A newsletter breaking down AI research, technology, and Australian property in plain English
XSec, a new deep architecture accepted at GameSec 2026, achieves 97.33% average accuracy across five security scenarios while producing deterministic
RAG study tests LLaMA, Mistral and Qwen to cut hallucinations in small business AI, but the arXiv preprint provides no benchmark numbers.
New arXiv preprint proposes Chained RLM, calling the same model with fresh context to stop early reasoning errors reaching the final answer.
OpenAI's improved GPT-5.6 Sol promises better accuracy and consistency in ChatGPT, with expanded free access. Here's who benefits and what's missing.
A prespecified study found zero unauthorized tool calls across 840 GPT-5.6 agent trajectories, but raising reasoning effort changed inspection behaviour
AI agents fail on long tasks because of five cognitive gaps named in an August 2026 arXiv preprint proposing a new architecture.
Self-improving AI agents that learn from stored memory can inflate rewards for wrong answers, then preferentially reuse their most confident mistakes
An open-weight 35B model hits 64% on a new video reasoning benchmark, beating Claude 4.5 Sonnet and GPT-5, but the team graded its own homework.
XGBoost outperformed neural networks and SVM in a malware detection benchmark, but the 98.62% figure is self-reported and unverified.
