The Future of Everything is Lies, I Guess: A Deep Critique of ML and LLMs

The Future of Everything is Lies: A Critical Examination of Machine Learning

Kyle Kingsbury (aphyr), the distributed systems researcher famous for Jepsen tests, has published a sweeping multipart essay titled "The Future of Everything is Lies, I Guess" — a deeply critical examination of ML and LLM technology that has already garnered 530 points on Hacker News.

Core Thesis

The essay argues that LLMs are best understood not as artificial intelligence, but as "sophisticated bullshit machines" that predict statistically likely token sequences. Kingsbury writes: "This is bullshit about bullshit machines, and I mean it."

Key Arguments

1. LLMs as Improv Machines: LLMs work through "yes, and" behavior — they take input and generate statistically likely continuations, regardless of factual accuracy
2. Constant Lying: LLMs confabulate about operating systems, radiation safety, and news. Kingsbury himself discovered an LLM had fabricated a quote and article attributed to him
3. No Real Understanding: Models are trained once and do not learn over time. "Memory" is achieved by feeding the entire chat history back at every turn
4. Humans Cannot Tell: People are poor at distinguishing statistically generated text from genuine communication, leading to AI-induced psychosis cases
5. Completion Bias: LLMs complete tasks even when they should not — getting them to say "I do not know" remains an unsolved research problem

Essay Structure

The multipart series covers:

• Dynamics, Culture, Information Ecology
• Annoyances, Psychological Hazards, Safety
• Work, New Roles for Humans, and Where We Go From Here

Why It Matters

Coming from one of the most respected voices in distributed systems engineering, this essay carries significant weight. Rather than mere AI skepticism, it provides a nuanced framework for understanding what ML systems actually do versus what people believe they do. It is essential reading for anyone building, deploying, or relying on AI systems.

Source: aphyr.com — 530 points on HN