I want to explore something that feels counterintuitive in today’s AI-driven world: moving away from Artificial General Intelligence (AGI) and going in the exact opposite direction. Why? There are two main reasons:
1. General intelligence is overrated.
If I’m building an AI agent for a specific purpose, I don’t need it to be a jack of all trades. I want it to be exceptional—like a laser-focused expert in one area. A specialized agent, fine-tuned for one task, has far more value in many cases. I’m imagining a system that’s good at understanding conversations just enough to do its job but incapable of wandering into unrelated territory. It wouldn’t try to answer questions it wasn’t trained to answer, and it wouldn’t pretend to know things it doesn’t.
2. It’s a straightforward way to avoid existential risks.
Creating an all-knowing, all-capable AI comes with serious risks. If AGI ever became self-improving, we could end up in Skynet territory—a runaway intelligence with catastrophic consequences. But what if we focused on creating tightly controlled, specialized AI systems instead? By “containerizing” knowledge into smaller, purpose-built models, we could avoid many of the dangers inherent in general intelligence. Not only would this make AI development safer, but it would also make regulation easier. Governments could vet these narrowly focused systems far more effectively than they can oversee sprawling, omnipotent LLMs.
Is a Blank Slate LLM Possible?
This brings me to my second question: can we create a “blank slate” LLM?
Imagine an AI that knows absolutely nothing but is still capable of understanding and holding conversations. It wouldn’t start with encyclopaedic knowledge of the world—just enough linguistic ability to talk to you. Think of Vision from Avengers: Age of Ultron. When he first comes online, he doesn’t fully understand who he is or what’s happening, but he’s aware enough to say, “I… am.” He demonstrates reasoning and awareness but has no knowledge base to draw from.
So, can we decouple language understanding from factual knowledge? Could we train an AI just to process language—without preloading it with the world’s information?
Why Build a Dumb Bot?
At first glance, it might seem counterproductive. Why would anyone want an AI this “dumb”? Well, I would. And I think a lot of others would too.
Here’s the thing: current LLMs rely heavily on guardrails to stay in check. Guardrails monitor input and output, trying to prevent the model from saying anything harmful, false, or off-topic. But these systems are far from perfect. For example, earlier versions of ChatGPT would refuse to repeat certain phrases, abruptly cutting off sentences midstream if they detected something sensitive.
Hackers and clever users routinely find ways around these guardrails. One famous exploit involved asking ChatGPT to act like an API and output answers in JSON, bypassing restrictions entirely. While OpenAI patched this, new workarounds will inevitably emerge. Guardrails are a reactive solution to a fundamental design problem.
A blank slate LLM avoids this issue altogether. By starting with zero knowledge, it’s inherently limited in what it can say or do. It wouldn’t need guardrails because it simply wouldn’t have the information required to generate harmful or irrelevant responses.
The Data Dilemma
Of course, there’s a catch. Training a blank LLM to be coherent without overloading it with unrelated data is tricky. Language is inherently tied to knowledge. To understand and produce coherent sentences, an AI needs at least some exposure to the world.
This reminds me of a few examples:
-
Vision from Avengers
Vision’s “I… am” moment captures the essence of this idea. He’s intelligent enough to process his existence but too new to truly understand it. A blank slate LLM could behave similarly—aware of its role and limitations but incapable of answering questions outside its domain. -
The Real-Life Mowgli
The inspiration for The Jungle Book came from a child raised in the wild, completely isolated from human civilization. This child adapted perfectly to the forest, communicating with animals through grunts and howls. However, upon returning to society, the child couldn’t grasp human language. Their experience in the forest wasn’t the problem—the lack of linguistic exposure was. This raises a question: does coherence require a broad base of unrelated knowledge, or can it emerge independently?
Rethinking AI Development
This brings me back to the original idea: could we create a system like Vision—a blank slate, self-aware enough to admit “I don’t know” but capable of learning exactly what it needs for a specific task?
Such a system would:
- Be tightly specialized, operating only within its designated scope.
- Answer “I don’t know” when confronted with out-of-scope questions.
- Eliminate the need for guardrails, as it wouldn’t possess the knowledge to cause harm.
- Be simpler to regulate and safer to deploy.
In essence, it would be the opposite of an AGI—a purpose-built agent designed for safety and precision.
Final Thoughts
Here’s the thing: this vision might be achievable, but not with LLMs as we know them. Current LLMs are a local maxima—a system where we dump massive amounts of data and hope it sticks. They’re powerful, but they entangle self-awareness and knowledge in a way that’s impossible to separate.
To build the kind of AI I’m describing, we’d need to start from scratch. We’d need a new approach that prioritizes modularity and specialization over generalization. Instead of creating one model that tries to solve everything, we could build a network of narrowly focused agents, each excelling in its domain.
Such a shift would make AI safer, more accountable, and better aligned with human needs. It’s a thought experiment for now, but it might just be the future of AI development.