Human limits and LLM parity

In my prior post, I explore what happens when AI’s convergence point isn’t human parity, but is well beyond human parity. In this post, I want to look at the opposite - what happens as humans try to obtain parity with LLMs, or automated systems in general. As AI systems barrel towards human parity, we’re approaching a philosophical and practical cliff. The question isn’t just when machines will match our cognitive abilities, but what happens when they surpass our most fundamental limitation: the amount of information we can simultaneously process.

It’s not easy to calculate the information storage or processing capability of the human brain, as our brains are not as discrete as computers. Neuroscience research suggests the human brain has a kind of theoretical storage capacity of approximately 2.5 petabytes (2,500 terabytes) based on the number of neurons and the exponential effect of neurons working together to store memories. The brain’s actual working memory - the information we can actively process and manipulate - the “RAM” if you like - is vastly smaller. Research from Caltech indicates that conscious thought processes information at around 10 bits per second, while our senses take in approximately 1 billion bits per second. We can hold a handful of concepts in memory at once. This creates an enormous funnel where massive amounts of information are continuously being filtered out.

Even more significantly, human recall is notoriously unreliable. While we might theoretically be able to store 2.5 petabytes over a lifetime, our ability to accurately retrieve specific information is limited by factors like memory decay, interference, and cognitive biases. In practice, we’re working with a fraction of our theoretical capacity at any given moment, with dubious recall ability. Essentially, we have a tiny amount of RAM, even if we have a large storage capacity on tape drive.

While human cognition is bounded by biological constraints, AI systems face no such restrictions. As of 2024, the total global data volume stands at 149 zettabytes (149 billion terabytes), with projections of 181 zettabytes by the end of 2024. This isn’t just a quantitative difference - it’s a paradigm shift. AI systems can potentially access and process information across this entire data landscape simultaneously, while humans are limited to working with vanishingly small subsets. If humans can compute across data the size of a grain of sand, AI can, at least in principle, compute across a whole beach of data.

This disparity creates an epistemological challenge. AI systems operating at this scale may discover patterns, formulate hypotheses, and even establish truths that require more simultaneous information processing than human brains can achieve. Consider:

• Complex multi-dimensional relationships that require holding thousands of variables in working memory

• Long-range correlations across massive datasets spanning decades of information

• Emergent properties that only become apparent when analyzing petabytes of interconnected data

AI might present us with conclusions that our brains literally cannot validate because we lack the cognitive bandwidth to hold all necessary information simultaneously. We’ll be asked to accept truths on authority - truths that may appear contradictory or impossible from our limited perspective.

This creates a future where humans must become comfortable with a new kind of intellectual humility. We’re entering an era where:

• A and B may appear mutually exclusive to human reasoning, but AI demonstrates their compatibility through higher-order analysis requiring more data than we can process

• Narrative conflicts won’t resolve through better human reasoning, but through acceptance that our reasoning capacity is fundamentally insufficient for certain problems

• Scientific validation may require trusting AI systems to perform analyses we cannot independently verify

To navigate this transition, we need to develop new frameworks for thinking about knowledge and truth:

1. Accept epistemic outsourcing: Recognize that some forms of knowledge will require AI mediation

2. Develop meta-cognitive skills: Learn to evaluate AI reasoning processes even when we can’t verify individual conclusions

3. Embrace cognitive pluralism: Accept that different levels of data processing may yield different but equally valid perspectives

4. Build trust architectures: Create systems for validating AI conclusions without requiring full human comprehension

The question isn’t whether we can keep up with AI’s capabilities - it’s whether we can adapt to a world where some truths exist beyond the reach of human verification. The future belongs not to those who can hold the most information, but to those who can wisely navigate the space between what they understand and what they must trust.