By this point, we know that AI can beat sophisticated Turing tests. But how does it “think,” exactly? Perhaps a better and less vague way to phrase it would be: how does the neural net use those artificial neurons to come up with its cognitive or quasi-cognitive behaviors?

Ok, let me take another stab at that: how does AI differ from the ways that humans think with biological brains?

It has to be evident that one major difference is access to information.

We humans are not born with access to information. We have to acquire it – in analog ways that are extremely inefficient. The AI, by contrast, is able to connect to the Internet and scan the known body of information on there in the blink of an eye – which gives the LLMs a tremendous advantage.

Having said that, we come to deeper questions about how AI entities and humans use what they’ve learned. Some scientists and research teams are exploring this in some interesting ways, ways that I think most people would find it hard to follow.

Take this paper on the scaling of AI , versus the scaling of human “performance.” How do you measure cognitive performance? These researchers are game to pursue this fairly broad question, which I find difficult, given the abstraction of even measuring human cognitive performance. We’re biological brains, in physical, biological bodies, with limited data, and all kinds of errant stimuli around us all the time, in a physical environment that is chaotic and disordered, to say the least.

But if you take all of that away and assume that you can track human “scaling,” what this paper seems to show is that humans do better than AI at scaling up. The basic premise here is that giving AI more tokens is like giving humans more time to do better on a task or problem. I’m going to let you read the rest of it yourself, with all of the math and “log-log” distributions, because to me, it’s the philosophical core of this that begs deciphering. In other words, we’re measuring human and AI “performance” toward what ends, exactly?

That leads to another interesting foray into AI cognition that’s oriented toward the creative side of thinking. It plays into our ongoing analysis of AI agents that are more autonomous than what came before them, their digital “ancestors,” so to speak.

My colleague Daniela Rus, director of the MIT CSAIL Lab, is on board with this, pioneering a project with an agent called Eko that is supposed to demonstrate “playful learning.” You can read more here , about how this approach shows how an agent can, in Rus’s words, “interact with its world physically, reason, form memories, and seek novelty.”

When an agent can, say, draw things and create its own little worlds, what does that mean? What part of the Turing test does that confront? Maybe it’s time for another moniker, a “playfulness test?” That’s not very cool-sounding. Maybe we should have a contest for the naming of this concept. Or maybe it’s just a deeper version of the Turing test, one that has humans asking: if AI can do all of this, what are we for?

Leaving aside that thorny onion, let me talk a little bit about what I encountered in the comments.

“The key experimental question is: what would distinguish a homeostatic memory mechanism from ordinary belief persistence or instruction-following?” wrote John R. Smith, commenting on the announcement of Eko in Rus’s LinkedIn feed. “Our Homeostatic Substrate Hypothesis proposes two falsifiable tests: HSH-H3, whether removing entrenched false beliefs restores more capability than additive correction; and HSH-H4, whether measured memory drift predicts subsequent behavioral lock-in. Both have explicit falsification conditions and must be evaluated against simpler competing explanations.”

He left this link to a paper .

If you’re not familiar with what a homeostatic memory mechanism is, you’re not alone.

This type of system maintains stability in the stored information that an LLM uses, resisting changes that disrupt established patterns. This effect might reinforce existing beliefs, regulate memory updates, and preserve consistency over time, potentially making incorrect assumptions difficult to correct or remove from memory. So if the machine hallucinates, it hallucinated for a reason, and that error is baked into its response system.

You can kind of understand this by thinking about how humans misremember things: you might have been told that a certain person lived on Chestnut Road, but you had been thinking of Walnut Road, and sure enough, that next time it comes up, Walnut Road is still your first answer. It’s just hard to self-correct something that’s stuck in the brain.

But AI is different right?

Well, yes and no. What all of this research shows is that we’re still in the very beginning stages of understanding what neural nets are doing when they go to work – and there’s still what we call the “black box problem,” where we don’t really see the workings of the LLM clearly. If you believe the AGI crowd, we never will, because the systems will continue to get “smarter” than us until we reach the era of god-like AI, and, according to some, merge with these immortal minds in what Kurzweil calls “the Singularity.”

Until then, this type of research helps us to at least scan the outlines of how AI works. Implementing or even just identifying something like a homeostatic memory mechanism gives us at least some knowledge of how to predict the behavior of models and agents, and if we put enough of those pieces together, just maybe we’ll be able to keep up.