The AI Strategist - 26.5.23
Moativation
In the news - Google has no moat - this much-referenced recent article suggests that even foundation model providers have no moat. Making the point that Open-source models are faster, more customisable, more private, and pound-for-pound more capable.
Why this matters - the controversial article certainly raises some interesting points. If the sector is so unstable that even the foundation model providers don’t have a moat, should startups and corporates pursue an Open-source strategy from the start? Or at lease a hybrid? Closed-API only brings long-term risk.
When I attended a Generative AI event (London) back in February a CEO of Generative AI business ($1bn valuation) said on stage stated that Open-source is ‘2-3 years’ behind proprietary data science research. He believed this was healthy and necessary. When it comes to LLMs this is spectacularly not the case. This article does a good job of explaining why. As Bob Dylan once said, ‘the times they are a-changin.'
When it comes to moats, all businesses will need to engage with the Open-source community to better understand potential defensibility. Innovators will need to take a long-term view on how the Generative AI tech stack will evolve and let this determine what short-term build/buy decisions they make. Taking a strategic view on moats and value creation will help organisations navigate a steady course through a sector which right now is developing at breakneck speed.
Right now there is a movement toward smaller open source models which perform well for specific tasks (ubiquity and low cost). The pace of advances in open-source LLMs is much faster than in the closed ecosystem because different teams can build on top of each other’s work. This puts even more emphasis on gathering moderate size datasets as a competitive moat.
Speaking of data, Stackoverflow and Reddit are starting to charge for API access. They realise how valuable specific dialogue datasets are for context and understanding. Organisations need to start organising and capturing conversational internal datasets.
2. What type of Generative AI company are you?
In the news - another way to think about moats and long-term growth is to consider what type of business you want to be.
This great article breaks down the three types of Generative AI business:
Model builders (e.g. OpenAI, StabilityAI, Co:here)
Significant funding
Access to capital, compute, data, and specialised talent
Have an API or provide models as a service
They rarely build products on their own models
They are threatened by the pace of innovation (competitors and open source)
Business strategy is subscriptions, usage-based billing, partnerships with large orgs
Model consumers (Jasper AI, Lensa)
Minimal barriers to entry
Consume models from open source or model builders
Typically have low defensibility and high churn
Low barriers to entry mean copy-cat services
Risk of building core business on an API (exit strategy)
Business strategy is subscribers, ad-supported, enterprise deals
Vertical model players
Have sufficient funding (compute, talent, data) to train their own specialised models
Then build a dedicated product on top of their own models
Disadvantages:
Foundation models are not as powerful as model builders
Speed of development is slower than model consumers
Benefit - these companies can operate in multiple modalities without being limited to existing models on the market
Business strategy is subscribers, ad-supported, enterprise deals (same as model consumers)
Why this matters - moats can be quite a paper/theoretical exercise. This framework, in my mind, comes before any moat-based thinking. It is a bit more existential. There is the internal view - your company culture, strengths, history of successful innovation and funding level etc. This brings an external lens to how Generative AI will evolve and where you want to position your business. Very unlikely you want to become a model builder. Most startups will be model consumers. Vertical model players trade-off more defensibility with more risk.
In terms of strategic implications:
Strategy 1 - Model consumers (closed API) - fast-paced, low defensibility.
Strategy 2 - Model consumers (open-source) - fast-paced, low defensibility, element of control.
Strategy 3 - Model consumers (both) - fast-paced, medium defensibility, a foundation model hedge. No intention to train own models in future.
Strategy 4 - Vertical Model Players (typically open-source) - slower, high defensibility, more expensive. Start with open source strategy with intention to train own models. If you are serious about this strategy, this read is a good starting point.
Strategy 5 - Pursue model consumer strategy with an open mind about becoming vertical model player. I.e. have defensive moats in place such as data gathering.
3. AI is eating the world
In the news - back in March Cohere put out this great blog about AI eating the world. It builds on thinking from Marc Andreessen. Cohere do a good job of talking through the Generative AI tech stack. The article concludes with a diagram outlining where ‘pockets of competitive advantage’ are available for AI startups.
Why this matters - it is now quite clear that Generative AI is going to lead to major disruption of business models. In this section I am going to talk few a recent developments to explain what ‘eating the world’ could mean in practice.
Implication 1 - Human-in-the-loop - Generative AI will help businesses/consumers automate tasks that are repetitive and mundane. Organisations will need to truly redesign processes around human-in-the-loop. This is a well-known phrase in designing AI systems but something that many organisations struggle to implement well.
Implication 2 - LLM Simulation - much ML up to now has been about inserting predictions into human processes. This was well covered in prediction machines. Sounded great but humans didn’t understand or trust the predictions. This led to the emergence of a category called decision intelligence. We can think of this as rewiring how strategic decisions are made and audited within organisations - using data and AI. Generative AI simulations will disrupt this yet again and are applicable to a wide array of use cases. It is highly relevant, for example, for newspapers monitoring spread of online discussions and when they might go viral. See diffusion of innovation, and Signal AI for further reading.
Implication 3 - Automation - the economic potential of Auto-GPT is significant. I.e. helping humans do things faster, more accurately, running parallel tasks, in different systems. Organisations need to think about the powerful potential of combining Auto-GPT with tested and proven traditional ML techniques.
Implication 4 - AI for everyone - Companies like Zapier are building AI automation tools to enable users to customise software to their needs. This is the next evolution of what C3AI have been doing in B2B Enterprise AI with their mission of no code AI using their platform. Interestingly Zapier productise this under 4 roles, 4 industries and 3 company sizes. They are pitching themselves as the glue between AI apps.
Key takeaway - it is expected that Generative models will be used to infuse generalised knowledge into other ‘traditional’ smaller ML models to solve more specific tasks. Without this they lack context and understanding. They can be thought of as components of any system. Businesses that win in the long run are likely to take applied and research data science seriously. This will allow them to figure out how these multiple models can work together in production safely and ethically.