Traduttore Traditore? LLM Language-Dependent Safety Answers in Community Context

Funding request to extend a promising research based on a pilot conducted during Apart Research’s Global South Hackathon

Token taxes as a mechanism for reducing AI-driven power concentration.

A policy memo, co-authored with the Institute for Public Policy Research, resolving the open technical, economic, and legal questions blocking real-world implem

Originality, Quantification, and Decomposition of Generalization.

Perform research to rigorously elucidate and quantify generalization versus memorization, and examine evidence of originality in LLMS.

Empirical econ - applied AI transition for quantitative econometrician

Fund demonstrated/rigorous quantitative researcher (already run reproduction/audit pipelines on published economics) for 6-month AI safety transition, shipping

AI Safety Nepal

Five-week AI safety fellowship + 3-month project mentorship

What Factors Influence Chain-of-Thought Faithfulness?

Measuring whether CoT monitoring fails when an influence reaches an agent through a tool return rather than the user message.

Creating a better way to compile to transformer architectures

Detecting LLMs hallucinations without a second model

A simple, cheap methods to know when an LLM answer might be wrong, whether you run models in production or just ask Claude or ChatGPT at home

Shutdown evaluations for AI agents

KernelBench: an independent benchmark for AI-automated GPU kernel engineering

Frontier-model sweeps, hardware-roofline scoring, and reward-hacking audits at kernelbench.com. Keeping an independent AI R&D automation benchmark alive.

Postdoc on verification protocols for frontier AI training

Low-overhead zero-knowledge proofs of properties of training

Theoretical research of inductive reasoning building on LTA

Concept control in transformers with Sparse Concept Anchoring

Develop a training-time method for transformers that puts concepts where you can find them, so removal has predictable efficacy and bounded side-effects.

Testing failure modes of debate-style AI control schemes

Proposed Action Auditor (PAA)

Building a local, low-latency security runtime that monitors token logprobs to block unaligned agent tool-calling trajectories before execution.

Emotional AI safety: personalised affective guardrails for companion AI

A run-level reproducibility audit of LLM safety-eval benchmarks

Testing whether published safety evals replicate run-to-run, and generalizing the audit method across benchmarks.

A neutral and unbiased AI governance and safety system for the world

AI safety which operates on both model/agent and global level - a unique solution as far as we know.

Mechanistic estimation of highly out-of-distribution behaviors

Debate training on LLMs as a reward-hacking mitigation