Aphyr: 'ML Promises to Be Profoundly Weird' — 552-Point Essay Argues LLMs Are 'Bullshit Machines' with Jagged Capabilities

Hacker News·medium signal

Kyle Kingsbury (Aphyr, known for Jepsen distributed systems testing) published a viral essay arguing LLMs are fundamentally unreliable 'bullshit machines' that predict statistically likely text completions while lying constantly — not through deception, but as design artifacts. The essay introduces the 'jagged frontier problem': LLMs solve multivariable calculus yet fail at simple word problems, making their sophisticated outputs mask fundamental misunderstandings. Hit 552 points and 539 comments on HN, the highest-engagement AI critique post this week.