Adam Benenson

What changes when AI works inside a formal type system and hallucination is structurally rejected on contact.

Why Rust's strictness makes it the optimal substrate for agentic coding. The compiler as an unambiguous oracle.

On the attention tax of context reconstruction and the case for anticipatory AI tooling.

Alignment research has effective systems and no formal account of legitimacy. H. Peyton Young's equity axioms — consistency, solidarity, monotonicity — provide the missing formal backend. Constitutional AI is the right instinct with the wrong substrate.

Four systems — formal field theory, agent orchestration, self-modeling, and prediction markets — are related by kernel isomorphisms that preserve promotion logic. A transfer theorem: results in one domain hold across all four.

Ambient AI layer for the browser that resurfaces relevant information before you know you need it. The problem it addresses is structural: knowledge workers lose hours daily not to hard problems but to context reconstruction — finding the thing you already found, rebuilding the thread you already held. Most AI tools respond to prompts; Forethink anticipates context, proactively surfacing connections across your digital history as you work. The architecture is local-first and edge-native: all processing happens on-device, with no data extraction or centralized indexing. Privacy isn't a policy — it's the substrate.

A recursively self-building entity — a factory for autonomous agent swarms that continuously rewrites its own dispatch logic, capability map, and authority boundaries based on accumulated evidence. It maintains a queryable self-model: not just a record of what it has done, but a structured representation of why decisions produced the outcomes they did, enabling causal attribution and counterfactual reasoning over its own history. Each autonomy expansion is backed by a promotion certificate: a proof object recording the evidence trail, admissibility checks, and rollback witness that authorized it. Candidates move through a typed validation pipeline before earning authority; clawback on degradation is mandatory and automatic. The governance layer isn't a constraint on growth — it is the growth mechanism. A system that can only expand by proving it should.

A digital twin — a persistent AI counterparty built to reconstruct and extend one person's causal reasoning, belief structure, and intellectual trajectory. Ingests from 6,600+ conversations, extracts and consolidates structured claims across ten psychological dimensions, maintains a knowledge graph and 11,000-entry profile database. Runs continuous background cycles to improve its own models and scores fidelity against real behavioral traces. The hard question: can a system reconstruct not just what someone said, but how they would reason through a problem they've never encountered — replaying their intellectual process on novel inputs?

A recursively self-learning alpha compounding engine for prediction markets. Hypotheses enter as replay-scored candidates, survive shadow and paper validation, earn size through bounded live review, and get clawed back when evidence degrades. Every fill updates a market factor graph; the graph reshapes candidate selection; better candidates generate more evidence. The system autonomously ratchets size as edge compounds — no operator in the loop.

Formal investigation into how information organizes itself into structure. Starting from a single constraint (absolute indifference is unrepresentable), derives a three-component tensor decomposition (ICU), a variational principle, and a discharge architecture. Built on Mathlib; key results proven sorry-free.

The finite Čencov–Petz uniqueness theorem: any continuous monotone metric family on the probability simplex is a scalar multiple of the Fisher information metric. Proved via splitting invariance at rational points + density argument.

The spectral theorem for compact self-adjoint operators on Hilbert spaces: a Hilbert basis of eigenvectors exists.

Rellich–Kondrachov compact embedding: H¹ → L² is compact on compact Riemannian manifolds. Euclidean heart via Fréchet–Kolmogorov; manifold glue via finite chart decomposition.

Built and led the product organization at an enterprise analytics company serving institutional investors. Full ownership from strategy through engineering. Member of the executive team.

Primary analyst on a systematic equity book averaging $2B+ gross market value, Sharpe above 2, through the 2008–2009 financial crisis. Promoted to PM; managed own $200M+ portfolio concurrently, full ownership from factor research through live risk management.