Compound AI and Fit-for-Purpose Edtech Applications

June 30, 2024

A debate is raging in the AI community between those who believe that large language models (LLMs) and their offspring can or will be able to do (more or less) anything, and those who believe that a wider diversity of AI methods are likely to be needed.

Similarly but much less visibly, in the edtech community, many seem to believe that the future lies in platforms that can address a broad spectrum of learning needs, while others like LearnerShape see a large role for narrower applications that address specific learning needs.

These two debates are closely connected, analytically and technically. On the latter, technical connection, the emerging approach of “compound AI” supports the view that the future lies with diversity – i.e. with AI not dominated by LLMs and with edtech not dominated by platforms.

What Compound AI Is and Could Become

In an influential February 2024 Berkeley AI Research blog The Shift from Models to Compound AI Systems, a group of 11 AI researchers:

“define[d] a Compound AI System as a system that tackles AI tasks using multiple interacting components, including multiple calls to models, retrievers, or external tools.”

In contrast, the authors explain that “an AI Model is simply a statistical model, e.g., a Transformer that predicts the next token in text”.

The blog then goes on to explain why compound systems are better than standalone models at delivering results on a variety of tasks, providing several examples of successful compound AI systems. For example, Google’s AlphaCode 2 code generation application combines LLMs with a code execution model and a clustering model.

Although much of the AI community and the public seem to be obsessed with the growing capabilities of LLMs – including ones from OpenAI, Google, Anthropic, Meta, Mistral and others – the compound systems approach is gaining traction, and is supported by views that some AI leaders have been expressing for years.

Facebook AI Research leader Yann LeCun tweeted in 2023: “On the highway towards Human-Level AI, Large Language Model is an off-ramp.” AI gadly Gary Marcus has been making this point for years, including in his 2019 book with Ernest Davis Rebooting AI: “Getting to broad intelligence will require us to bring together many different tools, some old, some new, in ways we have yet to discover.” And just this week, Bill Gates introduced the idea of AI “metacognition” in an interview:

“The big frontier is not so much scaling. We have probably two more turns of the crank on scaling … . That’s not the most interesting dimension. The most interesting dimension is what I call metacognition, where understanding how to think about a problem in a broad sense and step back and say, ‘Okay, how important is this answer? How could I check my answer? What external tools would help me with this?'”

This point of view has also been a guiding principle for LearnerShape from the beginning, leading me to conclude my first blog on this site in January 2020 by saying that:

“we have a conviction that there is a huge opportunity to leverage advances in machine learning and data science to improve the quality of our recommendations.”

Fit-for-Purpose Edtech Applications

The connection of these principles of compound AI to edtech is obvious. Our Head of Data Science Jonathan Street saw this connection in October 2020 when he wrote:

“A more flexible approach is vital to creating a productive, engaging learning experience. … At LearnerShape, we are driven by the conviction that this challenge can be addressed now, for the first time, by using AI … . AI allows us to deliver flexible learning pathways for any content, using any skills framework, on any platform.”

We are now accelerating the strong commercial potential of this vision with our PlaylistBuilder YouTube curation application. In cooperation with Google, we are solving an easy-to-explain but unaddressed problem of better curation of YouTube videos, with an easy-to-use solution that is both simple and highly powerful. PlaylistBuilder uses a multi-step curation pipeline that combines AI and other techniques. The application is currently in beta testing with leading universities, and we will move towards a wide release in the coming months.

PlaylistBuilder is just one big step in our broader vision of improving learning by enabling a wide range of applications that are tailored to specific learning needs. Ongoing advances in artificial intelligence, combined and remixed using the approach of compound AI systems, will be the rocket engine that speeds us along this course. 

Maury Shenk, Founder & CEO, LearnerShape