In their rich and thought-provoking review, Burkart et al. use impeccable scholarship to produce a heroic synthesis of multiple complex literatures. Their two main goals are to critically evaluate the question of whether general intelligence exists in nonhuman animals, and to evaluate the implications of general intelligence for current theories about the evolution of cognition. In our view, they accomplish the first goal extremely effectively, making a compelling argument that general intelligence is indeed widespread among animals. Regarding their second goal, they argue that existing data from vertebrates support the cultural intelligence hypothesis, which stresses the critical importance of social inputs during the ontogenetic construction of survival-relevant skills. However, the general intelligence explained by the cultural intelligence hypothesis is actually quite limited, so we must seek a more robust explanation for its evolution.

We believe that the cognitive buffer hypothesis (Allman et al. Reference Allman, McLaughlin and Hakeem1993; Deaner et al. Reference Deaner, Barton, van Schaik, Kappeler and Pereira2003; Sol Reference Sol2009a; Reference Sol, Dukas and Ratcliffe2009b; Lefebvre et al. Reference Lefebvre, Reader and Sol2013) offers a better alternative because it can account for phenomena the cultural intelligence hypothesis leaves unexplained. The cognitive buffer hypothesis posits that general intelligence is favored directly by natural selection to help animals cope with novel or unpredictable environments, where it enables individuals to exhibit flexible behavior, and thus find innovative solutions to problems threatening their survival and reproduction. In our view, Burkart et al. dismiss the cognitive buffer hypothesis prematurely. They argue that fundamental preconditions for the evolution of large brains include a slow life history and high survivorship, possible only in species not subject to unavoidable extrinsic mortality such as high predation pressure (van Schaik et al. Reference van Schaik, Isler and Burkart2012). However, much can be learned by considering apparent exceptions to “rules” like these, so we offer the octopus as one such exception.

Most octopuses are strictly solitary except when copulating, have very short lives, have countless predators, and produce thousands of offspring, most of which die. Nevertheless, they have some of the largest brains known among invertebrates (Hochner et al. Reference Hochner, Shormrat and Fiorito2006; Zullo & Hochner Reference Zullo and Hochner2011); they exhibit a great deal of curiosity about their environments (Montgomery Reference Montgomery2015); they recognize individual humans (Anderson et al. Reference Anderson, Mather, Monette and Zimsen2010); they exhibit pronounced individual differences (Sinn et al. Reference Sinn, Perrin, Mather and Anderson2001; Mather et al. Reference Mather, Leite and Battista2012); they use tools; and they play (Mather Reference Mather1994; Mather & Anderson Reference Mather and Anderson1999). Octopuses thus appear to exhibit a considerable amount of general intelligence without any opportunity whatsoever for social learning. Clearly, the cultural intelligence hypothesis cannot account for the general intelligence apparent in creatures like these.

Similarly, the cultural intelligence hypothesis offers little promise with respect to evolving general intelligence in machines. Computer scientists and robotic engineers have understood for decades that the embodiment of intelligent machines affects their ability to adapt and learn via feedback obtained during their interactions with the environment, mediated by sensors and activators (Brooks Reference Brooks1990; Reference Brooks1991; Sharkey & Ziemke Reference Sharkey and Ziemke1998; Goldman & de Vignemont Reference Goldman and de Vignemont2009). Most hypotheses forwarded to explain the evolution of intelligence in animals, including the cultural intelligence hypothesis, fail to address the question of how morphological traits outside of the nervous system might have shaped intelligence. In creatures such as octopuses and primates, mutations affecting nervous system structure or function, which might generate less-stereotyped and more-flexible behavior, are visible to selective forces in the environment because they can be embodied in the limbs. Thus, greater intelligence is likely to evolve in these animals than in those whose interactions with their environments are more highly constrained.

Roboticists have also realized that logic alone cannot generate much intelligent behavior in their machines, and that to achieve better performance, their robots must also want things. The skills discovered by evolutionary algorithms are diverse, and many such skills may occur within a single population of digital organisms, but individual agents are rarely motivated to acquire a large array of skills. As a result, most current evolutionary algorithms produce domain-specific intelligence in machines that rarely possess more than a small set of skills, and they are thus suited to performing only tasks that demand that particular skill set. Although an intrinsic motivation to explore the environment has been imitated in artificial agents via machine learning (Schmidhuber Reference Schmidhuber1991; Oudeyer et al. Reference Oudeyer, Kaplan and Hafner2007), the production of generalist learners within an evolutionary context remains highly problematic (Stanton & Clune Reference Stanton and Clune2016).

Any selection pressure that promotes behavioral diversity or flexibility within the organism's lifetime, including the ability to learn from experience, should theoretically result in enhanced general intelligence. Novel or changing environments should select for individuals who can learn as much as possible in their lifetimes, as suggested by the cognitive buffer hypothesis. Indeed, Stanton and Clune (Reference Stanton and Clune2016) recently developed an evolutionary algorithm that produces agents who explore their environments and acquire as many skills as possible within their lifetimes while also retaining their existing skills. This algorithm encourages evolution to select for curious agents motivated to interact with things in the environment that they do not yet understand, and engage in behaviors they have not yet mastered. This algorithm has two main components: a fitness function that rewards individuals for expressing as many unique behaviors as possible, and an intra-life novelty score that quantifies the types of behaviors rewarded by the algorithm. Agents are also provided with an intra-life novelty compass that indicates which behaviors are considered novel within the environment. The intra-life novelty compass may simply identify and direct agents toward areas of high expected learning because new knowledge often promotes the ability to perform new skills. Aligned with these results, we suggest that the primary value of the cultural intelligence hypothesis is to offer social learning as an intra-life novelty compass, but that this hypothesis provides neither the requisite fitness function nor anything analogous to an intra-life novelty score.

A viable hypothesis explaining the evolution of large brains and general intelligence should be able to account for general intelligence in any species where it is known to occur, and it should be able to predict the conditions under which we can develop machines with general intelligence as well. The cultural intelligence hypothesis simply cannot do these things.