Actually Effective Altruism

Inspiration

After reading the papers [1][2][3], we realised that there was a lot of potential for using LLMs and autonomous agents to visualize and look at possible chain of events spawning from initial scenarios. This has uncountable possible benefits to those suffering under consequences of corporate, govermental or even international politics

What it does

We made a AI agentic simulator to model the configuration and evolution of complex scenarios in which there exists deeply interconnected stakeholders and relationships.

How we built it

We used Svelte + TypeScript in the front end with Flask (Python) for backend to scrape pages on Google to place into a vector database for RAG. We used Sveltekit, SkeletonUI and other web technologies to build the UI which lets users tune the models/scenarios and see their outputs.

We then use a randomised beam search to explore the option space in a controlled and non-exponential manner, more closely resembling the natural human decision making process.

Challenges we ran into

The issue with removing bias from AI agent responses due to training data The limitations with AI to stay up to date with current news due to their information cut off Being able to parallelize workload between team members. ANTHROPIC

Accomplishments that we're proud of

Creating a highly dynamic user interface that intuitively allows the user to explore the decision space Eliminating the risk of LLM hallucination by using web search to vectorize facts in a RAG database Bidirectional graph visualizer to show relationships between states and the possible events that link them, then an interactive tree to model the intersection of multi-party decision trees.

What we learned

More about AI agents and LLM model limtations How to balance time constraints when working in a team More in depth understanding of web dev

What's next for Actually Effective Altruism

More sophisticated search algorithms with more grounding in both anthropological and technical research.

[1] = Political-LLM: Large Language Models in Political Science [2] = This Land is {Your, My} Land: Evaluating Geopolitical Bias in Language Models through Territorial Disputes [3] = Large Language Models Reflect the Ideology of their Creators

Built With

• ai
• machine-learning
• python
• rag
• svelte
• typescript