Beating hallucination: citations, verification and world models compared
Perplexity grounds answers in academic-style citations, Cursor looks to bug finding and formal verification, and Yann LeCun argues hallucinations come from LLMs' long tail and need world models to fix. Comparing these approaches shows readers how different builders try to make AI output trustworthy.
What three approaches to AI hallucination does the article compare? Show answer
Grounding answers in cited sources, verifying outputs after they are produced, and redesigning models so they plan with a model of the world.
How does Perplexity try to keep answers truthful? Show answer
It searches, extracts relevant paragraphs and has the model write an answer with footnotes, forbidding claims that were not retrieved.
What four causes of hallucination does Aravind Srinivas still see in a grounded system? Show answer
The model misunderstanding, stale or thin snippets, too much irrelevant detail, and irrelevant retrieved documents.
What is the main weakness of relying on citations? Show answer
A citation shows only that a source said something. Sources can be wrong, outdated or deliberately manipulated.
What is Cursor's shadow workspace? Show answer
A hidden copy of the editor where AI edits code and gets feedback from programming tools, such as type errors, without touching the user's files.
When does machine verification work well, according to the Cursor team? Show answer
When checking an answer is easier than generating it, as with code that can be compiled, tested or proved.
What objection do Cursor's own founders raise to replacing tests with proofs? Show answer
Intent is often hard to state formally, and side effects and outside services are hard to cover.
Why does Yann LeCun say hallucination is built into current models? Show answer
They generate one token at a time, each with some chance of error, so mistakes compound and rare prompts are never fully covered.
What does LeCun propose instead? Show answer
Systems that learn models of the world from observation and plan an answer in an abstract space before converting it to text.
Why is fluency a hazard? Show answer
People assume fluent language signals understanding, so errors that sound right are easily believed.
Where do Srinivas and LeCun fundamentally disagree? Show answer
Srinivas treats hallucination largely as a fixable pipeline problem. LeCun treats it as a flaw in how the models generate text.
Which kinds of answers deserve the most skepticism? Show answer
Long, fluent, free-form answers to rare or oddly phrased questions, where errors compound and training coverage runs out.