Aphyr essay warns of LLM lies

The sharpest takeaway is that the danger is not sentient AI, but boring, pervasive unreliability becoming the new normal.

• –LLMs are already good enough to be embedded into core workflows before teams have a real handle on their failure modes.
• –The essay’s strongest argument is systemic: when false output is cheap, the burden of checking, tracing, and correcting shifts to people.
• –That makes provenance, refusal, and auditability more important than raw benchmark gains for developers shipping AI features.
• –The piece is opinionated, but it captures a real product risk that many teams still underweight: trust erosion compounds faster than capability improvements.