Bio
<aside>
💡 Mel Andrews is a philosopher of science and technology investigating the human use and misuse of mathematical and computational tools in science and societal decision-making. Mel earned a doctorate in philosophy in 2025 from the University of Cincinnati with a thesis titled ‘The Math & the Territory: On the Scientific Uses & Abuses of Machine Learning & Other Mathematical Modelling Strategies.’ Currently, Mel is a postdoctoral research associate jointly affiliated with the Center for Information Technology Policy (CITP) at Princeton University and the Natural and Artificial Minds Initiative (NAM) of the Princeton AI Lab. Mel’s also the principal investigator on a grant from the Alfred P. Sloan Foundation’s Metascience and AI funding scheme assessing the epistemic credentials of AI in science evaluation.
</aside>
- Would you still consider LLMs with inference-time compute to be a type of “stochastic parrot” the way Bender would put it?
- As a philosopher of science, what has studying contemporary AI and analysing LLMs told you about fields such as linguistics, CogSci, and PhilMi and other fundamental questions, such as the nature of language, PhilSci, etc.?
- What, if any, would you say has surprised you the most?
- Apropos Andrews et al. (2024) and Andrews (2025), I am curious about your views on the contemporary LLM benchmarks used to evaluate foundation models, including those like ARC-AGI?
- What is intuition laundering?

“This is what is happening when we try to take theories and models from the cognitive sciences and apply them to software systems.”
- What is the problem of model transfer PhilSci has recognised?
“These concepts [“love”, “sadness, “Consciousness” and “Understanding”], importantly, cannot be properly operationalised wholesale. There is no definitive scientific test of whether or not ‘my cat loves me,’ or whether or not ‘trees can think.’”
- This reminds me of the notion of thick concepts in ethics (Williams, 1985/2015 & Philippa Foot).
“For all of the above reasons, I do not think we are likely to ever see properly scientific and non-question-begging research into the question of whether or not language models “understand” language (or anything, for that matter).”
- Would it be great for you to elaborate on this point of why we cannot scientifically determine whether language models “understand”? I wonder if we can ever pose this question for any kind of machine intelligence?
- I, for instance, have no doubt you [Mel Andrews] are an intelligence agent, but if you tell me to justify that belief, I would eventually end up on a hinge proposition (Wittgenstein et al., 2008).
The Math & the Territory: On the Scientific Uses & Abuses of Machine Learning & Other Mathematical Modeling Strategies
etd.ohiolink.edu
- To start somewhat anecdotally, if you’re open to sharing, I’d love to hear how your thesis went in this direction of “the human use and misuse of mathematical and computational tools, in particular, machine learning (ML) based methods.” It is fascinating, as it seems like a beautiful bricolage and synthesis of all the areas you published on, from the FEP to PhilML and AI ethics.
- Worth discussing the positive vision of the theory-laden conception of ML in science that you put forward.
- Relatedly, what is the right way to argue for using an ML-based mathematical modelling framework in neuroscience: the free energy principle (FEP)?
- Worth discussing conceptual reification. I see it everywhere, but it is also rather difficult to argue against such reification. Sort of like saying there are Markov blankets in the brain.
- How did AlphaFold succeed because it was, in fact, heavily Theory-Laden?
- How important is the history of science to you? I love this talk by Hasok Chang: https://youtu.be/zGUsIf9qYw8?si=eCZCDxTp8LhPRM2b