Boris Grozdanoff

Large Language Models (LLMs) operate within closed neural architectures, isolated from direct physical interaction with the external world. This lack of empirical grounding is a critical epistemic deficiency of empirical theories to explain the referential nature of LLM outputs and their hallucinatory fallibility. To address this, I propose a computational adaptation of Fregean semantics that integrates the semantics of transformer architectures with classical philosophy of language. In it I reject static token embeddings as the sole bearers of meaning in an LLM and I formally identify shifted, contextualized vectors generated via self-attention as rigorous computational analogues of Fregean sense. This structure provides an epistemological grounding for how transformers construct meaning, and then it demonstrates the new potential of ungrounded models to execute a priori revisions of deeply learned, highly probable “beliefs.” I integrate these contextualized Fregean senses into a Quinean-style web of beliefs,1 endowing the neural network with dual epistemic fallibilities: empirical and a priori. To demonstrate this mechanism, I model a paradigmatic historical shift, the a priori revision of the absolute simultaneity principle, as it would unfold within this modified Fregean vector space. I identify a multi-sense consistency mechanism which can correct errors a priori and I show that this approach exhibits a robust model capable of delivering epistemic success in artificial systems devoid of real-world grounding.

In one of the most influential empiricist account on the epistemic nature of thought experiments John Norton proposes a challenge: that no thought experirnent in science could be found that cannot be logically reconstructed as an argument. Norton’s account has two main theses, the epistemic thesis that all information about the physical world delivered through a thought experiment comes solely frorn experience and the reconstruction thesis that all thought experiments could be reconstructed as arguments. In the present paper I argue that in at least sorne cases Norton’s theses are incompatible with each other and therefore their combination could not form a reliable account. I try to show that sometimes the experience available could not justify the conclusion of a thought experiment and even contradicts it. I suggest an analysis of Einstein’s Train Thought Experiment as a counterexample to the challenge.

The paper critically examines an unpopular line of Frege’s view on numbers in the Foundations of Arithmetic. According to this view, which analyzes numbers in terms of properties and not in terms of extensions, numbers are properties of concepts vs. properties of objects. The latter view is held by Mill and is famously criticized in the Foundations. I argue that on the property account numbers cannot only be properties of concepts but they also have to be properties of objects. My main argument rests on purely metaphysical grounds. It stems from the motivation that were numbers only properties of concepts we would not have been able to explain mathematical truths about the physical world or those truths would have been miraculous. On pains that we do have mathematical truths about the physical world that are not miraculous we cannot agree with Frege’s property line about the metaphysical nature of numbers.