The man who named AGI thinks the companies racing to build it are burning billions on a dead end.
The Summary
- Ben Goertzel, who coined "AGI" in 2005, says Big AI's transformer obsession is "a waste of resources" because all LLMs are fundamentally doing the same thing.
- Transformer models can't learn continuously from new experiences. They reset to baseline parameters after each interaction, unlike human intelligence.
- As models scale, intelligence gains are becoming exponentially more expensive, with unclear ROI at current trajectory.
- The financial stakes are now so high that labs can't afford to seriously explore alternative architectures.
The Signal
The AI industry is in a curious position. It's discovered something that works, transformer models trained on massive datasets, and now it's pouring every available dollar into making that one thing bigger. OpenAI, Anthropic, Google, Meta, they're all running variations on the same play: pre-train a transformer on more data, add more parameters, throw more compute at the problem. The bet is that scale equals intelligence, and if you just keep stacking chips, AGI emerges like a phase transition.
Ben Goertzel isn't buying it. As someone who literally wrote the book on artificial general intelligence two decades ago (co-authored with DeepMind's Shane Legg, no less), he sees the current approach as fundamentally limited. The problem isn't that transformers don't work. They obviously do. The problem is what they can't do.
"All these LLMs are kind of doing about the same thing."
Here's the critical limitation: transformers don't learn continuously. When you talk to ChatGPT, it doesn't update its weights based on your conversation. It processes your input, generates a response, and then forgets. The next time you interact, it's back to its baseline trained state. Humans don't work this way. Every conversation, every experience, every failure updates our internal model of the world in real time.
This matters enormously for agents. An AI agent that can't learn from its mistakes in the field, that has to be retrained offline at massive expense every time you want to update its behavior, is an AI that can't truly adapt. It can follow instructions brilliantly. It can generate coherent text. But it can't grow from experience the way a human employee does after six months on the job.
Key constraints of the transformer paradigm:
- Billions in compute required for initial training
- No real-time learning from new experiences
- Each intelligence improvement requires full retraining cycle
- Diminishing returns as models scale up
The economics are starting to look suspect. Early transformer scaling showed clear gains. Double the parameters, measurably better performance. But those gains are getting more expensive. The question isn't whether you CAN make GPT-5 smarter than GPT-4. It's whether the improvement justifies another hundred billion in compute spend. And what happens when you're spending a trillion for a 10% improvement?
Goertzel's point about resource allocation hits hard. Because the financial stakes are so enormous, no major lab can afford to seriously bet on alternative approaches. If you're Anthropic or OpenAI, you can't tell your investors you're pivoting away from transformers to explore cognitive architectures that might work better but have no proven track record. You're locked in. The companies with the most resources are trapped by their own success into iterating on the same fundamental architecture.
The Implication
If Goertzel is right, the current crop of frontier models represents a local maximum, not a path to AGI. They'll keep getting better at specific tasks. They'll become more useful tools. But truly general intelligence, the kind that learns continuously and transfers knowledge across domains the way humans do, might require completely different architectures that nobody's funding right now.
For builders, this suggests two things. First, current AI agents will plateau in capability unless someone cracks continuous learning. Plan accordingly. Second, if you're working on alternative approaches to intelligence (cognitive architectures, neurosymbolic systems, bio-inspired models), the window might be opening. The transformer consensus is strong, but it's also expensive and possibly hitting diminishing returns. That's when paradigm shifts happen.