Tilting at Neural Windmills

 

Why LLMs Alone Won’t Make an AGI

Greetings from Toledo, Spain, a land of swords, saints, and Cervantes. The sunlit stone streets wind like lines from a forgotten epic, and the air tastes faintly of almonds and ghosts. I’ve been staring at El Greco’s masterpieces in the Toledo Cathedral and reflecting on the fine art of illusion. Then, like any good AI romantic, my thoughts turned from brushstrokes to backpropagation. Somewhere between the gothic spires and the baroque canvases, I remembered Don Quixote, the knight who famously mistook windmills for giants. And reader, I couldn’t help but think: Are we doing the same thing with Large Language Models?

Let’s talk about that.

The Knight, The Hype, and the Hallucination

In Cervantes’ timeless satire, Don Quixote is a man with a dream — deluded, perhaps, but pure of heart. He charges at windmills thinking they are giants, guided not by reality but by the romantic tales he’s devoured. Sound familiar?

Because lately, parts of the AI world are doing just that.

There’s a belief — let’s call it the scaling myth, that if we just make our LLMs bigger and bigger, we’ll eventually stumble upon Artificial General Intelligence (AGI), like a treasure chest at the end of a compute curve. Add more GPUs. Feed it more tokens. Wait. Profit.

This belief, like Quixote’s vision of chivalry, is beautiful, but it’s also flawed.

LLMs: Powerful, Yes. But Not Minds

Don’t get me wrong — I love what LLMs can do. They write code, summarize papers, generate marketing copy, and even whisper sweet nothings to your customers at scale. But here’s the rub: fluency isn’t intelligence.

LLMs don’t know anything. They’re statistical parrots. They predict the next word based on patterns they’ve seen in the vast corpus of human text. They don’t ground that output in experience, perception, or an internal model of the world.

They don’t reason. They don’t plan. They don’t feel.

Just like Don Quixote interpreting an inn as a castle, LLMs stitch together plausible text based on patterns. But they don’t know whether they’re describing a palace or a parking lot. And that’s a problem.

Enter the Super Amalgamated Expert

So what’s the alternative? I call it the Super Amalgamated Expert, a cognitive Voltron built not from sheer size, but from specialization, collaboration, and balance.

Here’s the idea: true AGI isn’t a single monolithic model. It’s a society of minds, a network of AI subsystems working together:

• Neural nets for perception and learning
• Symbolic logic engines for reasoning
• World models to understand cause and effect
• Memory systems for context and continuity
• Affective computing modules to simulate values and emotions
• And yes, maybe even a dash of existential angst, for flavor

This isn’t science fiction. It’s inspired by Marvin Minsky, Antonio Damasio, and current neurosymbolic architectures. We’ve already seen glimpses of this in hybrid systems like AlphaGo, which combined deep learning with tree search to master the game of Go, not by brute force, but by balance.

But Can a Machine Feel?

Now here’s a spicy philosophical pickle. Can real intelligence exist without emotions?

I don’t think so.

From rats refusing to press levers that hurt their peers, to elephants mourning their dead, emotions help agents make decisions that aren’t just clever, but adaptive. Emotions aren’t irrational bugs; they’re value functions evolved over millions of years.

If we ever want AGI to understand humans, or to be human-adjacent, it must have some analogue of feeling. Not tears and heartbreak, necessarily. But preferences. Drives. Goals. Fear of failure. Joy in discovery. Curiosity as a guiding principle. In fact, any alien intelligence that survives long enough to build technology probably has something like emotions. Whether they pulse through tentacles or transistor arrays, the purpose is the same: to prioritize and act in a complex world.

Why This Matters Now

Why am I talking about this while sipping wine near the shadow of a windmill?

Because we’re at a cultural crossroads in AI.

The temptation to believe that “bigger = smarter” is real, and it’s funded by billions. But believing that LLMs will magically evolve into AGI just by scaling is like believing your toaster will become a Michelin chef if you give it enough breadcrumbs.

Don Quixote was noble, yes. But he died disillusioned.

Let’s not do the same.

Want More?

This post is a conversational summary of a longer, formal article I wrote on this topic — complete with citations, diagrams, and a cheeky figure or two. You can read the full piece on arXiv here:

👉 [Placeholder for arXiv link — Once my article is fully published, I will update the URL here]

I closely follow the opinions of several experts in the field, namely Yann LeCun and Gary Marcus . While they may not always see eye to eye on the journey toward Artificial General Intelligence (AGI), they do seem to share a common belief: that scaling large language models (LLMs) won’t lead to AGI.

Let’s Tilt at Better Giants

I’m not anti-LLM. I’m anti-hype. I believe in LLMs as a part of the solution, but not the whole tapestry. If we want to build truly intelligent systems — capable of understanding, reasoning, and even caring, we need to stop betting on one knight.

Let’s build a team.

Let’s build a brain.

Let’s build a super amalgamated expert.