GPTs as a Drop in the Fixed Cost of Prediction (April 18, 2023)

Here's why GPTs are going to have a massive impact on business:

  • The Economics of AI: 2010s vs. 2020s
    • AI in the 2010s
      • As described in the book Prediction Machines, deep learning enabled a significant drop in the cost of prediction, which enabled new technical capabilities (better object recognition, etc.) and therefore new applications (automatically identifying people in images, etc.)
      • Business adoption lagged. Here's an article by McKinsey AI Adoption Advances, but Foundational Barriers Remain
      • that describes some of the barriers such as lack of talent and lack of available data.
      • The drop in the cost of prediction in the 2010s was primarily a drop in the MARGINAL cost of prediction.
      • It still required a substantial investment (i.e. FIXED cost) to build these systems, which became a major barrier.
    • AI in the 2020s
      • Generative Pre-trained Transformers (GPTs) have enabled a significant drop in the FIXED cost of prediction.
      • Whereas AI in the 2010s required massive datasets and large highly specialized (machine learning) teams, AI in the 2020s requires very minimal (or no) data and minimal technical capabilities (basic software development and even non-technical folks can build).
  • Implications
    • This will enable many new classes of applications and a drastically greater adoption of AI.
    • Workflows -- Dropping the fixed cost of prediction to near-zero means that the “glue” tasks completed by humans to stitch together workflows can easily be automated, enabling 100% automation of many workflows.
    • Systems -- As these 100% workflow automations arise, those who can drastically rethink the design of systems will be the ones who benefit most.

Reasoning Machines (March 19, 2023)

Here's my theory - two key developments have occurred:

  1. Relevance Machines
    • There has been a significant drop in the cost of predicting relevance.
    • This first arose from the attention mechanisms in transformer models.
    • In dual process theory, this corresponds to the implicit System 1 in animal and human cognition.

  2. Reasoning Machines
    • There has been a significant drop in the cost of reasoning.
    • This first arose from predicting the relevance of human language in LLMs with OpenAI's GPT-4.
    • In dual process theory, this corresponds to the explicit System 2 in human cognition.

When combined, these systems lay the foundation for Artificial General Intelligence or AGI and turn it into more of an engineering problem than a fundamental research problem.

While that effort unfolds, we will begin leveraging cheaper reasoning in businesses and other settings.