See our paper on arxiv

​​

Could Large Language Models feel pain or pleasure, or could they develop granular representations of such affect states? We developed a new behavioral approach, inspired by a paradigm from the animal sentience literature, to test the LLM behavior and decision-making beyond direct self-report. 

I was particularly interested in the extent to which a biologically-inspired behavioral approach could be adapted to testing LLMs, and whether one can plausibly claim a link to sentience candidacy via such tests. Our current conclusions are that 1) the LLMs are currently not sentience candidates, and 2) the experiments we conducted could lay a path toward a portfolio of research on artificial cognition and sentience, where several distinct lines of evidence would be required to support claims about the possibility of sentience in artificial systems.
 

This project was a collaboration between the Paradigms of Intelligence team at Google and Jonathan Birch's Foundations of Animal Sentience team.

preprint on PhilSciArchive

There has been much great recent work in philosophy on the questions of opacity and reliability of deep learning systems in science. My paper begins by taking their conclusions as a starting point, and asks "what follows from there for how we understand the question – what does it mean that something is known to science?".

​

One striking case of AI drastically transforming the production of scientific knowledge is AlphaFold, a system designed to predict protein structures. The problem of predicting how a protein's sequence of amino acids determines its three-dimensional structure—the protein folding problem—has historically been one of the most difficult and significant challenges in biological sciences.

My central question is how we can accommodate the AI-driven production of scientific claims under our conception of scientific knowledge. A situation where we have to say that we are doing cutting-edge science based on claims that are not scientific knowledge claims seems unattractive. On the other hand – it is difficult to accommodate the claims AlphaFold and similar systems produce under a traditional (empiricist & internalist) understanding of scientific practice and knowledge. 

​

I argue that AlphaFold can generate scientific knowledge and that scientific knowledge can be strongly opaque to humans, as long as it is properly functionally integrated into the collective scientific enterprise.

​

​

​​Portia spiders display sophisticated hunting strategies, leading to debate about the cognitive mechanisms behind their behavior. “Higher-cognition” explanations that appeal to working memory and mental simulation compete with “embodied-heuristics” explanations that aim to show how Portia’s embodiment and sensory apparatus simplifies the computations required. The evidence to date is indecisive: studies that may seem to shift the dial towards higher cognition, such as studies of numerosity, leave room for embodied heuristics.

I worked with Jonathan Birch to propose two lines of inquiry that could shift the dial: neurophysiological studies of brain mechanisms and comparative evidence from other arthropods that also use visual scanning.

A broader view we aim to advance is that intelligent behavior can be differently mechanistically realized, and in studying various mechanisms of animal intelligence we have a good reason to allow for such variations.

We find that it is likely for a general arthropod-style mechanism to allow for variation in intelligent problem-solving. In proposing a common “arthropod-style” cognitive ground-plan, we pursue a hypothesis that puts the core explanatory work on the mechanisms present in closely related lineages, and investigate the kinds of intelligence that can be achieved by those. 

​

My current project in progress is probing the concept of "agent-world dualism" in how building AI has traditionally been approached. I am looking into the accounts of reciprocal causation in biology and the work on major transitions in cognitive evolution to explain the nature of a properly embedded interaction between the agent and the world in the natural kinds of intelligent behavior.

I then aim to assess whether the strongest candidates for a resembling approach among the artificial systems today (e.g. Animal-AI Olympics but also the LLM-driven virtually embodied agents in 3D virtual environments) can overcome the dualism in question.