Obituaries of Generative AI or artificial intelligence in general have started trickling in across the media. This article in Fast Company predicts the onset of AI winter for the reasons that are typical in the early phase of any technology transition.
Those of us who saw the boom-and-bust or hype-and-disillusionment cycle of the Internet would remember similar articles predicting the demise of the world-wide-web. Here is one such prediction by none other than Robert Metcalfe! There are always a set of subtle but important levers that one should understand in order to correctly predict the future of any technology that is fast evolving. Do I see onset of AI winter? No! Will majority of the early AI players will cease to exist or see their dominance diminish in future? Yes!
While it took almost five-to-six years for the Internet hype to bust (1995 – 2001), it will take maybe one or two years for the Generative AI hype. Why? The hype around the Internet was triggered by an enabling product – the browser. The hype around Generative AI was triggered by an end product – chatbot (ChatGPT). Just like WWW is much more than what appears in your browser, AI (generative or otherwise) is much more than just a chatbot.
What then does the future hold for AI? Here are some of my predictions.
LLMs (Large Language Models) will be commoditized as their differentiation becomes as thin as a razor blade. The levers that control the differentiation are the training data and the model architecture. If OpenAI, Google, Meta, Amazon, and other big-tech players are using the same or similar public data and the same, but slightly tweaked, transformer architecture then how would the models be much different? Unless someone comes up with a novel architecture that can significantly improve the model performance and output with a drastic cost reduction, these tech-players will hit a wall and will not be able to recoup their investments. This is the primary reason why Sam Alton and others want to limit the competition through legislation and fearmongering.
Content creators will hide their content behind paywalls which will further restrict data for model training. The companies will try to overcome that by generating synthetic data but that will lead to more bias with risks of lowering the output quality of the models. If one LLM is trained on the data curated by another LLM (assuming data use permissions), they will likely become an echo chamber that further aggravates the bias problem.
Custom, private, and personalized models will rule the (AI) world. I mentioned this in my previous blog, but that was more targeted towards the enterprise. The same will hold true for consumers. Imagine a platform where you can spin up one or more language or other types of AI agents that are trained on your data and stay private, as applicable. One agent to read and summarize all content based on your interests, another one to watch and predict when to buy or sell an equity based on your investment portfolio and yet another one that not only alerts you on various job postings but also customizes your resume and applies on your behalf. I can think of multitude of use cases, both in enterprise and consumer space, where the early movers will rake in billions, if not more. Edge models that can reside on your mobile device to truly become a personal & private virtual assistant will have an edge (pun intended) over cloud-based models.
Domain specific AI models and delivery platforms will be another big piece of the pie. This is another area that gets me excited because the use cases will spout like mushrooms after a rain. Many entities that have the data or rights of use to data may create walled gardens while some may choose to make their models public. Watch this space closely as the monetization of domain specific AI will be a trillion-dollar play.
One subtle paradigm shift that no one, or maybe very few, are paying attention to is the shift from deterministic rule-based computing to predictions. This is nothing less than a tectonic shift in how we think about software and computing. More on this in another blog.