Generative AI continues to drive headlines, buzz, and massive amounts of capital. Most recently, Sam Altman is in talks to raise trillions (not a typo) to “revolutionize” the chips business. Meanwhile, the CEO of Amazon’s AWS business unit, Adam Selipsky, described the energy around generative AI as “dramatically overhyped,” and compared it to the dotcom bubble.
The AI landscape, unlike categories like sales tech, or marketing tech, spans a dramatically diverse landscape of business models. From foundation model platforms to GPT wrappers to core applications. Questions have come up about the long-term margins and economic potential for many of these businesses. Given the space is changing so rapidly, articulating the steady state potential of these technologies can be difficult.
Even Adam Selipsky, while believing the space is overhyped, still described the technology as valuable. But the excitement around the specific companies brings with it a cost:
“Many companies and organizations are struggling to understand, ‘Out of these hundred pilots or proofs-of-concept that I have going on, which ones do I take into production?’ And they're starting to see that it can be very expensive once they go into production.”
As we consider the potential future of the space, there are some core questions that may shape some of these ongoing dynamics:
- Future Dominant Architecture: In our interview with Aidan Gomez, the CEO of Cohere, he mentioned that, if in 5 years, the dominant AI architecture is still transformers, he would be very disappointed. So what will take its place? And how will that change the equation for AI going forward?
- Durability of Specific Models: A number of people have continuously pointed out the capital intensity of developing specific models (including us.) And competition continues to be fierce. Open and closed source models alike are vying for the top slot on various leaderboards, including Abacus AI, Mistral, Anthropic, OpenAI, and Google’s Gemini. Will prior models still find use cases that will allow companies to amortize those costs over time? Or are they all a flash in the pan that will ultimately become a race to the bottom?
- Availability of Training Data: We’ve written before about how knowledge communities, such as StackOverflow, have continued to see declining engagement from power users because LLM output is so good. But models rely on user-generated content from platforms like StackOverflow and Reddit to train their models. What happens when we push the limits of user-generated data? Does the focus shift to synthetic data? Or do we increasingly move away from language models that are so large?
These are just a few of the questions that will increasingly define the generative AI category. More critically, the question will become how each company playing in this space will respond to the ever changing landscape?
