AI Digest: Chatbots Make Mistakes in Financial Matters

The main developments over the past 24 hours — from issues with the reliability of AI responses to new methods for detecting hallucinations and working with long contexts.

🔥 Hot:

🔹 AI chatbots give incorrect answers to financial questions in most cases — This is especially critical for scenarios in which users rely on accuracy and make decisions based on the model’s response.

➡️ Useful resources:

🔹 Researchers propose detecting LLM hallucinations through the structure of information flows — The method analyzes attention graphs and looks for structural signs of bottlenecks in context transmission. 🔹 A new study separates changes in LLM representations from their causal importance after fine-tuning — The paper examines how fine-tuning changes attention and activations across layers, and whether these changes are related to the model’s behavior. 🔹 RBS-Attention reduces prefill costs for long-context LLMs — The method selects limited attention regions and accounts for the risk that block averaging may hide an important token.

➡️ Discussions and case studies:

🔹 A developer wrote an MCP server and asked an LLM to review its code — This practical experiment demonstrates a scenario in which the model serves as a reviewer of its own tooling. 🔹 LocalLLaMA users compare Ternary Bonsai 2 27B with compact local models — The discussion is aimed at users with limited GPU resources and specifically takes KV-cache size into account.

📝 If you would like to add other news and resources to the list, leave a comment.