The Supernova of Generative AI: What’s Next?

Written by

Kes Sampanthar Managing Director | Abhishek Gupta Senior Responsible AI Leader & Expert at BCG, Augmented Collective Intelligence Fellow at BCG Henderson Institute |

May 26, 2023 · 8-minute read

Generative AI has overtaken dinner conversations, enrapturing even those who previously never discussed technology, let alone AI. Organizations are convening special projects and working groups to experiment with this technology to find ways to harness the tremendous business potential of Generative AI. But, what comes next?

For those who want to skate to where the puck is going and peek into underlying economic forces, we propose a Twin-Flywheel Framework that highlights the technological, psychological, and market dynamics at play. And some actions you can take to jumpstart and crank those flywheels for success.

Stable Diffusion’s “The AI Flywheel”

The AI Flywheel

Tried and true model:

Algorithms of mass engagement start with interesting and unique content (and capabilities) that attract people. These people then generate data through their engagement with the system, providing valuable feedback. These enhanced systems generated more convincing content and capabilities that attracts more people. Rinse and repeat.

In the spotlight:

The current supernova is also fueled in part by the excessive media adulation that these systems have received. This has accelerated the pace at which people are flocking to these systems. Common complaints from people now include encountering a “High server load, please come back later” message which only heightens anticipation and FOMO feeding into the above cycle, a veritable desire to be a part of the zeitgeist. Social media is an attractive accelerant in this phenomenon since early adopters of such technologies like to be the first to share new innovations.

The Two-Sided Platform Flywheel

Platform mindset:

The current crop of Generative AI systems are exposed through platform APIs (with Stable Diffusion being additionally available as an open-source trained model). This brings into play the familiar platform dynamics of serving a two-sided marketplace. On the one hand, the platform needs to attract users, and on the other hand, it needs to attract developers who build useful services using that platform.

Twin forces:

Both users and developers contribute to spinning the flywheel. Users are often responsible for virality of the outputs from these systems, particularly when they have a massive online following. This drives many more users to come and engage with the services and the platform itself. This spills over to the side of developers who are now more incentivized to build their services on this platform given the large user base. At the same time, more developers can lead to a diversity of services on the platforms attracting more users. As developers invest in building services on the platform, they generate more technical documentation, tutorials, tips and tricks that progressively lowers the barriers for other developers to come and build services on it.

The First-Mover Advantage

Defensible moats:

Platforms who are able to jumpstart the flywheels, spin them, and crank them up have an accelerated first-mover advantage. They can garner more users and media attention which helps with both the AI and Platform flywheels. The data and algorithmic advantages that this creates lead to a strong moat. This elevates the cost of entry for other platform competitors into the market. This is partly the reason for VCs investing deeply into Generative AI platforms, notably OpenAI floating the idea of raising $29b through an IPO.

Adoption curves:

While these flywheels exist naturally on the AI and Platform side of things, they need a jolt to get them going. Initially, the platforms focus on attracting the early adopters, both on the developers and users’ side. These are people who love experimenting with new technology on the market. Crossing the chasm to bring in the majority, where the bulk of the momentum lies for both these flywheels, requires understanding how to accelerate the journey along this adoption curve. With these systems being in the spotlight, the profiles of early adopters might also be shifting, something that hasn’t happened many times before.

Kickstarter effect:

Given the strength of the first-mover advantage, the race for pole position has accelerated, building upon trends in the market. We have seen a rise in early adopters and a leftward shift in the traditional adoption curves. Crowdfunding platforms have risen over the last decade, with more people accessing innovative products and services in conjunction, the lean startup movement is leading companies to launch ‘minimum viable products’ (MVPs) earlier into the marketplace to experiment and learn rapidly. These dual forces of crowdfunding and lean startup have led to a shortened t-minus to launch.

Stable Diffusion’s “Defensible Moats”
Stable Diffusion’s “Jumpstarting the Flywheel”

Jumpstarting the Flywheels

Race to utility:

The key question the platform needs to answer is how to attract users and developers? It begins with offering utility (duh!), but helping them reach it quickly! Platforms like ChatGPT from Open AI have succeeded even though similar (attenuated) capabilities existed in the past. This is because the outputs from ChatGPT provide direct and instant value. It doesn’t take too long to get a text output back from ChatGPT. And what it replies back with is usually immediately useful (though it can be wrong, caveat emptor).

Speeding up experimentation:

Compare this to Stable Diffusion where it takes a while (in internet time) to get back the image output and sometimes it isn’t what we wanted. And this is very important. Altering the text input prompt based on the output we got back to go for another round can become a frustrating experience. This is the differentiator as to why ChatGPT caught the wider public’s imagination as opposed to Stable Diffusion (which came earlier and frankly, generates more interesting outputs!).

The diffusion model:

No, not the one used in Stable Diffusion, but Rogers’ Five Factors that explain why adoption happens. Platforms that succeed need to show relative advantage. ChatGPT showed that with its utility, and getting to it quickly (e.g., through cohesiveness and completeness of responses, though not in all cases). The other factor from this framework that speeds up achieving utility is trialability. ChatGPT with its free offering made it a no-cost effort to continuously play around with it, fueling the diversity of things that users did with it.

Spinning the Flywheels

Make it intuitive:

Once the flywheels have started spinning, they need help to keep going. Intuitiveness of the interface to the platforms and services is critical for building this momentum. It makes it easier for new users to experiment with the systems. Given the potentially accelerated adoption curve and evolving early adopter profiles, something that is easy to use helps edge out competing platforms. Drawing from Rogers’ Five Factors, reducing complexity increases accessibility for users who want to see what the hype around Generative AI is all about. This is reinforced through compatibility with interface styles that are familiar to users (e.g., a chatbot) and trialability, i.e., you won’t break the system when you mess around with it.

Learning from peers:

Humans are social learners, we are wired to create by imitating those around us. Midjourney leaned in on this tendency by deploying their system as a chatbot in a Discord channel. Users can observe how others are prompting the system and see the results that they are generating. (This also makes the wait times a bit more tolerable.) How can users share their creations? The Discord interface makes it easy and integrated, lowering the barrier to observation and learning from peers. It sparks more ideas in the user base making them more likely to come back and experiment with the system. The underlying dynamic is being replicated by other platforms like ChatGPT that now show the history of chats that the user engaged in, allowing them to go back and compare results and figure out how they can do better.

Stable Diffusion’s “Interface to the Platforms”
Stable Diffusion’s “Harnessing the Supernova”

Cranking Them Up

Harnessing the supernova:

Rarely do we have such a mania and confluence of interest and energy in a field, especially one as technical as AI, and in particular Generative AI. Platforms that have the ability to rapidly gather and iterate on feedback will be the ones who will corner the market. Midjourney does a great job with this: David Holz has mentioned that his team observes the interactions and usage patterns of the community on Discord which gave them insights that led to their very successful v4 release. This extends the Observability factor towards the platforms themselves as a mechanism of improvement.

The Twin Flywheel Framework sheds light on where we can expect the industry to go and what organizations who want economies of flywheel-spin should do to get ahead of the puck. Reach out to us and let us know how you are experimenting with Generative AI and how you imagine the ecosystem is going to evolve.

All images in this post were made using Stable Diffusion


"*" indicates required fields


Tell us how we can partner together.