André Curtis-Trudel

Large language models (LLMs) exhibit impressive performance across a range of apparently cognitive tasks. Mentalists holds that this performance is best explained by the fact that LLMs have mental states, while anti-mentalists hold that this performance should be explained some other way. In this note we address representationalist folk mentalism, which holds (a) possessing a folk mental state like belief or desire is a matter of having an internal representation with appropriate content and (b) that LLMs have folk psychological states of this sort (or at least robust precursors to such states). Although representationalist folk mentalism appears attractive, we argue that neither probing nor intervention studies have uncovered representations of the relevant sort in state-of-the-art LLMs. However, while it would be premature to accept representationalist folk mentalism, our argument provides a roadmap for mechanistic interpretability research going forward.

We advance a novel version of the No Miracles Argument (NMA), tailored explicitly for AI, on which predictive success provides support for the existence of hidden quasi-representations—entities that would count as representations if they embodied aboutness relations. Our primary claim is comparative: reframing the AI-specific NMA in terms of quasi-representation yields a weaker, and thereby more plausible, argument than our previous representation-based formulation, one that inherits whatever force no‑miracles reasoning has in the traditional case. An added advantage is that our new NMA is compatible with selective realism and anti-realism alike, because quasi‑representation can be cashed out in thin terms. To illuminate our approach, we consider how quasi-representations might underpin the predictive accuracy of GraphCast, a cutting-edge AI system used for global weather forecasting. We then defend our quasi-representational NMA against concerns of underspecification and triviality. Finally, we explore the consequences for explainable AI (XAI) if our NMA is sound.

A key premise driving the problem of unconceived alternatives is that contemporary scientists are no better than their predecessors at envisioning serious rivals to even the most well-confirmed scientific theories. Some realists reject this, arguing that present-day science is capable of more severe tests and more comprehensive searches of the space of theoretical alternatives than were previously possible. One way to support this response appeals to the fact that much contemporary science is computational, facilitated by large-scale digital computers. However, we argue that advanced computational tools and methods often impose a tradeoff between generating high-resolution predictions and performing lower-resolution scans of possibility space, which may exacerbate rather than attenuate concerns about unconceived alternatives. We illustrate by considering a specific, highly computational branch of contemporary science: gravitational-wave astronomy.

Unlimited pancomputationalism is the claim that every physical system implements every computational model simultaneously. Some philosophers argue that unlimited pancomputationalism renders implementation ‘trivial’ or ‘vacuous’, unsuitable for serious scientific work. A popular and natural reaction to this argument is to reject unlimited pancomputationalism. However, I argue that given certain assumptions about the nature of computational ascription, unlimited pancomputationalism does not entail that implementation is trivial. These assumptions concern the relativity and context sensitivity of computational ascription. Very roughly: relative to a specific, contextually salient way of regarding a physical system computationally, the claim that that system implements a specific computational model is as non-trivial as one could reasonably want.

According to the standard no miracles argument, science’s predictive success is best explained by the approximate truth of its theories. In contemporary science, however, machine learning systems, such as AlphaFold2, are also remarkably predictively successful. Thus, we might ask what best explains such successes. Might these AIs accurately represent critical aspects of their targets in the world? And if so, does a variant of the no miracles argument apply to these AIs? We argue for an affirmative answer to these questions. We conclude that if the standard no miracles argument is sound, an AI-specific no miracles argument is also sound.

What is computational explanation? Many accounts treat it as a kind of causal explanation. I argue against two more specific versions of this view, corresponding to two popular treatments of causal explanation. The first holds that computational explanation is mechanistic, while the second holds that it is interventionist. However, both overlook an important class of computational explanations, which I call limitative explanations. Limitative explanations explain why certain problems cannot be solved computationally, either in principle or in practice. I argue that limitative explanations are not plausibly understood in either mechanistic or interventionist terms. One upshot of this argument is that there are causal and non-causal kinds of computational explanation. I close the paper by suggesting that both are grounded in the notion of computational implementation.

A skeptical worry known as ‘the indeterminacy of computation’ animates much recent philosophical reflection on the computational identity of physical systems. On the one hand, computational explanation seems to require that physical computing systems fall under a single, unique computational description at a time. On the other, if a physical system falls under any computational description, it seems to fall under many simultaneously. Absent some principled reason to take just one of these descriptions in particular as relevant for computational explanation, widespread failure of computational explanation would appear to follow. This paper advances a new solution to the indeterminacy of computation. Very roughly, I argue that the computational identity of a physical system is determinate relative to a contextually specified way of regarding that system computationally—known as a labelling scheme. When a system simultaneously implements multiple computations, it does so relative to different labelling schemes. But relative to a fixed labelling scheme, a physical system has a unique computational identity. I argue that this relativistic conception of computational identity vindicates computational explanation in the face of simultaneous implementation.

This article advertises a new account of computational implementation. According to the resemblance account, implementation is a matter of resembling a computational architecture. The resemblance account departs from previous theories by denying that computational architectures are exhausted by their formal, mathematical features. Instead, they are taken to be permeated with causality, spatiotemporality, and other nonmathematical features. I argue that this approach comports well with computer scientific practice and offers a novel response to so-called triviality arguments.