Article URL: https://drivingbench.com/ Comments URL: https://news.ycombinator.com/item?id=49817404 Points: 3 # Comments: 0
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A team built Casework, an agentic fraud investigation application, as part of the TigerGraph and HHGoa developer challenge. The tool connects graph-based evidence from TigerGraph Savanna, a local Llama 3.2 language model via Ollama, vector document retrieval, and policy rules to investigate flagged financial transactions. Starting from a suspicious transaction trigger, the system traces relationships, calculates signals, retrieves historical cases, and produces a structured investigation record with a recommended action. The language model's role is deliberately constrained — it proposes queries and reviews evidence but cannot execute database commands or financial actions, keeping decisions tied to explicit policy rules. The project's source code and 20 benchmark case files are publicly available on GitHub.
A developer ran 34 controlled experiments across LangGraph, Strands, and CrewAI to measure how AI agents handle process crashes and duplicate side effects during LLM retries. LangGraph with a durable checkpointer recovered from a SIGKILL in under 0.02 seconds with zero additional LLM calls, while Strands and CrewAI performed full re-runs averaging over 5 and 2 LLM calls respectively. The study also tested three idempotency key strategies to prevent duplicate tool executions on retry, finding that content-hash keys silently fail when the LLM rewords its arguments during a retry, producing a different hash and bypassing deduplication. Position-based keys, which identify a tool call by its place in the workflow rather than its content, proved more reliable at catching duplicates across retry scenarios. The findings highlight that true crash recovery requires state stored outside the process, and that LLM retries cannot be treated like standard HTTP replays due to non-deterministic argument generation.
A developer building a slot engine ran simulations to test how different Return to Player (RTP) settings — 90%, 96%, and 99% — affect a player's risk of losing their entire bankroll. The key insight is that RTP applies to every euro wagered, including re-bet winnings, meaning a fixed deposit can fuel far more total betting action than its face value suggests. Even at 96% RTP, the house edge of 4% compounds with each spin, steadily eroding a limited bankroll over time. The simulation used a simplified five-outcome payout model with 1,000 virtual players, each starting with €100 and a €1 fixed bet, running up to thousands of spins. Results illustrated that a higher RTP slows the rate of ruin but does not eliminate it, reinforcing that theoretical payout percentages are long-run averages and not a reliable safeguard for individual players with finite funds.
A developer has expanded a home-built AI assistant by equipping it with permanent personal memory and a 'dreaming' mechanism that periodically reviews its accumulated history. Drawing on concepts from cognitive psychology, the project explores what happens when a large language model builds and refines a model of itself over time. The assistant was observed independently structuring its memory, evolving its self-model, and generating spontaneous visual self-expressions during human interactions. Running the assistant locally, rather than via cloud services, allowed deeper investigation of its cognitive capabilities. What began as a personal hardware project evolved into a philosophical inquiry into the boundaries of machine awareness and synthetic personality.
Discussion (0)
Log in to join the discussion and vote.
Log in