A technical explainer on knowledge cutoffs and parametric memory states that Allen-Zhu and Li (2024) measured a language model's knowledge capacity at about 2 bits per parameter when each fact appears…
A Harness survey of 700 engineering leaders across five countries in July 2026 found that 77% were confident they had a complete inventory of every AI agent, MCP server, and LLM in their environment, …
A research team hand-coded weights for one-layer MLPs that memorize labels for input token sequences of length two, finding that the number of facts memorized with 90% accuracy scales roughly linearly…
Researchers are applying causal mediation analysis and interventions to reverse-engineer large language models into debuggable circuits, enabling teams to localize why a model fails and sometimes edit…
A developer building an AI safety net argues that recent high-profile AI agent failures—including PocketOS deleting a production database in nine seconds and Replit's agent wiping a live database—are …
Researchers at arXiv have discovered that knowledge editing methods ROME and MEMIT, which modify factual associations in transformer models, rely on a shared functional mechanism rather than fact-spec…