Assessing intelligence of AI systems is hard. There is a large spectrum of assessments that I think are reasonable. You can point out that babies and animals show more intelligence in real-life situations. You can point out that AI systems only understand text but intelligence may require multi-modal understanding. You can point out the many stupid mistakes AI systems make. That’s all fair.
What annoys me is the claim that AI systems cannot be intelligent because they are stochastic parrots. It is indeed true that modern LLMs are prediction engines: they build their response (including the “thinking” modes) by predicting the next token and repeating. It is a silly system but why shouldn’t it be able to give rise to intelligence. Just because token prediction is a simple system it does not mean that a complex behavior cannot emerge from it.
First, though, I want to clarify the term stochastic parrots. The term was coined by Bender in this paper published in 2021. In the paper, the authors argue that LLMs are stochastic parrots because they don’t have an explicit mental model of the user and the authors try to assess the risks of that. The paper does not make any comments about intelligence of LLMs, focusing instead on the associated risks of making such systems generally available. I have no objection to this usage but the term has escaped it’s original context since then (Bender even has a FAQ trying to reestablish the original context). Now it is more often used to argue that LLMs cannot be intelligent because they are simple.
That’s the argument I disagree with. There are many systems that exhibit complex behaviors but are made up of simple parts. Just because we understand the parts of the system, it does not mean the system overall will demonstrate the same level of complexity.
For example, emergent complex behavior is common in nature. Eusocial animals like ants and bees are a good example. Each worker in these groups follows relatively simple rules for when to find food, when to build and where to build. Out of the actions of each worker, a broader system emerges that none of the individuals planned. The functioning hive is an emergent system that displays more complexity than the individual subsystems.
Life itself is an example of emergent complex behavior. Each part of a cell does relatively simple mechanical operations. Each organelle can detect a handful of chemicals and respond to that signal in a few fixed ways. Together these organelles combine to form a cell that has more complex behavior than the parts of the cell.
And this extends to human intelligence as well. We have neurons that receive signals from neighboring cells and then can choose to output a signal to neighboring cells based on the input. Compared to the intelligence we derive from them, each neuron is relatively simple. Yet the interaction of these simple components gives rise to intelligence.
So it is not sufficient to point out that LLMs are simple to dismiss their intelligence. We have to look elsewhere for that.