Gary Marcus Questions LLM Sustainability In AI Crash Debate

Co-intelligence has never just about scale alone but just as important the ease of entry and diverse pathways to explore different ways of knowing.

Low floors High ceilings Wide walls

Recursive AI provides that easy entry, the loop function grows complexity and by building sideways as well as up we explore multiple frameworks.

The remix takes the best of what's come before to create the new, something that has the possibility to exist if patterned.

By patterning we uncover structure and structure exists outside of time.

@mres @medialab @ProjectZeroHGSE

I tried building a simple to-do list app with a basic Claude Pro account. It technically works, but the process is cumbersome, fragile, and far from error-free. The common claim that “everyone will be able to build software this way in the future” for now is wishful thinkig.

People often compare AI to the iPhone, which dramatically simplified computing for the masses. But that analogy falls apart, and not only when it comes to programming through natural language. Creating software is fundamentally different from tapping icons on a screen.

Beyond the familiar “billions of people will become more productive” narrative, there are still significant barriers to turning vague instructions into reliable, maintainable applications.

At the moment, the most visible progress seem is Sally only in meme generation, content farms, and the broader “AI slop” economy. When it comes to building actual software, the limitations become apparent very quickly.

Even for an app with only a few lines of code, I ran into token limits surprisingly fast, forcing constant context management and repeated explanations. Instead of eliminating complexity, the process often feels like replacing one set of technical challenges with another.