David Chalmers

Mechanistic interpretability is the program of explaining what AI systems are doing in terms of their internal mechanisms. I analyze some aspects of the program, along with setting out some concrete challenges and assessing progress to date. I argue for the importance of propositional interpretability, which involves interpreting a system’s mechanisms and behav- ior in terms of propositional attitudes: attitudes (such as belief, desire, or subjective probabil- ity) to propositions (e.g. the proposition that it is hot outside). Propositional attitudes are the central way that we interpret and explain human beings and they are likely to be central in AI too. A central challenge is what I call thought logging: creating systems that log all of the rel- evant propositional attitudes in an AI system over time. I examine currently popular methods of interpretability (such as probing, sparse auto-encoders, and chain of thought methods) as well as philosophical methods of interpretation (including those grounded in psychoseman- tics) to assess their strengths and weaknesses as methods of propositional interpretability.

This paper offers both a theoretical and an experimental perspective on the relationship between connectionist and Classical (symbol-processing) models. Firstly, a serious flaw in Fodor and Pylyshyn’s argument against connectionism is pointed out: if, in fact, a part of their argument is valid, then it establishes a conclusion quite different from that which they intend, a conclusion which is demonstrably false. The source of this flaw is traced to an underestimation of the differences between localist and distributed representation. It has been claimed that distributed representations cannot support systematic operations, or that if they can, then they will be mere implementations of traditional ideas. This paper presents experimental evidence against this conclusion: distributed representations can be used to support direct structure-sensitive operations, in a man- ner quite unlike the Classical approach. Finally, it is argued that even if Fodor and Pylyshyn’s argument that connectionist models of compositionality must be mere implementations were correct, then this would still not be a serious argument against connectionism as a theory of mind.

Consciousness fits uneasily into our conception of the natural world. On the most common conception of nature, the natural world is the physical world. But on the most common conception of consciousness, it is not easy to see how it could be part of the physical world. So it seems that to find a place for consciousness within the natural order, we must either revise our conception of consciousness, or revise our conception of nature. In twentieth-century philosophy, this dilemma is posed most acutely in C. D. Broad’s The Mind and its Place in Nature. The phenomena of mind, for Broad, are the phenomena of consciousness. The central problem is that of locating mind with respect to the physical world. Broad’s exhaustive discussion of the problem culminates in a taxonomy of seventeen different views of the mental-physical relation.1 On Broad’s taxonomy, a view might see the mental as nonexistent, as reducible, as emergent, or as a basic property of a substance. The physical might be seen in one of the same four ways. So a four-by-four matrix of views results. At the end, three views are left standing: those on which mentality is an emergent characteristic of either a physical substance or a neutral substance, where in the latter case, the physical might be either emergent or delusive.

Why is two-dimensional semantics important? One can think of it as the most recent act in a drama involving three of the central concepts of philosophy: meaning, reason, and modality. First, Kant linked reason and modality, by suggesting that what is necessary is knowable a priori, and vice versa. Second, Frege linked reason and meaning, by proposing an aspect of meaning (sense) that is constitutively tied to cognitive signi?cance. Third, Carnap linked meaning and modality, by proposing an aspect of meaning (intension) that is constitutively tied to possibility and necessity

There has been much interest in the possibility of connectionist models whose representations can be endowed with compositional structure, and a variety of such models have been proposed. These models typically use distributed representations that arise from the functional composition of constituent parts. Functional composition and decomposition alone, however, yield only an implementation of classical symbolic theories. This paper explores the possibility of moving beyond implementation by exploiting holistic structure-sensitive operations on distributed representations. An experiment is performed using Pollack’s Recursive Auto-Associative Memory. RAAM is used to construct distributed representations of syntactically structured sentences. A feed-forward network is then trained to operate directly on these representations, modeling syn- tactic transformations of the represented sentences. Successful training and generalization is obtained, demonstrating that the implicit structure present in these representations can be used for a kind of structure-sensitive processing unique to the connectionist domain.

This chapter describes three relatively specific forms such as destructive uploading, gradual uploading, and nondestructive uploading. Neuroscience is gradually discovering various neural correlates of consciousness, but this research program largely takes the existence of consciousness for granted. It presents an argument for the pessimistic view and an argument for the optimistic view, both of which run parallel to related arguments that can be given concerning teletransportation. Cryonic technology offers the possibility of preserving our brains in a low‐temperature state shortly after death, until such time as the technology is available to reactivate the brain or perhaps to upload the information in it. Reconstructive uploading from brain scans is closely akin to ordinary (nongradual) uploading from brain scans, with the main difference being the time delay, and perhaps the continued existence in the meantime of the original person.

Does the capacity to think require the capacity to sense? A lively debate on this topic runs throughout the history of philosophy and now animates discussions of artificial intelligence. Many have argued that AI systems such as large language models cannot think and understand if they lack sensory grounding. I argue that thought does not require sensory grounding: there can be pure thinkers who can think without any sensory capacities. As a result, the absence of sensory grounding does not entail that large language models cannot think or understand. I also consider to what extent quasi-sensory grounding can at least boost the performance of a language model.

This chapter provides a rich philosophical discussion of superintelligence, a widely discussed piece that has encouraged philosophers of mind to take transhumanism, mind uploading, and the singularity more seriously. It starts with the argument for a singularity: is there good reason to believe that there will be an intelligence explosion? Next, the chapter considers how to negotiate the singularity: if it is possible that there will be a singularity, how can we maximize the chances of a good outcome? Finally, it considers the place of humans in a post‐singularity world, with special attention to questions about uploading: can an uploaded human be conscious, and will uploading preserve personal identity? The chapter also analyzes whether there are certain constraints on design of artificial intelligence (AI) and AI+ that we might impose, in order to increase the chances of a good outcome.

[This is an edited version of a keynote talk at the conference on Neural Information Processing Systems (NeurIPS) on November 28, 2022, with some minor additions and subtractions.] There has recently been widespread discussion of whether large language models might be sentient or conscious. Should we take this idea seriously? I will break down the strongest reasons for and against. Given mainstream assumptions in the science of consciousness, there are significant obstacles to consciousness in current models: for example, their lack of recurrent processing, a global workspace, and unified agency. At the same time, it is quite possible that these obstacles will be overcome in the next decade or so. I conclude that while it is somewhat unlikely that current large language models are conscious, we should take seriously the possibility that successors to large language models may be conscious in the not-too-distant future.