LLMs Are Still Toxic, Stuck in the Past, and Bad at Math
Summary
The piece argues four persistent limitations of LLMs—math accuracy, recency, memory, and toxicity—remain despite tooling progress. It explains how harnesses, tool-calling, retrieval-augmented generation (RAG), and reinforcement learning from human feedback (RLHF) with synthetic data and guardrails mitigate these issues, and it discusses long-running agents and context-management as the path forward. It emphasizes engineering around the model rather than fixing the weights alone for practical SMB AI work.