A researcher explores whether gzip, a standard compression utility, can perform language modeling by leveraging the mathematical equivalence between compression and prediction. Using beam search over byte sequences scored by gzip's compression length, they demonstrate that gzip can generate text continuations that show understanding of source material, despite producing imperfect output.
A new approach uses personal computer use data to train local LLMs that predict user judgment and writing patterns, reducing the effort required to prompt AI agents. In a two-week study, a specialized model achieved 17.1% semantic accuracy on next-write predictions at $0.3 per call, with a continually trained version reaching 3.0% accuracy at $0.01 per call, suggesting potential for scaling.
Jev is a machine learning tool-call classifier that monitors agentic AI executions. While useful for measuring progress and reducing time/spend prediction errors by ~18-19%, it struggles with security (easily bypassed) and detecting lazy agents, making it best suited for observability rather than threat detection.
An article describes building a machine learning model to predict Survivor winners and eliminations. The model uses logistic regression trained on historical data from 50 seasons, analyzing features like confessional time, votes received, advantages, and age to forecast each player's probability of winning or being voted out next.
HeronLoom is a tool that transforms social media posts and documents into a 3D graph for analyzing trends, sentiment, and predictions using iterative clustering and LLM-based analysis. It tracks how topics and narratives evolve over time while preserving conflicting interpretations and maintaining evidence-based confidence scores with citations to source material.
Large Language Models are stateless prediction machines that lack memory between API calls. To create the illusion of conversation, the entire message history is sent with each API request in a messages array, where each message is labeled with a role (user, assistant, or system) to provide context the model cannot infer on its own.