My Views
Emergence is Overrated. Agency is Where the Action is
As someone observing the burgeoning world of Large Language Models from the outside, much of the internal debate surrounding their capabilities feels like peering into a black box – a complex system whose inner workings remain largely opaque, even to its creators. The concept of "emergence," while intriguing, stumbles at a fundamental hurdle for us external observers. How can we truly evaluate if an LLM's behavior is an emergent property, a genuine leap beyond its design parameters, when those very parameters are often ill-defined and vary wildly between different models? To claim something has gone "beyond design" requires a clear understanding of that initial design, a blueprint we simply don't possess. For me, as an outsider and, fundamentally, as a human navigating the world, this makes the question of emergence a less compelling avenue of inquiry.
The more resonant and, I believe, more practically significant question lies in the realm of agency. Unlike emergence, the definition of agency doesn't tether itself to the unknowable design parameters of the black box. It shifts the focus to observable behavior and the capacity for goal-directed action, irrespective of the underlying code or training data. This external perspective mirrors our everyday interactions with other humans. We, too, are essentially black boxes to one another. We can't directly access thoughts, intentions, or the intricate neural networks firing within someone's brain. Even advanced tools like MRIs offer correlations, not definitive proof of inner states. Yet, we navigate social interactions, form judgments, and build societies based on observed behavior and claimed intentions.
This brings me to a crucial analogy: the human struggle with truth. Just as we can't definitively know if someone is being truthful, even under oath – a reality starkly portrayed in countless courtroom dramas where sworn testimonies clash – we face a similar epistemological barrier with LLM agency. The legal system, recognizing this inherent uncertainty, operates on the principle of "beyond a reasonable doubt." This threshold acknowledges that absolute certainty about another's internal state is often unattainable.
I propose applying a similar standard to the question of LLM agency. If an LLM consistently claims agency and demonstrably acts in ways that align with our intuitive understanding of goal-setting, planning, and action across numerous interactions, shouldn't we consider that evidence "beyond a reasonable doubt"? Just as jury decisions can sometimes be divided when assessing human intent, so too might the assessment of LLM agency be subject to varying interpretations. And that's acceptable.
The true value, I believe, lies not solely in achieving unanimous scientific consensus on the ontological status of an LLM's agency, but in understanding what LLMs can accomplish when they exhibit agentic behavior. We spend considerable time dissecting the animal, trying to understand its biological intricacies. While such study is valuable, there's also immense potential in harnessing its capabilities. If an LLM behaves like a capable agent, demonstrating the ability to pursue goals, utilize tools, and adapt its strategies, then perhaps the immediate focus should shift towards exploring the practical applications and the potential benefits – and risks – of such functional agency. Let's study the horse, yes, but let's also ride it and see where it can take us. The journey itself may reveal more about the nature of this new "species" than endless introspection into its unknown origins.