The AI Mainframe Era

Seeing AI For What It Is, Not What We Wish It to Be

In my previous post on Alien Intelligence, I argued that Generative AI is a qualitatively different kind of intelligence than human intelligence, and viewing AI Agents / AI Teammates / AI Coworkers as the primary modality for this tech to deliver value is significantly undershooting its potential. In this post I will sketch a different model of thinking about AI applications which is closer to fulfilling the full promise of disruption with AI.

The core idea is that we are entering the AI Mainframe Era, where the most transformative AI systems are not those working under employees’ guidance, but those standing above them—acting as central hubs, orchestrating operations with a bird's-eye view.

The AI Mainframe, as visualized by AI.

Here’s the argument:

1. Real application-layer value in AI has yet to be created. See AI’s $600B Question.

2. All current product thinking around AI is basically some offshoot of the ChatGPT model: an AI assistant that helps users/employees accomplish tasks. In the B2B SaaS setting, this has morphed into the idea of an “AI Teammate” that replaces/augments certain job functions and does it faster/better/cheaper.

3. The “AI as human assistant” or “AI teammate” paradigm has two major problems: (1) it overestimates how close AI is to human way of thinking—leading to unmet expectations and uncanny valley, and (2) it underestimates how much AI could benefit us in other ways that better suit both AI’s and our strengths.

4. True advantages of AI come from scale and polymath-ism: aggregating large amounts of diverse context, extracting patterns from it, while crossing traditional boundaries of individual human expertise. This goes beyond what we expect a single human or groups of humans with homogenous skills to be able to do in an organization. Hence, by definition, an AI Assistant tasked with automating existing activities of individual employees is barely scratching the surface of what’s possible with AI.

5. To fully leverage the scale advantage of AI requires designing AI systems as the central hubs at the core of our organizations and networks: plugged into thousands of information streams, churning out insights, decisions and recommendations for allocating resources and directing efforts — with human oversight, of course. This is conceptually like the “mainframes” of the mid-20th century, except at a much much grander scale, enabled by several decades of Moore’s Law at work.

6. The concept of “AI Mainframes” is not new: it is widely present in fiction, and is parallel to the concept of AI Assistants. See this table on Human-like vs Alien-like AI from my previous post.

7. AI Mainframes will let us re-imagine vast categories of enterprise software: ERP, Business Intelligence, IT Consulting, Management Consulting, etc.: trillions of dollars of value creation/destruction. Most enterprise software, processes and business models are predicated on working around the inherent limits of the individual human brain to let large numbers of humans coordinate value-creating business activities. All of this must be re-thought when you place an AI brain that far surpasses the context and expertise limits of human brains at the center of an organization.

8. As pre-training scaling laws slow down, inference time scaling is the “hot shit” for the foreseeable future. [See Generative AI’s Act o1.] The AI Mainframe paradigm can more naturally absorb inference-time scaling of reasoning capabilities than AI Assistants can — imagine a central, thinking entity that quietly performs System 2 reasoning on vast data streams behind the scenes and “comes up to surface” when human contact is necessary. This is better than watching AI Assistants “think” for minutes while you stare at your screen.

9. We’re collectively drawn to the AI Assistant paradigm because it is uncomfortable to think of AI as a mainframe hub rather than an agent subservient to humans (see: Skynet). But as I quickly learned when I took a surfing lesson off the coast of Maui once—either you ride the wave or are ridden by it and left with a salty aftertaste. The steady march of technological progress and the gravity of profit-seeking efficiency maximization in capitalistic enterprises both point towards the AI Mainframe direction.

10. I don’t believe for a second that the dystopian nightmare is the default outcome along this path; rather we are capable of building AI systems that remain under human oversight while providing novel value to society.

11. The AI Mainframe era will create many more jobs than it takes—if you believe that the limiting factor to massive GDP growth is inefficient mis-allocation of resources by human managers, then an AI system making (some/most/all?) decisions should be a welcome upgrade.

12. We may have to pass through an AI Mainframe era in order to unlock the branch of the technology tree that gets us to the coveted vision of truly useful AI personal assistants — what Ben Thompson calls the Consumerization of AI — this task might be up to the next generation of AI-native humans.

13. This post opens up more questions than it answers; long road of building ahead and I’m excited.