Mayank Kejriwal

Advocates of modern decision theory advocate for making decision- making explicit and quantifiable. The core prescription is that we should introspect our intuitions, assign numbers to our beliefs, and subject these to formal analysis. This approach promises clarity, reduced bias, and improved decision outcomes. Yet I argue that this prescription rests on a hidden and problematic assumption: that translating tacit knowledge into explicit form preserves the original knowledge. I contend that, in many cases, this translation does not merely represent our intuitions differently; rather, it flattens and transforms them in ways that can entail genuine knowledge loss. Drawing on classical epistemology—particularly Polanyi’s concept of tacit knowledge and Dreyfus’s critique of formalization—I argue that certain kinds of understanding cannot be adequately captured by explicit criteria and numbers without forfeiting essential dimensions. When we anchor ourselves to explicit representations, we risk losing touch with the original implicit source from which our judgment emerges. This paper explores this tension and argues for a more nuanced view: that the value of formalization must be weighed against the costs of epistemic transformation. I do not claim formalization is always harmful, but rather that the decision theory community should engage more carefully with the philosophical problem of knowledge translation.

Despite significant progress, the AI alignment literature has largely overlooked a philosophical dilemma: \textit{humans and machines represent and communicate uncertainty in fundamentally incompatible ways.} To formalize the arguments around this dilemma, we introduce a four-quadrant taxonomy of uncertainty that partitions uncertainty according to two dimensions: whether uncertainty arises from a human or machine agent, and whether it remains internal or is communicated externally. The taxonomy comprises \textit{human inherent uncertainty} (Bayesian credence), \textit{human self-reported uncertainty} (linguistic hedges), \textit{model inherent uncertainty} (frequentist risk), and \textit{model self-reported uncertainty} (preference-optimized language). We show that this taxonomy reveals an epistemological gap: humans assign uncertainty through grounded, causal understanding, while models assign it through risk-minimization over training data. When models trained with RLHF generate hedged language, humans interpret it as Bayesian belief vis-à-vis frequentist risk-penalty. We close with a practical case study and argue that bridging this epistemic gap is essential - but unfortunately neglected in current discourse - for establishing trust in complex human-AI collaboration modes.

How do abstract singular terms like "the Pythagoras Theorem" refer? While theories of reference have been extensively developed for concrete singular terms (proper names, definite descriptions), the question of reference to abstract objects remains under-explored. This essay argues that abstract singular terms do refer, obeying similar principles as concrete singular terms while presenting distinct philosophical challenges. We examine three prominent theories of reference—Millianism, the Frege-Russell Theory, and Kripke's causal theory—and demonstrate that each faces serious difficulties when extended to abstract entities. Millianism requires commitment to mathematical Platonism; descriptivism struggles with reference-fixing and the problem of degraded understanding; pure causal theory fails to account for variation in speakers' conceptions of abstract terms. We propose that a causal-descriptive hybrid theory, wherein causation plays a constitutive role in the sense of abstract singular terms, best explains how speakers can meaningfully refer to abstract objects despite incomplete or evolving understanding. This account illuminates how ordinary language users refer to mathematical theorems without explicit knowledge of their formal content, suggesting that causation is not merely a supplementary feature of reference but essential to understanding how abstract singular terms function in discourse.

In "Is Power-Seeking AI an Existential Risk?" Joe Carlsmith provides an intriguing analysis of the potential pathways from advanced artificial intelligence to existential catastrophe. This essay contends that while the report's framework offers a fundamental contribution to AI safety, its application of longtermist principles is incomplete. By focusing almost exclusively on catastrophic failure modes, it overlooks the symmetrical potential for "uber-beneficence" i.e., outcomes of profound and durable positive value. This critique deconstructs Carlsmith's model of the power-seeking AI, arguing that it relies on a monolithic conception of agency that does not sufficiently appreciate insights from recent research into diverse agent architectures and emergent goals. Ultimately, a more complete longtermist strategy must not only build guardrails against catastrophe but also actively chart a course toward a flourishing, AI-assisted future. An overriding focus on downside risk alone may inadvertently preclude the most positive outcomes humanity could achieve.