AI in the wild
When intelligence equals experience, could AI be trained in the world instead of the lab?
Can AI “experience” the world outside the computer? If so, what might this mean for AI, and for us?
AI experience is the topic taken up by David Silver and Richard Sutton in a forthcoming book chapter. It reminded me of the very first article I posted on Medium, an in-depth analysis of the term “experience”. I presented that article in a 2019 talk titled “Experience beyond the human”.
So here I will revisit the talk, the article and the topic of experience, in light of Silver and Sutton’s chapter. I will summarize their argument, offer a critique and some implications, by forming a definition of the term “experience” suited to AI, while acknowledging its limitations.
All this provides a foundation for some of our current work in spatial intelligence, or what we might call “experiential AI”.
Why experiential AI?
As you probably know, current AI foundation models, like the one that powers your favourite chatbot, were basically trained on all the data on the internet. This has some benefits, because the internet serves as an archive for all sorts of human wisdom and behaviour – good and bad.
Using all this human-created data meant that AI didn’t have to go out and find it. And it reduced the need to create abstract conceptual models, the authors write. But with AI models’ need for ever-more data, the companies that built them quickly hoovered up pretty much the entire internet. The result, the authors write, is that “the amount of human-generated data is approaching a limit.”
And because more data equals greater accuracy for current models, new data sources are required. Synthetic data is an option: data generated by AI by recombining existing data – all that AI-generated text, image and video content is finding a second life as training data. But this approach is less likely to capture all the complexity, noise, and edge cases of real-world phenomena. Without the quality and diversity of real-world data, an AI is more prone to errors. That limits its predictive power, if a model is overfitted to its training data and less able to account for edge cases.
There is clearly more data in the real world than currently on the internet, so we could simply scan more books and digitize more insights from human experts and others. This is no doubt happening. But the other problem that Silver and Sutton identify is that data frozen in books, minds and archives is static: it might have been created at particular times and places and gone out of date, carrying with it historical biases. Silver and Sutton write:
if an [AI] agent had been trained to reason using human thoughts and expert answers from 5,000 years ago it may have reasoned about a physical problem in terms of animism; 1,000 years ago it may have reasoned in theistic terms; 300 years ago it may have reasoned in terms of Newtonian mechanics; and 50 years ago in terms of quantum mechanics. Progressing beyond each method of thought required interaction with the real world: making hypotheses, running experiments, observing results, and updating principles accordingly.
Training with all that human-created data prepared AI models well for brief interactions between humans and chatbots and other generative AI systems. But what happens as robots increasingly inhabit the human world? Also, the authors write, what about “long streams of grounded, autonomous interaction”? What about properties of the world that persist and change over long periods of time? That would bring AI closer to natural intelligence, they say.
Going natural
This is where experience comes in: it refers to an agent’s experience of its environment. The authors describe experience as “data that is generated by the agent interacting with its environment”. This matches my definition, in which an “agent” can be any human or nonhuman entity that engages sensorially with the world. Sensors are used by computational agents in a similar way to our sight, hearing etc. Sensory perception, according to environmental psychologist J.J. Gibson, “is a keeping-in-touch with the world, an experiencing of things rather than a having of experiences. It involves awareness-of instead of just awareness.” (I’ve written about his work here.)
Over time, Silver and Sutton write, AI agents “will inhabit streams of experience,” and this too conforms to my understanding of experience as unfolding over time. “The notion of experience as a continuous flow implies linearity and direction,” I wrote in my article, “connecting past, present, and future.”
Thus, “intelligence” becomes synonymous with “experience”. Both are situated in the world, not in the brain of an individual agent.
This is like the negation of the self in Japanese philosophy. The contemporary philosopher Emmanuele Coccia holds a similar view: that beings cannot be separated from their context – the fish from water, nor the human from the air and all that surrounds us. “It is in front of the world, in front of nature, that the human being can truly think,” he writes.
Transcendental philosophers like William James go even further, equating experience with riding a wave, or the wind, following the natural contours of a trail or even a sidewalk. Experience is letting the world pass through you, and harnessing this power for your own locomotion – both physical and metaphysical.
Contemporary philosopher Benjamin Bratton echoes this sentiment in taking a broader view: “Natural intelligence also emerges at an environmental scale and in the interactions of multiple agents. It is located not only in brains but also in active landscapes.” Natural evolution is “heredity interacting with experience, filtered through the inherent randomness of development.” [source] Since natural intelligence evolved this way, Bratton proposes that AI too needs to evolve out in the world.
There’s related evidence supporting this, regarding another kind of nonhuman agent: animals. Primatologist Christine Webb shows that when we remove animals from the wild in order to study, eat or otherwise exploit them, we essentially transform them into machines, as well as perpetuate the myth of human exceptionalism. Now think about how almost all AI is developed in research labs instead of natural environments.
A systems approach situates agency in components which are themselves systems, in nested, communicative relations. Bratton specifically points to rapid advances in machine learning combined with shrinking hardware to suggest that a human-centred approach is insufficient for understanding the increasingly global, distributed nature of computing systems.
In my own work, environmental experience relates directly to learning – including machine learning. For example, “experiential learning” can involve a cycle consisting of direct experience, reflecting on that experience, abstracting that knowledge into general principles, and then applying it in practice. (Education researcher Richard Kolb details this process in his book Experiential Learning.)
I found this process to relate directly to the steps of computational thinking, as developed originally by computer scientist Jeanette Wing and used by, for example, Google: deconstructing a dataset, looking for patterns in the pieces, abstracting the findings into a general theory, and applying that to design an algorithm. I found these steps so useful, I had my students apply them to all sorts of non-computational areas.
Computational thinking and experiential learning both involve collecting real-world data, then channeling that into practical activity. Some of my research has therefore involved “experience recording” technologies. These include sensors to capture data on location, time, temperature, light and sound levels and more. Mobile phones count here, since they usually contain all these capabilities.
Intelligence in context
The idea, as applied to human learning, is that context is important.
In current AI systems, context refers to how much data a system can hold in memory, so that your past conversations and actions inform present ones. For experiential AI however, it means grounding its knowledge in real-world data. This not unlike the ancient Hindu and Greek traditions of empiricism, where truth can only arise from phenomena you observe or encounter first-hand.
AI aside, computer science has mostly interpreted context as spatial and temporal data – especially location. But I found value in Falk and Dierking’s tripartite definition that divides context into personal, social and physical spheres. Specifically, when I adapted their model and applied it to both human and nonhuman agents, I found that meaningful learning happens when one agent connects its personal context with another agent’s. Personal context, for both humans and nonhumans, means individual history, preferences and particularities – the sum of one’s experience, in other words.
So experience means life experience. For humans this is clear: you can read all the books you want (we could call this “wisdom”), but when it comes to riding a bicycle, driving a car or navigating a new environment, then actual, real-world experience (“street smarts” is one term) is much more valuable.
For humans, this is backed up by research. Even for relatively abstract activities like playing chess, there is evidence that having a mental model of the game is far less useful than simply having played many previous games. Chess Grandmasters are experts because for any given board configuration, they’ve seen similar ones before and so know the best move.
Similarly for “innate” ability. It used to be believed that inherited genes accounted for around 80 percent of our potential, but subsequent research flipped this around, so that scientists now believe that “culture” or real-world learning instead accounts for around 80 percent, genetics being relegated to the remaining 20 percent.
The rewards of reinforcement
What that means for AI is that the more training data, the better. We knew that. But again, for understanding and navigating the world outside the screen, it means “life experience” in the real world, not training simulations or collected wisdom.
In practice, this specifically places value on reinforcement learning: Silver and Sutton are experts in this. In a simple sense, this means rewarding an AI for “right” answers. And if the right answers are not known at the start, the agent must discover through trial and error which approach works best, and it tests its assumptions against external reality, rather then received human wisdom. “To discover new ideas that go far beyond existing human knowledge,” the authors write, “it is instead necessary to use grounded rewards: signals that arise from the environment itself.” More specifically, from an “agent’s interactions with both the user and the environment as input”.
If this sounds like a dystopian scenario of AI becoming more autonomous and detached from humans, consider that for a nonhuman agent, the “environment” might actually be a human. A continuous stream of experience could include a person’s “health, educational, or professional needs towards long-term goals over the course of months or years,” Silver and Sutton write.
Alternately, an AI could autonomously conduct ongoing experiments in materials science, medicine, or design. “By continuously learning from the results of their own experiments,” in the authors’ optimistic view, “these agents could rapidly explore new frontiers of knowledge, leading to the development of novel materials, drugs, and technologies at an unprecedented pace.” In all cases, rewards come from “an abundance of grounded, real-world signals”.
Okay, but…
We could critique this view for being overly task-centered, focused on finding single answers to questions or problems that might have more than one solution. This is an issue with design more broadly, within a larger culture that values productivity and endless growth, as I’ve written before.
More specifically, the authors’ focus on superhuman performance and “superintelligence” assumes a direct comparison between humans and AI – something I have also written strongly against. So-called AI has no “understanding” of any given topic, nor can its “experience” of the world be at all close to ours.
Nonetheless, if we simply accept the idea that “experiential data” is more useful for AI training than static human-generated archives, then let’s dig deeper into “experience”, and what it might mean for AI.
Language and its limits
In my previous article, I unpacked the term “experience” by drawing on definitions from relevant philosophy and anthropology. I came up with a useful definition structured around processes of transduction, transformation, and transactions.
As a verb, individuals experience particular internal states or external phenomena. To be present means looking, listening, recognizing. “We may have 20 years of experience, or 20 years of accumulated experiences,” I wrote, “we speak of individual experience and collective experience.” Experience as a verb is subjective, while as a noun it refers to “a bounded phenomenon or situation that many presume can be empirically studied and designed.”
Given this, we can say that AI “experiences” internal states (inference, “thinking”, “reasoning”) and external phenomena via sensing capabilities. To a large degree, experience is sensory experience. As such, “the language of experience doesn’t respect the norms of description,” writes philosopher Federico Campagna.
Therefore, I contrasted direct experience with interpretation, which is when experience gets translated into language. This holds true for AI, as in our current work in spatial intelligence, wherein AI is trained to make meaningful interpretations of real-world spaces based on relevant theory that it has been trained on.
Related here are stories, as interpretive devices. Simple linear stories can be constructed by AI systems based on experiential data it collects – for example, a series of time-stamped locations. I wrote, “Experience itself can be considered a story“.
But I noted that experience can also be nonlinear and can distort one’s sense of time. Think of an emotionally intense experience you have had, and how time might have felt faster or slower during, or upon reflection.
Getting emotional
Here we hit a big roadblock, when we come to such experiences. Computers cannot, and never will, have emotions. I believe this strongly. They can only measure, predict and simulate them based on numeric data. A system might be able to guess what an emotionally intense experience might be, based on a training dataset. But it can never truly have such an experience.
Anthropologist Victor Turner describes “moments when life is lived most intensely” as aesthetic experiences. An emotional response to such an experience signals the inherent unpredictability of such experiences. This is where my first dimension of transformation comes in. “Any experience that does not violate expectation,” according to the philosopher G.W.F. Hegel, “is not worthy of the name experience.” Contemporary philosopher Byung-Chul Han echoes this: ““Experience transforms. It interrupts the repetition of the ever same.”
Aesthetics is another subjective human quality closely tied to emotions. On one level, it can simply refer to a “sensory knowing of the world” according to Anna Munster. But philosopher Jaques Ranciere identifies it as not only subjective but political. My colleague Carolina Rito writes that
aesthetics is understood as a domain of social and political life in which the forms of the sayable, the thinkable, and the doable are part of how we experience and make sense of the world. Ranciere argues that “aesthetics acts as configurations of experience that create new modes of sense perception and induce novel forms of political subjectivity.”4 Ranciere rightly points to the domain of aesthetics as operative in all the forms of political and human experience, also extended to all abstract and physical forms, e.g., from pixels and molecules, to non-human beings.
In this sense, we could say that computers count as abstract, physical, non-human beings. But they are aesthetic only in terms of human sentiments (subjective, emotional responses) towards them. We might say that some nonhuman entities, like birds and plants, have subjective “tastes” in an aesthetic sense, but again computers can only mimic or predict these in a quantitative manner. Any transformation that occurs in an AI system as a result of sensory input could only be called aesthetic in terms of a statistical average.
Therefore, my second dimension of experience is more fitting for AI systems: transduction. This refers to the conversion of one form of energy, signal, or information into another. That includes sensory information, which can be converted biochemically into emotional responses in humans, or electrically into digital data in computers. Such a process can transform the human or computational system into a new state, and this constitutes experience. A neural configuration, whether biological or simulated, is changed, with certain pathways and connections strengthened.
The specific means by which this happens is transactional: For philosopher John Dewey, “‘experience’ refers to transactions between us and the objects and events that make up the world in which we act.” Kolb similarly prefers the term “transaction” for his experiential learning cycle, “because the connotation of interaction is somehow too mechanical, involving unchanging separate entities that become intertwined but retain their separate identities.” For him, “transaction” implies a more fluid relationship in which both sides are changed.
From transactions to transformation
The term “transaction” might sound too economic. Indeed, “experience” has become commodified, as marketers have identified a shift in purchasing behavior among millennials, away from products and towards intangible things like travel, concerts, and immersive entertainment.
Conversely, now AI companies stand accused of plotting to “steal human experience” and turn it into proprietary data. Silver and Sutton’s proposed “Age of Experience” for AI adds another dimension. If a fairly autonomous AI agent can go out and collect data from the world, they write,
this provides fewer opportunities for humans to intervene and mediate the agent’s actions, and therefore requires a high bar of trust and responsibility. Moving away from human data and human modes of thinking may also make future AI systems harder to interpret.
The optimistic view is that such agents could, on the other hand, detect and solve problems before humans do, or modify their behaviour automatically when it doesn’t align with human values and goals. The test case is the “paper clip scenario”, which shouldn’t materialise because an experiential AI would correct itself before it consumes all the resources on Earth.
A more pessimistic take is that AI agents could extend the commodification of human experience, if they start experiencing the world for us, relaying it to us via language, images, video or whatever. Personally, I believe that people are still biologically driven to value direct experiences.
Moreover, if we locate experience in our interactions in and with the world, rather than in our heads, then the world is inseparably inside us and we are inside it, as Coccia contends. Amidst all the talk of autonomous AI agents, we shouldn’t forget or give up our own agency. “If the world is in all its beings,” Coccia writes, “this means that every being is capable of radically transforming the world.” Silver and Sutton’s proposed “Age of Experience” for AI should complement, not commodify, our own.
It sounds crazy, but doesn’t our current way of training AI systems also amount to commodification? What would happen if they were developed and trained in the wild?