Imagining a new incarnation for AI: The unprompted expert
True knowledge transfer rarely relies on a learner asking the right questions
Generative AI has done more than maybe any other technology to help solve the ‘blank page’ or ‘cold start’ problem when it comes to content production and idea generation. It’s the first time in history that we’ve developed a technology that can capably act in a co-creator role with human beings, as opposed to just an enhancement or accelerant to human origination. But it still operates primarily on a pull model – you have to prompt or query it in most of the dominant user experience paradigms we have for generative AI today.
On a recent episode of the always excellent No Priors podcast, Harvey CEO and co-founder Winston Weinberg articulated an under-appreciated point about how we learn from experts: We gain the most from smart people around us by just spending time literally with them around us, rather than by proactively going to them and asking them specific questions in isolated bursts, exclusively inquiring along a self-directed vector.
There are a number of limitations that arise naturally out of a dynamic where you have a non-expert or non-specialist person trying to learn from someone more knowledgeable in a specific area. One of those is commonly referred to as ‘not knowing what you don’t know’: Basically, you’re not going to necessarily ask the right questions or seek out the most relevant information from a source if you’re not already at least somewhat familiar with the lay of the land. Somewhat related is the problem of contextual gaps, or not having a fulsome picture of everything that surrounds a given answer to a specific question and therefore mis- or under interpreting the response.
Given the era in which generative AI emerged, and the paradigms that it was best set up to compete with, a query-based interface made the most sense for general use tools, including ChatGPT. It’s still also the best way to use AI for most people most of the time, for most tasks. But in the category of general enrichment and knowledge acquisition – a category I’d argue is growing – I think other models stand to offer considerable benefit vs. a request-based chat interface where most of the onus is on the person asking questions to know what’s important to focus on.
Some more recent AI-based products approximate what I think of as a more autonomous knowledge companion: OpenAI’s voice mode is good at prompting additional conversation through clarifying questions it appends to the end of its answers, for instance. Google also has its much-vaunted NotebookLM, which can generate topic-specific, naturalistic podcasts on-demand from whatever materials you choose to feed into it.
NotebookLM’s subject matter boundary (which is determined by the materials you put into it) is a feature not a bug for its intended use cases, which range from study guides to research and writing assistance. Yet it also makes it more prescribed than I think is strictly useful in a situation where your goal is to improve your general familiarity on a particular topic, in a way that is less about a specific directed outcome and more about building context.
I think a rich, under-explored product model for delivery of generative AI would be a kind of library of experts, which can perhaps be queryable much like ChatGPT or Claude are today, but with a very different primary delivery method that’s actually much more like traditional terrestrial radio or broadcast television.
Key to this working well would be that the AI expert is essentially self-directed in its knowledge sharing – perhaps responsive to prompts or queries in terms of very generally orienting its ‘thinking' out loud, but mostly meandering down a path guided by its own ‘interests’ as they pertain to its discipline. The experience should be more or less equivalent to asking, say, a professor of contemporary poetics what they’re interested in or paying attention to right now in their field, and then maybe asking them clarifying questions when they venture into jargon, or seeking expansion in case they hit on something you find particularly fascinating.
As I mentioned earlier, some tools out there now do resemble this very closely, but I don’t think it has received enough attention as a ground-up, intentional design exercise with the intent of building an AI interface that is, from first principles, about spontaneous and self-directed sharing of expertise.
If, however, I’m wrong and someone is already building this out there and you know about the project, please do point me in their direction because I’d love to see what they’ve found.