near future. Undoubtedly, some of them will be the first winners!
You can probably tell that we’re very excited about Generative AI and we hope sharing our thoughts and framework on the topic is helping to starting or steering your business. We are actively looking for startups tackling the most value-adding use cases and verticals with a killer go to market strategy and of course, a relentless speed of execution. If you’re a Founder of an early stage Generative AI business, we want to hear from you! Share your business with us here.
Bonus for those who stuck with me until this point: So, the real question you want to know is did I really write all of this? I promise, I’ve written every word here but as I constantly suffer from writer’s block and procrastination, I honestly do believe that this would’ve been faster with some AI help. Perhaps next time. No, definitely next time!