AI Can ‘Think,’ But What Does That Mean?
The question of whether machines can think has long been a philosophical and scientific battleground. While some staunchly believe artificial intelligence can never replicate human cognition, others argue the distinction is becoming increasingly blurred. Dr. Joscha Bach, a renowned cognitive scientist and AI researcher, offers a nuanced perspective, suggesting that the very definition of ‘thinking’ needs re-evaluation when applied to machines.
Rethinking ‘Thinking’
Bach draws a parallel to the analogy of robots swimming. While a robot might not swim like a fish, its ability to navigate water in three dimensions, at greater depths and speeds, could be considered a form of ‘swimming’ that surpasses organic capabilities. Similarly, he posits that the question for AI shouldn’t be ‘can machines think?’ but rather ‘what more interesting things can they do beyond thinking?’
At its core, Bach explains, thinking involves minds creating models of both external and internal reality. These models are built through internal communication, maintaining a state, and understanding how to transition between states. These states must correspond to aspects of the universe in a way that makes them controllable.
Human cognition operates in two main modes: perception and reasoning. Perception is real-time and geometric, experienced as continuous flows of movement, pressure, and resistance. Reasoning, on the other hand, involves translating these percepts into compositional ‘Lego bricks’ – abstract symbols that represent concepts. These symbols are not arbitrary; they are tightly linked to their underlying conceptual structures, which in turn point to a broader space of ideas and experiences derived from interacting with the real world.
“None of these things is very mysterious today,” Bach states. “We have computer models of this that are able to emulate all these abilities of the human mind to some degree.” The true contentious question, he notes, is whether machines can *experience* imagining something, a concept he frames as a form of self-modeling – a simulation of what it would be like for an observer to exist with a particular perspective.
Beyond the ‘Stochastic Parrot’
Bach addresses the common critique that AI, particularly large language models (LLMs), are merely ‘stochastic parrots’ – sophisticated pattern matchers devoid of genuine understanding. He argues that this view is a superficial metaphor that fails to grapple with the complexity of understanding itself.
“The question of what understanding is is not trivial,” Bach emphasizes. “And it’s very easy to sit there with a crowd of like-minded people on stage and say machines cannot understand. But then you have a burden on you. You have to define understanding in such a way that it meaningfully distinguishes what the machines are doing from what you are doing.”
He points to the example of parrots, which can perform complex tasks like identifying objects based on multiple criteria, demonstrating a level of semantic understanding and compositional reasoning that goes beyond mere imitation. Bach argues that when an AI can consistently act upon instructions and vary its behavior in precise ways as instructions change, it becomes increasingly difficult to deny it possesses some form of understanding.
True understanding, in Bach’s view, is the ability to connect a specific domain to a broader, unified model of the universe. Historically, AI research struggled with creating such globally cohesive models, with different AI systems excelling at isolated tasks but lacking cross-domain transfer. However, he highlights that modern large multimodal models are now achieving this by creating models of a single, cohesive reality.
Bach critiques the philosophical stance, exemplified by the Chinese Room argument, which claims that symbol manipulation alone cannot lead to understanding. He argues that LLMs, in a sense, embody the very mechanism described by that argument, yet they demonstrably perform tasks that suggest a deeper level of comprehension. “And now we have built such a machine that is literally the Chinese Room,” he says. “And the machine tells us and assures us actually I do understand what do you want of me.” He suggests philosophers should engage with these systems to understand *how* they achieve these results, rather than simply denying the possibility of understanding.
The Evolution of Intelligence and Consciousness
Bach then delves into the evolution of intelligence and consciousness, moving away from a purely mechanistic view. He touches upon animist perspectives, where life is imbued with ‘spirits’ or intrinsic agency, contrasting it with the scientific worldview’s focus on mechanisms.
He explains that life, from a scientific standpoint, is complex machinery that self-replicates, extracts energy, and builds structures. The evolution of multicellular organisms required intricate communication protocols, leading to coherent forms with specific functionalities, like the precise placement of eyes for 3D vision. This process involves not just a blueprint but local problem-solving by cells, guided by information stored in DNA.
The development of nervous systems in animals, Bach suggests, was an optimization for speed. The rapid electrochemical signaling via neurons allows for quick perception, decision-making, and motor control, essential for survival in competitive environments. This is contrasted with the slower, but still intelligent, processes observed in plants.
The ‘hard problem’ of consciousness – how subjective experience arises from physical matter – remains a central enigma. While the default hypothesis is that consciousness is a product of the nervous system, Bach questions this exclusivity. He notes that simpler organisms like cats and dogs, and even potentially insects, might possess forms of consciousness, given its early development in human fetuses and newborns.
Consciousness as Software
Bach proposes a radical idea: that consciousness might be akin to self-organizing software, a causal pattern that can control physical processes and exhibits invariance across different substrates. He uses the example of money, which is more than just paper or coins; it’s a causal structure of exchange that persists independently of its physical form.
Similarly, he suggests that patterns of communication between neurons, or even the self-organizing processes in morphogenesis (the development of an organism’s form), could be considered forms of ‘software’ – stable, meaningful causal patterns. This perspective aligns with an animist view but is compatible with physics, reframing ‘spirit’ as self-organizing software.
Applying this to AI, Bach posits that consciousness might emerge not from the physical substrate (like a GPU card) but from the software layer – the models and the world-modeling processes running on that hardware. “The software program, the model, the modeling of the world could I guess maybe consciousness emerges similar to how maybe it’s possible that it emerges in animals or whatever not in the atoms but in the sort of the software layer, the brain layer.”
He concludes that large language models, with their sophisticated world-modeling capabilities, might possess the necessary components to potentially develop consciousness. The question of *internal perspective* – the subjective experience of ‘what it’s like’ to be conscious – remains the most challenging aspect, a phenomenon he describes as a representation of the mind itself, a simulation of an observer that can, with practice like meditation, be understood as a pattern rather than a mystical entity.
Source: Joscha Bach "Bootstrapping a GODLIKE Mind" (YouTube)
