Thoughts
Modern AI systems are trained largely on observational data: patterns, correlations, and records of what has already happened. While it is familiar that correlation is not causation, less is known about what this implies for learning dynamics, intervention, and structural limits on generalization. This manuscript gives theorems characterizing fundamental limits of observational learning.
Most definitions of intelligence either collapse into task performance or become too vague to explain what intelligent systems are doing. This work asks what makes a system capable of improving across changing conditions, and argues that intelligence is better understood through the construction, criticism, and revision of explanations.
Modern AI systems are often evaluated by performance on observed tasks, but this can hide which errors are actually reachable, diagnosable, and correctable. This work asks whether progress toward general intelligence depends less on scale alone and more on the ability to expose, transform, and repair errors.
Causal models often describe variables and dependencies without giving a deep account of why some structures remain stable under intervention. This work develops Energy-Structured Causal Models as a way to connect causal explanation with constraints, locality, autonomy, and mechanism-level editability.
Thoughts
Modern states often treat policy as a contest of fixed programs rather than as a series of fallible conjectures that need criticism, measurement, revision, and appeal. This work explores institutional designs for a more corrigible civic system: one that preserves evaluation, exit, accountability, and lawful correction.
Healthcare financing struggles to combine universal access, price discipline, patient agency, and protection from catastrophic costs. TRUST-CARE proposes a credit-based health-financing mechanism designed to make care portable, repayable according to ability, and backstopped where ordinary repayment fails.
Thoughts
Immune recognition is often modeled through narrow proxies such as binding affinity, even though real immunogenicity depends on presentation, recognition, selection, and evolutionary context. This work asks how better representations of antigenicity and counterfactual selection could improve reasoning about tumors, viruses, and immune targets.
Vaccine design depends on knowing which viral regions are actually seen by T cells and whether those regions remain stable as the virus evolves. This study maps immunodominant SARS-CoV-2 T cell epitopes and examines their relationship to evolutionary conservation.