Meta introduces an "organizational second brain" architecture that captures and compounds expert knowledge through a decoupled knowledge layer and self-improvement loop, eliminating the need for model retraining.
Ai2 shares lessons from building Shippy, a maritime AI agent, emphasizing that in high-stakes domains, reliability matters more than the model itself, and demonstrating an architecture of 'soul, skills, config' for verifiable, maintainable agent systems.
Research reveals LLMs rely on text style rather than tags to distinguish instructions; destyling drops injection success rates from 61% to 10%.
The key to scaling enterprise AI isn't better prompts or larger models, but 智能体 Logic: using deterministic software engineering primitives to constrain and steer LLMs for reliable, cost-effective execution.
Large language models for optical character recognition lower error rates but produce stealthier hallucinations instead of obvious garbled text, rendering legacy validation tools obsolete and demanding new evaluation paradigms and system architectures.