May 15, 2025
LLMs are not whole brains
Tldr: Debates about whether (on their own) LLMs are “truly intelligent” or “comparable to human intelligence” are often misaligned. Intelligence emerges from modular interactions—stochastic LLMs plus extensive deterministic infrastructure.
This systems-level view provides a clearer, more useful framework for understanding, evaluating, and—crucially—building artificial intelligence.
The impulse to immediately analogise LLMs to human brains isn’t particularly compelling to me. Are LLMs genuinely intelligent? The answer is always: to some degree. But honestly, that’s true for basically everything. Ants show intelligence to some degree. Old-school, rule-based expert systems? Yeah, them too. Even your thermostat, if you squint a little. Intelligence lives on a spectrum. Debating if something’s intelligent can hide a more interesting conversation: how exactly is it intelligent, and in what context?
I find it useful to think of LLMs less as standalone "holistic" intelligences (though that framing can be helpful for evals), but as specialised parts of a cognitive system, like specific regions of our brains. Broca’s area handles language (which is itself quite broad); the hippocampus manages memory encoding. Similarly, LLMs specialize in language comprehension, creativity, summarization, and more. They clearly exhibit what we recognize as “intelligence,” at least partially, often in a specialized way, though their surface area keeps expanding.
A single brain region isn’t sufficient to form a complete human mind. Similarly, an LLM doesn’t need to handle every cognitive task itself, especially not from a product perspective. It needs complementary systems, which right now mostly means deterministic software. Just as your cerebellum reliably handles balance through repetitive motor commands, deterministic software does countless essential tasks: routing, filtering, iterating, evaluating, controlling things. Today’s SOTA systems already blend stochastic and deterministic components seamlessly. OpenAI’s Swarm, agent frameworks, Google's Alpha Evolve—they all leverage multiple models working together, combining probabilistic reasoning with deterministic logic and external memory. But even the so-called "basic" deployments—like the GPT-4o endpoint you hit—rely on intricate, rarely acknowledged infrastructure. Every prompt touches not just the model, but a vast supporting ecosystem: specialised hardware (GPUs, TPUs), layered software stacks, intelligent routing, optimised networking.
Does an LLM meaningfully exist without this support? Compute resources, memory, bandwidth, uptime fundamentally enable LLMs. They're oxygen and nutrients. Without them, even the smartest model would just sit there, inert.
Sidebar: I strongly feel that the stochastic nature of LLMs isn’t just a quirk or limitation—it’s a fundamental strength, especially when it comes to language. Human communication is inherently probabilistic; it’s ambiguous, nuanced, and deeply context-dependent. Every conversation we have involves constant subconscious estimations: “What does this person most likely mean?” Probability is woven into every sentence we speak or hear. My gut tells me this is precisely why stochastic systems naturally excel at language—they’re built to handle complexity, ambiguity, and uncertainty. Deterministic programming languages, by contrast, struggle because they demand precise instructions, leaving no space for uncertainty.
Anyway, it’s fair to argue the brain analogy isn’t perfect. Our brains dynamically rewire through plasticity, while today’s LLMs mostly rely on explicit retraining. Yet even this difference isn’t absolute. Models already get fine-tuned and regularly swapped out in production systems. We don't usually call this neuroplasticity, but it fundamentally is—an intelligence evolving in response to new information. Automating this process isn’t even particularly hard to imagine.
Other arguments could focus on how LLMs aren’t bound by biology—how they scale beyond biological constraints, learn faster, handle more inputs simultaneously, and never get tired—but those aren’t really important here. What’s important is that the analogy to the brain remains incredibly valuable—not as a perfect metaphor, but as an engineering framework. Like computer chips, operating systems, or complex software suites, practical intelligence arises from modularity and integration.
That’s how practical intelligence gets built—piece by specialised piece, layer upon interconnected layer, sustained and optimised through thoughtful engineering.