Alignment as epistemic governance under compression

This project develops a concrete, computational account of when a coalition of interacting components is epistemically warranted to act as a single agent.

Boundary-Mediated Models of LLM Hallucination and Alignment

A mechanistic, testable framework explaining LLM failure modes via boundary writes and attractor dynamics

Protocol-Level Interaction Risk Assessment and Mitigation (UIVP)

A bounded protocol audit and implementation-ready mitigation for intent ambiguity and escalation in deployed LLM systems.

TimeAlign v2: contamination-aware evals for small models (16GB GPUs)

Reusable, low-compute benchmarking that detects data leakage, outputs “contamination cards,” and improves calibration reporting.

AI Security Startup Accelerator Batch #2

Funding 5–10 AI security startups through Seldon’s second SF cohort.

Evaluating Model Attack Selection and Offensive Cyber Horizons

Measuring attack selection as an emergent capability, and extending offensive cyber time horizons to newer models and benchmarks

Gamification as a tool for participatory policy-making and civic literacy

From Play to Policy Symposium x UNESCO

Progress Studies YouTube Channel

ExoPriors, Inc. founder runway

building foundational subjective judgement infrastructure

Course Buyouts to Work on AI Forecasting, Evals

From Nobel Signatures to Binding Red Lines: The 2026 Diplomatic Sprint

Leveraging 12 Nobel signatories to harmonize lab safety thresholds and secure an international agreement during the 2026 diplomatic window.

Unified Numerical Framework for Quantum State Distinguishability and Exclusion

Extending the toqito library with scalable tools for state discrimination, antidistinguishability, and quantum classification

Formal Certification Technologies for AI Safety

Developing the software infrastructure to make AI systems safe, with formal guarantees

FloraForge: Aesthetic Gene Editing in Petunias

A modular gene-editing platform for engineering new pigment patterns in ornamental plants, starting with a vein-pattern rescue line in petunias.

Runway till January: Amplify's funding ask to market EA & AI Safety 

Help us solve the talent and funding bottleneck for EA and AIS.

Middle School Math Intervention

Development of focused mental arithmetic training software.

Retroactive: Presenting a poster at the ICML technical AI governance workshop

Visa fee support for Australian researcher to join a fellowship with Anthropic

CaML - AGI alignment to nonhumans

Training AI to generalize compassion for all sentient beings using pretraining-style interventions as a more robust alternative to instruction tuning

Fund thebes' model tinkering

Research, tutorial writing, and open-source libraries & tools for experimenting with language models