In this paper, we examine the relationship between subjective experience, self-awareness and scientific knowledge through the lens of cognitive science and the evolution of causal thinking. We present a developmental (historical) account of how scientific knowledge emerges from two alienation processes: the organism’s differentiation from its environment (enabling the manifest image) and from its own body (enabling self-awareness and the scientific image). By exploring these evolutionary dynamics, we build on Sellars’s integrative vision, suggesting that the scientific image, rather than replacing the manifest image, is fundamentally dependent on it. This approach sheds light on enduring philosophical concerns about agency, free will, and the tension between subjective experience and scientific objectivity.

Is the human mind a similarity machine or a probability engine? Perhaps it’s both – but if so, how do these mechanisms work together? These are perennial questions from the cognitive psychology of rationality. The influential heuristics and biases program treated similarity-based reasoning as an efficient, but ultimately irrational shortcut for probabilistic, or Bayesian reasoning. By contrast, I argue that similarity and probability are complementary principles of reasoning, drawing on the relationship between Conceptual Spaces (CS), a similarity-based framework for conceptual content, and Predictive Processing (PP), a Bayesian account of cognition. While CS and PP may operate at different Marr-levels, I focus on recent advances that show how CS can generate priors for Bayesian reasoning, offering a concrete mechanism for integrating similarity and probability. This approach resists reductionist interpretations, such as Tenenbaum and Griffiths’ account of perceptual categorization, by emphasizing the semantic interpretability that similarity provides for Bayesian models of reasoning.

Benchmarking, the evaluation of machine learning (ML) models based on predictive performance and competitive ranking, is a cornerstone of ML research and an increasingly prominent tool in scientific arguments. This paper argues that benchmarking constitutes a scientific epistemology, offering a powerful framework for scientific inference. We identify four core types of inferences drawn from benchmarks: those about the best (1) model, (2) learning algorithm, (3) deployment decision, and (4) prediction. We demonstrate that the validity of each of these inference relies on additional assumptions, analogous to ensuring construct validity in psychological tests. Through case studies in image recognition, life outcomes prediction, and weather forecasting, we examine these assumptions and their implications for inference validity. Finally, we discuss the social roles of benchmarks in organizing scientific communities and their potential threats to validity, offering strategies to mitigate these challenges and improve benchmark design and interpretation.

This paper critically examines Shagrir’s semantic view of computation according to which representational contents play an indispensable individuative role for vehicles of computation. I show that attempting to use representational contents to individuate vehicles of computation may lead to incoherent computational descriptions. I use an illustrative case-study – locust looming-object avoidance behaviour. A crucial component of the mechanism underlying this phenomenon is the lobula giant motion detector (LGMD) neuron, which multiplies angular velocity of an approaching stimulus with the negative angular size of the same. I show that interpreting the computation performed by the LGMD in semantic terms is exceedingly difficult. The representational contents of the outputs are not related to the representational contents of the inputs by the mathematical operation proposed by the model. The representational contents of the outputs do not individuate the vehicles of computation – these seem to be determined by the prevailing disciplinary norms (membrane voltages, instant firing rates). Furthermore, the representational content ascribed to one of the inputs seems to be unavailable for at least one of the functionally relevant class of stimuli (disconnected sequentially appearing dots). Thus, the utility, or even practicability, of semantically individuating the computation performed by the LGMD is questionable.

Philosophers and cognitive scientists inspired by the work of Vygotksy (1986) typically hold that complex thought results from the internalization of social interactions in development. More specifically, they propose that internalized speech facilitates various cognitive processes, including meta-cognition, reasoning and planning (Frankish, 2018; Kompa & Müller, 2022; Gauker, 2018). However, what exactly is internalized has not been spelled out in detail. In the literature on inner speech, it is usually held that what is internalized is a kind of dialogical structure (e.g. Fernyhough, 1996). In order to get a better handle on the structures involved, I present some ideas on discourse structure (Sanders et al., 1992; Thompson & Mann, 1987), and show how they can be applied to naturalistic data of overt private speech (Nelson, 1989). I will argue that these ideas allow us to describe 1) inner speech structure above the level of individual speech acts and 2) the function of individual inner speech utterances as part of a larger train of thought.

Could the genesis of artificial intelligence be the exodus of natural intelligence? Our current effort of engineering ever more capable artifacts mainly evokes concerns about the nature of artificial intelligence. But perhaps more pressing is what that reveals about the nature of our natural intelligence. We seem inclined and incentivized to utilize intelligent tools that solve our cognitive tasks for us. This is made possible by developing technologies that can receive parts of our cognitive skills which we would usually exert internally for those tasks. Such interaction strategies may profoundly decrease our own cognitive engagements and responsibilities, ultimately rendering us extracted cognizers. The hypothesis of extracted cognition states that we have a tendency to seek external artifacts that solve our cognitive tasks rather independent of us, namely by making or letting them capture, mimic, and eventually replace those cognitive skills we would otherwise employ and train internally. Three questions shall lead us to this idea: First, how do we make intelligent artifacts? Second, how do we use intelligent artifacts? Third, how do we thereby become extracted cognizers?