The Rise of the Cognitive Architect

A longer read about how Asimov’s Robopsychologist might be coming to life in AI as the Cognitive Architect.

Last week at AI Engineer World Fair in San Francisco, engineers and technologists presented cutting-edge AI technology being developed. The conference was motivated by a think piece published by @swyx and @FanaHOVA, defining the role of the AI Engineer (AIE). This role sits at the intersection of product and technology, specializing in applying Large Language Models to software development.

There was a common trend in some of the work presented: a need to overcome these models' more messy, probabilistic nature by making them more reliable and engineering-like. This is no surprise since AI Engineering focuses on applying these models to real-world problems. From building a question-answer engine like Perplexity to a tool for better note-taking like Granola, a certain level of determinism is needed to create products that can withstand real-world use.

While at the conference, I also followed @hwchase17 on Sequoia's Training Data podcast, as he talked the notion of cognitive architectures. While this was part of a larger conversation about turning LLMs into reasoning agents, this notion of cognitive architecture pointed toward the idea that using LLMs is as much about human-to-machine cognition as it is about technology.

Human-machine cognition means understanding human agency and translating it into instructions, considering the intricate steps we take when performing tasks (both conscious and subconscious), and translating them into machine code—that is, model language.

Now, a small detour into science fiction:

In Asimov's I, Robot, robopsychology is the field of understanding and troubleshooting the complex psychological and behavioral issues that arise as robots (read: AI) become more sophisticated and integrated into human society. Dr. Susan Calvin, a robopsychologist, navigates the dilemmas posed by the laws of robotics.

Here, I see a parallel between Asimov's study of robotic psychology and the challenges of training AI models to do what we want them to do. We've built a technology we're still working to understand. @mustafasuleyman even suggests it isn't just a technology, but a new kind of digital species—what an idea!

Given the challenges of engineering with language models and their incredible power, there might be a gap in applied AI waiting to be filled by Cognitive Architects—a role most likely already in existence in research labs at OpenAI, Anthropic, or Google.

Much like an architect works alongside a civil engineer to research how humans live, the Cognitive Architect may work with computer scientists to study human cognition and its relation to model cognition. From examining the boundaries of the model's probabilistic knowledge (latent space) to using the right keywords and canonical definitions to express instructions, to decomposing complex human tasks into the flows of actions, tools, and models, the Cognitive Architect might be part of the recipe to applying these models in ways that don't just tame it down to an engineering feature, but also harness their true nature:

To predict where we're most likely to go next.