The uniting of Principal Investigators (PIs) and their disciplines into a coherent research team and associated research program is one of the important accomplishments (Figure 9 in Chapter 2). Researchers from all disciplines involved work side by side in a single building, share laboratories, use the same office space, participate in the same graduate program, and work together on joint research projects. SCIoI will build upon and deepen this spatial and intellectual cohesion via the methodology described in this section.

Moravec’s Paradox

SCIoI’s approach to intelligence research remains influenced by Moravec’s Paradox. Already in the 1980s, robotics and AI researchers such as Hans Moravec, Rodney Brooks, and Marvin Minsky realized “it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility” (Moravec, 1988). Implications of this observation are central to SCIoI. First, the complexity of understanding and reproducing a skill is expected to correlate with the evolutionary time scale over which it developed. SCIoI will therefore include an emphasis on fundamental sensorimotor processes and aspects of their embodiment. Second, from the evolutionary history of animal behavior, we can postulate that intelligence must be based on–in fact, must include–these foundational sensorimotor skills. Therefore, within SCIoI, we will consider insights into sensorimotor processes and their embodiment as an important basis for the study of higher cognitive functions.

Weak decomposability–A defining characteristic of intelligent systems

Herbert Simon categorized complex systems based on their degree of decomposability (Simon, 1969). In a system consisting of multiple components, decomposability expresses the relative complexity of the components and their interactions. If the components of a system work independently of each other, Simon calls this decomposable. If the complexity of interactions is significant, but low relative to the complexity of the components, Simon defines it as near decomposable. Almost all engineered systems fall into this category.

Compelling evidence gathered in SCIoI suggests that intelligent systems are weakly decomposable. A system is weakly decomposable if its components and the interactions between these components are both complex. To understand such a system, it is not sufficient to understand the system’s components, because a substantial part of the system’s behavior results from the interactions of its components.

Weak decomposability has important implications for studying intelligence. “Complex, integrated systems are basically non-decomposable” (Schierwagen, 2012). We prefer to re-label “basically non-decomposable” as”weakly decomposable” because a truly non-decomposable system would possess no structure and that can be ruled out. The “decompositional analysis [of weakly decomposable systems] is inherently destructive to what makes the system complex– such a system is not decomposable without losing the essential nature of the complexity of the original system! [emphasis added]” (Schierwagen, 2012).

To investigate intelligence in artificial or biological systems, we must keep “the essential nature” of intelligence intact. This means that, in addition to studying components of intelligence (vision, reasoning, memory, learning, etc.), we must study the interactions of these components. This requires that we study intelligent systems in rich and ecologically relevant situations so that the complexities of the system manifest themselves in observable and interpretable behavior (Calhoun and Hady, 2021). This implies that artificial, together with “biological behavior needs to be studied and modeled in context, that is, in terms of the real problems faced by real animals [and robots] in real environments” (Webb, 2001).

Our approach to intelligence research must be tailored to the weak decomposability of intelligent systems to stand a chance of identifying the islands of structure depicted in Figure 1. An important consequence of weak decomposability is that a reductionist approach to intelligence research cannot succeed.

A synthetic approach to intelligence research

To help unite the disciplines of intelligence researchand to ensure we would be able to address the weak decomposability of intelligent systems, we implemented a synthetic approach to intelligence research. This approach requires that all scientific insights be turned into synthetic artifacts. This include robot systems, computer simulations, or mathematical formalisms. The artifacts represent a mechanistic encoding of research results, serving as a lingua franca for the transfer of research results between disciplines.

Mechanistic encoding does not only describe the properties of the observed behavior, but it can also produce the behavior, both in the original context and in novel contexts. This will become an important precondition for testing the ability of intelligent systems to generalize.

Since the inner workings of synthetic artifacts are articulated explicitly in mechanistic encoding, they become inspectable, comparable, and classifiable. Across the synthetic models produced within SCIoI, we found repeated patterns of information processing. Thus, these patterns satisfy an important requirement of a principle, namely that they determine a fundamental quality of intelligence, a general insight that transcends a single implementation or instance.

Given these insights enabled by mechanistic models, we begin to view intelligence differently. Reminding ourselves of Simon’s quote that “complexity, correctly viewed, is only a mask for simplicity” (Simon, 1969), we believe that this mechanistic view makes important steps toward disentangling the lines of Figure 1, enabling the discovery of islands of structure.

A mechanistic view of intelligence

To this day, scientists have not agreed on a definition of intelligence. A report published by the Board of Scientific Affairs of the American Psychological Association, the largest and most influential psychological organization in the world, has concluded that “although considerable clarity has been achieved in some areas, no such conceptualization has yet answered all the important questions, and none commands universal assent. Indeed, when two dozen prominent theorists were recently asked to define intelligence, they gave two dozen, somewhat different definitions” (Neisser et al., 1996). This quote remains valid even today.

The vast majority of conceptualizations of intelligence focus on behavior (Sternberg and Kaufman 2011; Legg and Hutter 2007; Goldstein et al., 2015). In the establishment proposal, we wrote: “Our scientific strategy begins with a preliminary, behavior-centric definition of intelligence. Behavior is intelligent if it is adaptable, general, cost-effective, goal-directed, and can be performed in the real world.” This preliminary definition reflected SCIoI’s continued focus on behavior as a starting point for understanding intelligence.

Over the course of SCIoI, we realized that a mechanistic characterization of intelligence is much more likely to capture the islands of structure we are looking for, in the visual analogy of Figure 1. This departure from a behavior-centric definition toward a mechanistic characterization marks a major step for SCIoI. It enabled much of the progress we made in understanding the principles of intelligence. However, we will, of course, continue to rely on observations of behavior as our source of information about intelligence. We will not be content simply with descriptive characterizations, but instead, will focus on identifying mechanistic models capable of producing intelligent behavior rather than describing it.

Principles of intelligence

According to the Oxford American College Dictionary, a principle is “a general scientific […] law that has numerous special applications across a wide field,” describing “a fundamental quality determining the nature of something.” Such a principle captures relevant and simple regularities, as illustrated by the islands of structure in Figure 1.

Recurring mechanistic patterns identified via the synthetic approach to intelligence research represent possible mechanistic principles of intelligence. We currently believe that a principle of intelligence must satisfy the following four properties:

1. Generality: A principle of intelligence is one that, in accordance with the definition of principle above, pertains to intelligent behavior. It generalizes in response to different tasks, environments, species, and levels of abstraction.
2. Mechanistic behavior generation: A principle of intelligence characterizes the process of producing behavior. This process can be viewed as a computation (Brock, 2024).
3. Regularity: A principle captures a regularity, i.e., a reproducible and predictable relationship between the agent’s embodiment, features of the environment, and the behavioral consequences of their interactions. It limits the degrees of freedom in the space in which the solution is to be found and can be viewed as a lower-dimensional manifold in that space. Such a regularity enables robust and general behavior. Regularities correspond to novel islands of structure in the analogy of Figure 1.
4. Constructiveness: A principle represents a design guideline for intelligent agents.

We identified candidate principles in SCIoI along with comprehensive empirical support across species, tasks, environments, and levels of abstraction.

A mechanistic definition of intelligence

In the view of intelligence proposed here, intelligent agents are characterized by the principles they implement to produce intelligent behavior. This is an important departure from the behavior-based characterizations of intelligence that are standard in the literature (Sternberg and Kaufman, 2011; Legg and Hutter, 2007; Goldstein et al., 2015). SCIoI‘s mechanistic definition of intelligence is one of the defining characteristics in our emerging discipline. We believe it is a central enabler for much of the progress we have achieved during the first funding period.

Why is a mechanistic approach to intelligence such a game changer?

Earlier, we explained the role that synthetic artifacts play in our efforts to unite the disciplines. We also stated how these synthetic artifacts have accelerated the progress within SCIoI, consistent with predictions made in the literature (Allen, 2014). But are there also principled arguments to support our mechanistic approach?

Behavior arises from interactions between an agent and the environment. Both determine the resulting behavior. The behavior changes if the agent changes, even in the slightest, and if the environment changes, even in the slightest. The combinatorics of all possible changes of agent and environment are daunting. However, if intelligence is characterized in terms of behavior, then this is the task we must face: We must characterize the space of all possible behaviors. But this is not the route SCIoI takes: we characterize the agent instead, with important implications.

Any characterization of behavior, agent, or environment must be an abstraction, i.e., it must ignore some aspect of the “full truth.” A suitable characterization accomplishes this by describing regularities that capture the observed phenomenon. In physics, such regularities are called laws, such as f = m · a. In Figure 1, these regularities correspond to the regions with parallel lines–the targets of our scientific inquiry: the islands of structure.

Let us now assume that we have found such regularities in the space of agents and in the space of environments. A characterization of behaviors inherits the complexity of both of these spaces, as it combines them into a more complex space. Consequently, if we could get away with characterizing just one of the spaces, we would face a much simpler problem. Describing the agent is easier and more compact than describing the agent’s possible behaviors, since the description of an agent does not need to consider environmental variations.

Viewing the intelligent agent itself as the conceptual target for defining intelligence, as opposed to attempting to describe the intelligent behavior produced by the agent, leads to a significant reduction in the complexity of the science required. At the same time, since we can combine insights across all species and tasks, we can accumulate much more data and evidence to find relevant patterns. The impact of this shift in perspective has fueled our substantial progress toward understanding intelligence. This then is our current, still preliminary definition of intelligence:

An agent is considered intelligent in a particular ecological niche if it leverages the niche-relevant principles to produce behavior.

We can now simply state that the agent produces behavior, as the principles necessarily ensure that within the ecological niche it will be “adaptable, general, cost-effective, goal-directed behavior in the real world,” citing parts of our initial behavioral definition of intelligence. The restriction “within a particular ecological niche” acknowledges that no agent will be able to perform intelligently under all circumstances. Agents must be tailored to their task and environment.

Generalization as a test of intelligence

Intelligence enables agents to apply generalized experience and knowledge to novel situations. This includes solving new and old problems under new circumstances and with novel approaches. Generalization is a defining behavioral trait of intelligence.

Our mechanistic characterization of intelligence does not explicitly consider variations of the environment, it only focuses on a description of the agent itself. However, we consider environmental variations as the set of conditions under which the agent must demonstrate its intelligence. Rather than including the environment in the description of intelligence, we consider this variation the litmus test for intelligence. An agent is only intelligent if it is capable of generalizing across a substantial and task-relevant range of environmental variations. We will therefore use the generalization abilities of agents as a test of intelligence. Generalization must be enabled by the principles of intelligence via the regularities they encode.

The analytic/synthetic loop

The synthetic approach to intelligence research led to a novel, mechanistic characterization of intelligence, in which generalization becomes a key indicator of intelligence. At the core of this view are the principles of intelligence that capture islands of structure. It therefore becomes pivotal for SCIoI to be able to effectively identify such principles. The analytic/synthetic loop (A/S loop) leverages the complementary strengths of the analytic and synthetic disciplines to accelerate the identification of mechanistic principles.

The study of biological intelligence reveals insights about “actual” intelligence, which we only know with certainty to exist in biological species. Research with biological species, however, incurs significant experimental and ethical challenges. For example, it is difficult to study specific behaviors in isolation and there is limited access to the internal state of animals. Experiments cannot be easily repeated with a single individual, as learning occurs during each trial. Some types of experiments, such as ablation studies and lesion approaches, cannot easily be conducted with biological species. Even though we get a chance to study “true” intelligence, the experiments are complex, and results can be difficult to interpret.

This is quite different for artificial species. Currently, it is difficult to argue that artificial agents exhibit “true” intelligence. Therefore, the opportunity to gain insights about intelligence is reduced compared to biological species. However, there are advantages: Experimental conditions can be controlled, the internal state of the system (robot, computer program, etc.) is accessible, and experiments can be conducted that would be impossible or unethical with biological subjects.

The A/S loop combines the analytic and synthetic approaches into a unified research methodology to facilitate the identification of the principles of intelligence. Insights from the biological disciplines are encoded in synthetic artifacts. The synthetic artifact then becomes the subject of analytic scrutiny itself, with fundamentally different experimental opportunities. These opportunities deliver new insights that would have been impossible or much more difficult to obtain by studying the biological subject directly. The new opportunities to analyze a synthesized biological aspect of intelligence generate novel predictions that, in turn, can be tested in biological species, delivering new insights to be encoded synthetically. This closes a loop that accelerates progress, relative to efforts purely confined to the analytic and synthetic domains.

The A/S loop means “understanding biology to build robots, and building robots to understand biology” (Webb, 2001). It has become a trademark of SCIoI research and many of the results were enabled by it. The A/S loop offers an effective methodological approach to identifying the principles of intelligence.

A comparative approach

Given the scientific progress achieved in the analytic/synthetic loop, we will extend this methodological concept.

Our comparative approach (Harvey and Pagel, 1991) to intelligence research is illustrated in Figure 2. We study intelligent systems not in isolation but relative to one another. By relating the differences in systems to the differences in observed behavior, we obtain insights not only about the two systems themselves, but we can also formulate hypotheses about the relationship of the dimensions of comparison. In other words, in addition to learning about single points in the design space of intelligent systems, we learn characteristics and formulate hypotheses that apply to a region of the design space. The comparative approach therefore leads to an increase in the efficacy of SCIoI’s research methodology.

Figure 2: Illustration of the comparative approach: The colored regions represent the different dimensions along which comparisons can be made. The overlap in regions of different colors illustrates that comparisons can also be made among different characteristics. The many opportunities for comparisons permit the selection of the most promising one for each research question.

We consider many dimensions along which to compare intelligent systems. Their biological and artificial genesis represents one such dimension; this dimension of comparison leads to the analytic/synthetic loop. In SCIoI, we will include many additional dimensions of comparison within the comparative approach to intelligence research (Figure 2):

› analytic and synthetic disciplines
› time (ontogeny and phylogeny)
› biological and artificial species
› cultures
› behaviors
› environmental conditions
› tasks
› combinations of these dimensions

We anticipate that the comparative approach will accelerate our investigations into the principles of intelligence. It will also enhance the cohesion among the PIs, as PIs from different parts of intelligence research will collaborate closely.

Tinbergen’s four questions–adding the temporal dimension

The temporal dimension, i.e., the comparison of intelligent systems across different stages of ontogeny and phylogeny, is an important addition within the comparative approach.

SCIoI relies on behavioral experiments and observations as the primary source of information about intelligence. One of the central tenets of studying behavior, which has withstood the test of time, are Tinbergen’s four questions (Tinbergen, 1963) which represent four complementary explanatory categories–or levels of analysis–for behavior.

Figure 3: Tinbergen’s four questions

The four questions help explain SCIoI‘s research program and how it has evolved over the years. The focus of research in SCIoI was so far on what is referred to as the “static view.” It corresponds to the two top questions in Figure 3. The “causes” are captured by the mechanistic implementation of the agent. The “function” represents the purpose of the behavior and is related to the task.

Now, we also emphasize the “dynamic view”, i.e., the temporal and developmental aspects that explain intelligent behavior. With this added emphasis, we now cover all aspects of Tinbergen’s four questions. This will enable us to leverage the full analytic power behind this approach to behavioral analysis. This broadening of our research agenda is critically enabled by the inclusion of new PIs from Leipzig (Eusemann, Liebal, Haun), who have extensive experience with the comparative approach and in particular the temporal dimension within it.

Finding the islands of structure

Let us return to the visual metaphor of Figure 1. It captures the notion of finding the “pattern hiding in apparent chaos” by removing the complexity that masks simplicity (Simon, 1969).

We are now able to articulate the high-level plan of identifying intelligence-related islands of structure. SCIoI‘s mechanistic definition of intelligence focuses on mechanistic principles, i.e., principles that describe the inner workings of the agent and characterize how the behavior is generated, instead of describing what kind of behavior is generated. We believe that this is the perspective that is best suited for achieving an understanding of intelligence. Here is why:

1. Mechanisms to produce “real” intelligence are implemented inside the agent. Intelligent behavior unfolds when these mechanisms interact with the environment. Since we want to understand the mechanisms, we must observe behavior in ecologically relevant environments. Only then will we be able to observe the manifestations of components and their interactions of weakly-decomposable systems. This explains why the combination of mechanistic principles and tests of generalization suffice when characterizing intelligence.
2. Our attempt to characterize intelligence via the specification of an agent’s internal workings greatly simplifies the task at hand. We can ignore the combinatorics that arise from the myriad ways in which the environment might vary.
3. Mechanistic characterizations facilitate the synthesis of intelligence in artificial agents.
4. Mechanistic characterizations provide a concrete language of mutual understanding of intelligent phenomena between They produce testable predictions in different disciplines, in species, and on different levels of abstraction.
5. Mechanistic characterizations generalize better than descriptive models of behavior.

There is a key challenge we face that cannot be ignored. So far, we simply spoke of mechanistic principles, as if the level of abstraction for describing principles would be obvious. This is far from the truth. Finding the right level of abstraction is in fact the key challenge when revealing the islands of simplicity. With reference to the analogy of Figure 1, we must move up and down in the image until the island of structure for intelligence becomes apparent.

When we strive to define a mechanistic model for an intelligent agent, we are referring to the model that is the most appropriate for the intelligent solution to a task. Such a model accurately captures the regularities involved in the phenomenon we study or in the task the agent solves. This implies that in our preliminary, mechanistic definition of intelligence, we must specify what we mean by “mechanistic” as it varies with the specific aspect of intelligence we investigate. Our previous mechanistic characterization of intelligence can thereby be extended:

An agent is considered intelligent in a particular ecological niche, if it leverages the niche-relevant principles to produce behavior. These principles are considered mechanistic with respect to the ecological niche if they capture the regularity exploited by the agent to produce the behavior.

Therefore, the term “mechanistic” does not refer to a fixed level of abstraction but to the level of abstraction that best describes the mechanisms for producing the observed behavior–the level that reveals the island. The notion of mechanistic principles on different levels of abstraction is therefore essential for finding the islands of structure.

Bottom-up, top-down, and middle-out

The ability to formulate mechanistic principles on different representational levels of abstraction enables a highly flexible approach to innovation.

To avoid the formidable complexity of biological agents in ecologically plausible environments, many scientists study narrow phenomena. This permits an accurate, quantitative analysis of the phenomena. The validity of the analysis rests on the assumption that a higher-level understanding of animals and their behavior can be obtained from studying these narrow aspects. But the hopes associated with this bottom-up approach when studying intelligence have not fully materialized. Despite amazing progress in explanations of many narrow aspects of animal behavior, there has only been slow progress in providing higher-level explanations that tie together these fragmented explanations. Nevertheless, SCIoI will continue to contribute to science with a bottom-up approach, by studying precisely defined phenomena under controlled experimental conditions (see  Figure 4 and caption).

Figure 4: Different modeling approaches in intelligence research: The top layer of the figure represents complex phenomena with complex context; the bottom layer represents elementary, more isolated phenomena. Arrows indicate top-down modeling (blue) and bottom-up modeling (red). Shading indicates the amount of evidence supporting a model, with light colors indicating less evidence. The width of the colored region represents the generality of a model. To discover principles of intelligence, SCIoI will continue performing bottom-up research, i.e., studying specific phenomena under well-controlled conditions (panel-left). We will also perform top-down research to reflect the weak decomposability of intelligent systems (panel-middle). The formulation of candidate principles at intermediate levels of abstraction also enables a third approach, called middle-out (panel-right), in which hypothesized islands of structure serve as starting points for local top-down and bottom-up investigations, pushing outwards from points in the middle (panel-right). This offers a new way to discover islands of structure. As we cover the phenomenological space associated with intelligence incrementally, we build a complete understanding of intelligence.

SCIoI also employs a top-down approach for studying complex and intelligent behavior. When introducing weak decomposability as a property of intelligent systems, we argued that we must study the interactions between components to understand behavior. To experimentally observe the effect of components and their interactions, we must choose experimental settings that are sufficiently challenging to the agent for the interactions to be reflected in observable behavior. We therefore must study biological and artificial agents in ecologically relevant and sufficiently complex environments (Webb, 2001). Since this will produce widely varying behavioral patterns, a detailed and predominantly quantitative approach to the analysis will be inadequate. We must be able to identify broader patterns rather than detailed variations. Therefore, we will exploit high-level, established knowledge and human insight to generate testable hypotheses for principles in a top-down manner (see panel-middle, Figure 4).

The formulation of mechanistic principles of intelligence in SCIoI will enable a third approach to research and modeling. We call this middle-out . Mechanistic principles characterize regularities at intermediate levels of abstraction, anywhere between fully integrated, complex phenomena and isolated, narrow phenomena (see caption of Figure 4). This means that principles capture information about a region of the phenomenological space lying between the complex, integrated phenomena at the top of Figure 4 and the bottom of the figure, representing relatively isolated, elementary phenomena. Furthermore, mechanistic principles are designed to generalize to new contexts and varied conditions. This generality of mechanistic principles enables the local exploration of the phenomenological space from the region characterized by the principle outwards, as illustrated in the right-hand panel of Figure 4. The mechanistic principles of SCIoI enable a novel methodological approach to intelligence research by supporting bottom-up, top-down, and middle-out research and modeling. This will enable a more rapid exploration of the phenomenological space associated with intelligence, leading toward a complete understanding of intelligence.

Characterizing biological and artificial intelligence

We discussed how principles help identify islands of structure. We also explained how, starting from these islands, we can explore the hypothesis space about intelligence. We now turn to the question of how to use principles in explaining intelligence. What would such a principle-based characterization of intelligence look like?

Figure 5: Illustration of how biological intelligence has evolved in a conceptual space spanned by the principles of intelligence

To characterize intelligence based on principles, we must be able to separate intelligence from non-intelligence. We must also be able to articulate why this boundary exists and what the characteristics are that give rise to that boundary. To illustrate this, we imagine a space whose dimensions span the characteristics of systems, i.e., the principles implemented in the system. In Figure 5 , this space is drawn in the horizontal plane. There are certainly many dimensions/principles required to characterize any kind of intelligent system. But since this is a thought experiment and we merely care about illustrating the concept, we content ourselves with only two dimensions, labeled “Principles” in the figure. Additional axes indicate that there are actually many more dimensions. Any imaginable intelligent system now corresponds to a point in this space, whose coordinates are sufficient to specify the system in its entirety.

Let us first look at biological systems, illustrated as the green, coral-like structures in Figure 5. These systems evolved over long time scales (the vertical axis). We consider the period between today and a point t=0 in the past at which intelligent systems already existed.

Consider the slice t=0 of the space shown in Figure 5. There are five green regions, indicating parts of the design space in which intelligent systems existed in biology. We refer to these regions as species, being fully aware that the notion of species in biology is disputed (Hey et al., 2003). Within SCIoI, a species is a set of system designs sharing fundamental characteristics that differentiate it from other designs.

As evolution progresses along the vertical time axis, these regions can change their shape or move in the plane. These changes represent changes in system design, either as a response to changing environmental conditions or as an adaptation to the ecological niche. There can be convergence and divergence events. Convergence corresponds to several species beginning to share the fundamental system traits relevant for intelligence because of evolutionary pressures. Divergence corresponds to a species separating into two different system designs to occupy different ecological niches. This evolution of the green regions gives rise to the green “trunks” that capture the evolution of intelligent systems over time.

Our goal is to characterize relevant regions of system designs, such as the green trunks, using principles of intelligence. The differentiation between different trunks arises from the set of principles being different for each of them. This implies that principles of intelligence are not universal. They are beneficial in specific ecological niches and therefore characterize the system designs capable of establishing themselves in those niches.

Figure 6: Illustration of how biological and artificial systems can be characterized in the same conceptual space

Figure 6 includes artificial systems. Like biological ones, they are characterized by the principles they realize. These principles separate AI systems into “species.” This way of characterizing AI systems leads to explainability, because we can explain the system’s behavior based on the principles.

Figure 6 shows that biological and artificial species can occupy an overlapping conceptual space. This corresponds to them addressing overlapping ecological niches with an overlapping set of principles. But artificial species are not bound to the process of evolution—we can design artificial systems to deliberately explore the design space beyond biological examples. The inclusion of artificial systems thus provides SCIoI with an important opportunity to gain insight into principles not yet discovered by biological processes.

A variety of biological and artificial species

Figure 7: Variety of biological and artificial species in SCIoI

For a large number of problems there will be some species of choice, or a few such species, on which it can be most conveniently studied.
(adapted from Krogh’s Principle Krogh, 1929)

This idea, originally formulated by Claude Bernard (Bernard, 1865), is central to our comparative approach to intelligence research. Our ability to gain insights into intelligence will be greatly accelerated by an appropriate selection of species to study, either biological or artificial. To support a broad research agenda, the PIs and collaborators of SCIoI have experience with and access to the following 14 biological and 12 artificial “species templates” (because the species are not fully specified yet).

Mammals: Humans (Abdel Rahman, Brass, Hanning, Haynes, Hertwig, Haun, Kurvers, Lazarides, Liebal, Onnasch, Rolfs, Wiese); nonhuman great apes: chimpanzees, bonobos, gorillas, orangutans (Haun, Liebal); pigs (Eusemann, Haun, Liebal, Thöne-Reineke); mice (Lewejohann, Sprekeler, Thöne-Reineke)

Birds: Cockatoos (Auersperg (external PI), Brock, Kacelnik); fish-eating birds (Krause, Romanczuk); chickens (Eusemann, Thöne-Reineke)

Fishes: Clonal Amazon mollies, sulphur mollies, Trinidadian guppies (Krause, Romanczuk)

Insects: Honeybees (Landgraf, Lewejohann)

Robots: Nao humanoid robots (Hafner); Pepper humanoid robots (Hafner); mobile manipulators (Brock, Toussaint); soft robots (Brock); Kilobots swarming robots (Romanczuk); Thymio II wheeled robots (Romanczuk); Robofish (Krause, Landgraf)

LLMs: ChatGPT-style decoder models (Akbik); BERT-style encoder models (Akbik); PaLM-E-style models (Toussaint); multimodal models (Brock)

Reinforcement learning agents (Akbik, Brock, Hafner, Romanczuk, Toussaint)

This diversity in artificial and biological species will enable researchers in SCIoI to follow Krogh’s principle, picking the model systems most suitable for the question at hand. This diversity in model systems will also facilitate identifying the constraints that lead us to the islands of structure illustrated in Figure 1.

Summary of main methodological advances in SCIoI

The definition of principles of intelligence provides a new methodological focus. While we so far could only speculate about principles, we now have aligned the research program with our new insights and resulting implications. The revised methodology has already proven to be highly effective.

Given the explosive progress in AI, we have now included a branch of the research program dedicated to AI systems, enabling us to leverage these recent developments. Here, we consider AI systems to be instances of artificial species, analogous to biological ones.

Given the new branch investigating artificial species, we will intensify our efforts to apply analytic methods to artificial systems. AI systems have become so complex that studying their behavior becomes challenging. SCIoI’s methodology, with its synthetic approach and the analytic/synthetic loop, is well-suited to addressing this challenge.

The comparative approach with its many different dimensions is now an integral part of our methodology, significantly expanding the opportunities for insight into intelligent behavior and the principles that produce it.

We have approximately doubled the number of biological species in our investigations, expanding our opportunities for discovery within the comparative approach.

Our methodology now explicitly includes the temporal component of Tinbergen’s four questions, meaning that we now fully cover all levels of analysis for behavior. This is supported by the new PIs from Leipzig who have the relevant expertise (Eusemann, Liebal, Haun).

Given our understanding of principles of intelligence, we have improved our ability to devise principle-driven engineering practices for engineering intelligent systems.

The insights obtained so far and the resulting methodological advances described here enable us to address higher-level cognitive functions and more complex intelligent behaviors. While the focus has so far been on relatively low-level behaviors, our research program now includes projects on higher-level cognitive functions, for example, on metacognition and theory of mind (ToM).

A historical note on SCIoI and the cognitive revolution

SCIoI’s research methodology follows the tradition of the cognitive revolution which lead to the founding of cognitive science (Miller, 1956). This revolution produced a shift in focus from the characterization of behavior to the characterization of the inner workings of the mind. SCIoI advances the field of intelligence research much more comprehensively. Rather than attempting to identify the architecture of a single cognitive system (the human mind), SCIoI seeks to characterize intelligence as a whole in its entirety. At first, this sounds much more challenging. We contend that it leads to a significant simplification and disruptive progress. By raising the level of abstraction of the overarching question and by broadening the perspective across all forms of intelligence and levels of abstraction, we can gather more and diverse evidences more easily across different species and contexts. This makes it much easier to identify relevant patterns, corresponding to candidate principles of intelligence. Moreover, SCIoI’s methodology carefully distinguishes between descriptive models of behavior in the tradition of cognitive science and psychology, and truly mechanistic models that generalize. These aspects are all embedded into SCIoI’s research methodology. The resulting shift in scientific approach brings about fundamental progress, as we have demonstrated so far.

Positioning within the research area

Positioning with respect to disciplines

Cognitive science: SCIoI shares with cognitive science the goal of understanding cognition and intelligence based on an interdisciplinary research program. The conceptual foundations of cognitive science remain valid in SCIoI, as described earlier when we discussed the cognitive revolution. However, cognitive science did not develop into a separate discipline, as initially intended. Instead, it retained a substantial overlap with psychology. Núñez et al. (2019) determined that most papers are authored by psychologists and cite mostly other psychology papers. Faculty members are mostly psychologists by training, and most courses taught in cognitive science departments are psychology courses. In contrast, SCIoI is building a new scientific discipline from a cohesive research project with a set of PIs working together in a single location on joint projects using a shared infrastructure. SCIoI generates a deep conceptual collaboration between the involved disciplines, as evidenced by the data in Figure 9. The substantial progress made in SCIoI toward understanding intelligence and cognition validates our methodological approach to a unified science of intelligence.

Artificial Intelligence: SCIoI is also a science of Artificial Intelligence (AI). AI has reached a level of complexity that makes it nearly impossible to fully understand AI systems, given methods available in AI today. We respond to this challenge by applying SCIoI’s methodology to artificial systems as well: SCIoI investigates the principles of intelligence shared by all forms of intelligence, both artificial and biological. In that sense, AI systems are model systems in SCIoI, similar to biological species. SCIoI will study state-of-the-art AI systems, in conjunction with biological intelligent systems, to identify the shared principles of intelligence.

The transformative developments in AI technology are undeniable. They already play a key role in SCIoI. Our research leverages these technological advances in many ways. Our team includes three PIs who develop this technology (Akbik, Brock, Toussaint) and many are applying it (Akbik, Brock, Lewejohann, Romanczuk, Sprekeler, Toussaint).

SCIoI aims to develop an approach to artificial intelligence that is fundamentally different from today’s deep learning foundation models. This is urgently needed: Researchers have estimated that biological intelligence is six orders of magnitude more energy- and data-efficient than today’s AIs (Smirnova et al., 2023). In addition, the need for exponential amounts of data remains a barrier for progress (Udandarao et al., 2024).

There are also strategic reasons for Germany and for Europe to invest in alternative approaches to AI. Currently, among the top 15 organizations training foundation models, not one is based in Europe (Maslej et al., 2024, Figure 1.3.17, page 60). Two foundation models are from Germany, compared to 102 from the United States of America and 20 from China (Maslej et al., 2024, Figure 1.3.18, page 61). The technology is owned and dominated by large American corporations (Google, Meta, Microsoft, OpenAI) (Maslej et al., 2024, Figure 1.3.17, page 60). If successful, an alternative approach to AI would enable Germany and Europe to leapfrog such developments and establish itself as a leader in AI.

The cost of training large AI models is skyrocketing. It increased from 12 million US$ for PaLM in 2022 to almost 200 million US$ for Gemini Ultra in 2023 (Maslej et al., 2024, Figure 1.3.21, page 64), making the technology accessible to only a few. The financial cost is largely caused by enormous electricity requirements and therefore translates into CO₂ emissions. An alternative, energy-efficient approach to AI, such as the one developed in SCIoI, might contribute to democratizing the technology and reducing its impact on climate change.

AI is liable to be the most transformative technology invented by humans. There should be several alternative and complementary paradigms to support this transformation. AI technology developed in SCIoI can ensure that this transformation is sustainable and safe. SCIoI’s conceptual approach to intelligence can enable AI systems to show the abilities currently exclusive to biological systems, like generality, autonomy, and versatility. In addition, the resulting technology promises to exhibit the advantageous performance parameters of biological systems, such as very low energy consumption and low data requirements. The technology enabled by SCIoI will also be explainable, increasing the transparency and auditability of AI technology.

Computational, behavioral, and general biology: By studying intelligence as a complex phenomenon exhibited by biological agents, SCIoI naturally has strong overlaps to sub-fields of biology concerned with observing, analyzing, and modeling behavior. Experimental research within SCIoI employs methods from behavioral biology, psychology, and neuroscience, potentially adapting and advancing them for the purpose of investigating principles of intelligence.

The comparative approach, which will take center stage as a methodological tool in SCIoI, already plays a key role in evolutionary biology. Its scope in SCIoI extends beyond the typical usage in biology by explicitly including artificial “species” as objects of comparison, and by explicitly comparing the role of principles of intelligence across different scales in multi-level systems. And so, within SCIoI, Darwinian evolution is only one context for comparison, besides cultural evolution, or the diversity emerging from the technological innovation in AI.

Computational models of behavior from biology often serve as a starting point for corresponding modeling within SCIoI (see e.g., Klamser and Romanczuk (2021); Gomez-Nava et al. (2023)). However, SCIoI is mainly concerned with identifying mechanistic models of behavior at the right level of abstraction, capturing the regularities of relevance for intelligent behavior. Therefore, we are not concerned with statistical models as for example employed in bioinformatics (Ewens et al., 2005), and the level of description differs significantly from modeling approaches established in biology, such as those focusing on biomolecular (Mogilner et al., 2006) mechanisms (Hastings, 2013) or (adaptive) dynamics at the level of populations . Finally, based on the study of principles of intelligence, one aim of SCIoI is to propose new mechanistic models of behavior, which we expect will be highly relevant for disciplinary research in biology (Lang et al., 2023).

A central objective of SCIoI is the development of a new scientific community which performs across traditional disciplines, including biology. Through identifying and investigating general principles of intelligence and how they underlie the generation of behavior, SCIoI will provide the basis for testing the implementation of these principles in other biological systems like plants or bacteria that have been argued to exhibit aspects of intelligence (Jacob et al., 2004; Trewavas, 2003). In consequence, SCIoI research will establish new links to other sub-fields of biology not yet represented in SCIoI so far, such as plant biology or microbiology.

Complex systems, applied mathematics, and statistical physics: The field of complex systems spans many disciplines, including applied mathematics or statistical physics. It focuses on understanding mechanisms of self-organization and the emergence of complex structures and dynamical patterns from the interaction of many components or agents. It often relies on complex networks as a framework for describing such systems (Boccaletti et al., 2006).

The range of real-world systems motivating research on complex systems is quite broad. The discipline includes fundamental research into the origin of structure in physics and biology (Haken, 2006), the study of complex dynamics of social networks (Vega-Redondo, 2007), the scaling of cities (Bettencourt, 2013), and even cosmological patterns (Krioukov et al., 2012). Typically, it remains a theoretical science, primarily aimed at explaining dynamical patterns exhibited by different classes of complex systems, identifying universal laws or scaling relationships in these systems, and not aiming to uncover principles underlying the generation of adaptive, intelligent behavior in biological and artificial systems.

When analyzing and modeling intelligent, distributed, or collective systems, SCIoI relies on similar methods as complex systems research, such as adaptive network models (Poel et al.,
2022). Despite methodological overlaps, by bringing together experimental and theoretical researchers to uncover principles of intelligence and by investigating functional implications of the observed self-organized dynamics for intelligent behavior, SCIoI distinguishes itself from the larger, more diffuse field of complex systems.

Positioning with respect to research centers

The Center for Brains, Minds and Machines (CBMM) and the Quest for Intelligence at the Massachusetts Institute of Technology (MIT) are networks of researchers across analytic and synthetic disciplines of intelligence research. While they provide a platform to interact across disciplines, they do not define a shared research strategy or specific joint projects. The focus on neuroscience, in contrast to SCIoI’s focus on behavior and mechanistic characterizations of intelligence, represents fundamental methodological differences from SCIoI.

Similarly, the Leverhulme Centre for the Future of Intelligence (LCFI) is a center connecting interdisciplinary researchers within the University of Cambridge. The activities within the center address the challenges and opportunities of AI. There is less emphasis on intelligence as a biological phenomenon. However, the research thread Kinds of Intelligence raises questions also addressed by SCIoI. The research approach of LCFI is focused on cognitive sciences and philosophy. In contrast to SCIoI, they do not target a mechanistic understanding of intelligence and lack the synthetic approach.

The Santa Fe Institute (SFI) has several research themes related to SCIoI’s approach, including Complex Intelligence and Invention & Innovation. However, the specific research foci differ from SCIoI. SFI focuses on human social and hybrid human-AI systems on a larger scale and on a theory-minded, complex systems approach to intelligence and computation. Further, SFI’s structure is quite different from that of SCIoI. It hosts a small number of core faculty, and a larger number of external faculty, that form a rather loose community.

The Alan Turing Institute performs research in data science, AI, machine learning, and data engineering. It focuses on applications in the public sector (national security, environment, health). Its agenda is not directed toward a basic science of intelligence. Biological intelligence and the analytic approach are not represented.

The Berlin Institute for the Foundations of Learning and Data (BIFOLD) is dedicated to research in big data management and machine learning. The focus of BIFOLD is to develop these technologies for application in industry and the natural sciences. Its target is not a basic science of intelligence itself, and it lacks the consideration of biological intelligence and the analytic approach. BIFOLD is co-located with SCIoI in Berlin and complementary in its approach to AI. SCIoI cooperates with BIFOLD.

The Machine Learning for Science Cluster at Universität Tübingen focuses on applying machine learning methods to other sciences, such as physics, medicine, chemistry, and climate research. Its goal is therefore not a basic science of intelligence itself. Biological intelligence and the analytic approach are not represented.

Hessian.AI focuses on modern AI technologies (the “Third Wave of AI”), basic AI research, its transfer to industry, and its impact on society. The target of Hessian.AI is not the basic science of intelligence itself; it lacks considerations of biological intelligence and the analytic approach.