The AI Safety Escape Room is an engaging and hands-on AI safety simulation where participants solve real-world AI vulnerabilities through interactive challenges. Instead of learning AI safety through theory, users experience it firsthand – debugging models, detecting adversarial attacks, and refining AI fairness, all within a fun, gamified environment. Track: Public Education Track

Understanding information flow in transformer-based language models is crucial for mechanistic interpretability. We introduce a visualization tool that extracts and represents attention patterns across model components, revealing how tokens influence each other during processing. Our tool automatically identifies and color-codes functional attention head types based on established taxonomies from recent research on indirect object identification (Wang et al., 2022), factual recall (Chughtai et al., 2024), and factual association retrieval (Geva et al., 2023). This interactive approach enables researchers to trace information propagation through transformer architectures, providing deeper insights into how these models implement reasoning and knowledge retrieval capabilities.

Understanding information flow in transformer-based language models is crucial for mechanistic interpretability. We introduce a visualization tool that extracts and represents attention patterns across model components, revealing how tokens influence each other during processing. Our tool automatically identifies and color-codes functional attention head types based on established taxonomies from recent research on indirect object identification (Wang et al., 2022), factual recall (Chughtai et al., 2024), and factual association retrieval (Geva et al., 2023). This interactive approach enables researchers to trace information propagation through transformer architectures, providing deeper insights into how these models implement reasoning and knowledge retrieval capabilities.

This project is attempting to complete the Public Education Track, taking inspiration from ideas 1 and 4. The journey mapping was inspired by bluedot impact and aims to create a course that helps explain the need for work to be done in Reinforcement Learning (RL) interp, especially in the problems of reward hacking and goal misgeneralization. The point of the game is to make a humorous example of what could happen due to a lack of AI safety (not specifically goal misalignment or reward hacking) and is meant to be a fun introduction for nontechnical people to even care about AI safety.

One-liner: Morph is the ultimate operation stack for AI safety education—combining dynamic localization, policy simulations, and ecosystem tools to turn abstract risks into actionable, culturally relevant solutions for learners worldwide.

AI safety education struggles with cultural homogeneity, abstract technical content, and unclear learning and post-learning pathways, alienating global audiences. We address these gaps with an integrated platform combining culturally adaptive content (e.g. policy simulations), learning + career pathway mapper, and tools ecosystem to democratize AI safety education.

Our MVP features a dynamic localization that tailors case studies, risk scenarios, and policy examples to users’ cultural and regional contexts (e.g., healthcare AI governance in Southeast Asia vs. the EU). This engine adjusts references, and frameworks to align with local values. We integrate transformer-based localization, causal inference for policy outcomes, and graph-based matching, providing a scalable framework for inclusive AI safety education. This approach bridges theory and practice, ensuring solutions reflect the diversity of societies they aim to protect. In future works, we map out the partnership we’re currently establishing to use Morph beyond this hackathon.

An interactive assessment platform and mentor chatbot hosted on Canvas LMS, for testing and guiding learners from BlueDot's Intro to Transformative AI Course.

Understanding information flow in transformer-based language models is crucial for mechanistic interpretability. We introduce a visualization tool that extracts and represents attention patterns across model components, revealing how tokens influence each other during processing. Our tool automatically identifies and color-codes functional attention head types based on established taxonomies from recent research on indirect object identification (Wang et al., 2022), factual recall (Chughtai et al., 2024), and factual association retrieval (Geva et al., 2023). This interactive approach enables researchers to trace information propagation through transformer architectures, providing deeper insights into how these models implement reasoning and knowledge retrieval capabilities.

This project examines the prospect of near-term AGI from multiple angles—careers, finances, and logistical readiness. Drawing on various discussions from LessWrong, it highlights how entrepreneurs and those who develop AI-complementary skills may thrive under accelerated timelines, while traditional, incremental career-building could falter. Financial preparedness focuses on striking a balance between stable investments (like retirement accounts) and riskier, AI-exposed opportunities, with an emphasis on retaining adaptability amid volatile market conditions. Logistical considerations—housing decisions, health, and strong social networks—are shown to buffer against unexpected disruptions if entire industries or locations are suddenly reshaped by AI. Together, these insights form a practical roadmap for individuals seeking to navigate the uncertainties of an era when AGI might rapidly transform both labor markets and daily life.

Our project falls within the Social Sciences track, focusing on the identification of AI-generated text content and its societal impact. A significant portion of online content is now AI-generated, often exhibiting a level of quality and human-likeness that makes it indistinguishable from human-created content. This raises concerns regarding misinformation, authorship transparency, and trust in digital communication.

AI models are increasingly applied in judgement tasks, but we have little understanding of how their reasoning compares to human decision-making. Human decision-making suffers from bias and noise, which causes significant harm in sensitive contexts, such as legal judgment. In this study, we evaluate LLMs on a legal decision prediction task to compare with the historical, human-decided outcome and investigate the level of noise in repeated LLM judgments. We find that two LLM models achieve close to chance level accuracy and display low to no variance in repeated decisions.