Last week, we hosted Max Rimpel, Partner at General Catalyst, for an evening of dinner, drinks, and lively debate at Tribe House. (This is part Tribe’s series on investing in ML – let us know if you'd like an invite).
Our conversation ranged from licensing problems to digital avatars. But (no surprise here) a lot of it focused around the potential of generative AI and how it’s changing the landscape. Below are some key takeaways from the discussion:
Building Moats in Generative AI
There’s been a lot of talk recently about moats and generative AI. Essentially, how do companies carve out a competitive advantage when they’re building a product or a company on an existing model? And there’s been even more talk about whether moats exist in the first place or they’re just a way for investors to feel more secure in their investments.
The discussion landed on: moats are real, but there’s no such thing as an infinite moat. The key is to constantly adapt to stay relevant. Even for FAANG companies with enormous scale there’s inherent risk in being a one-product company. Companies need to diversify and adapt.
Uncovering Long-term Value in AI
We’re at an inflection point where most products will have some form of AI in them in the next five to ten years. But the reality is: the lives of consumers won’t be fundamentally disrupted by what we’ve seen in the last six months – despite the hype.
Max also predicted that long-term value is going to accrue in verticals that are less explored right now. For example, the video and 3-D animation industry has huge potential. Foundational models have the potential to become part of everything. In the same way databases became a standard for building applications, AI has the potential to be the "database" for the application layer that will be produced in the coming years.
Spotting “Snake Oil” in AI
For Max, the key to identifying “snake oil” in the AI industry is to look at the entrepreneur’s experience and motivations. They should be passionate and have a deep understanding of the problem they are trying to solve. AI should be an enabling technology and not the focus of a product. In many ways, the best products will be the ones where the user doesn’t have to think about the AI behind the product. It’s just there, helping them achieve their goal.
The Foundation of AI
Generative models are poised to become the foundation of most ML and infrastructure. Generative models and existing classification models will be used as components in other systems, allowing them to infuse generalized knowledge into smaller models. They will be used to solve more specific tasks by providing the underlying knowledge and context that the smaller models may lack. This could have a big impact on long-term edge cases in areas like autonomous driving and document processing.
Predictions for the future
Predictions for the future were wide ranging, but a few clear possibilities emerged:
- A few big companies have the means and skills to build foundational models for AI. They will generate their own data sets and control the sphere. Average companies will not have the resources required to build their own models, so they’ll license a menu of existing models.
- An industry emerges that’s focused on fine tuning these existing models for companies. They could charge for personalization by industry, users, or group.
Navigating Data & Licensing in AI
Data and licensing is currently a gray area for a lot of verticals. For example, one reason the music industry is behind compared to visual art generation is that musicians have stronger legal protections on IP.
Recommendations and AI-generated content
Will generative models like ChatGPT flood the internet with valuable content? Or is it just junk and noise? If content continues to proliferate, two things are clear: the future of recommendations and content discovery will have to change in order to handle the flood of AI-generated content. And the cost of running these models may require the development of A’Is that filter out meaningful content from the noise.
All images created using Midjourney
.png)
.jpeg)
.png)
%2520(1).jpeg)
.png)
.png)
.png)
.png)
.png)
.png)
.png)
.png)
.png)
