Why LLMs make things up, and three ways builders fix it
Stephen Wolfram, Yann LeCun, Marc Andreessen, Aravind Srinivas and Jimmy Wales explain hallucination as a consequence of how language models generate plausible text, and point to computation, retrieval with citations and human-checked sources as remedies. Knowing the cause helps users judge when an answer can be trusted.
Why do language models make things up? Show answer
They are built to continue text in a way typical of what people have written, one token at a time. No separate step checks the result against the world.
How does Stephen Wolfram contrast ChatGPT with computation? Show answer
He calls ChatGPT wide and shallow, a shallow computation over a large amount of training data. Computation is deep and can produce things never computed before.
Why does Yann LeCun say errors grow with the length of an answer? Show answer
Each token carries some chance of leaving the set of reasonable answers. If those errors are independent, the probability of staying correct shrinks exponentially with length.
Why do people believe made-up output, according to LeCun? Show answer
We are fooled by fluency. We assume fluent language implies human intelligence, and he says that impression is false.
What is the first fix, used by Wolfram? Show answer
Hand exact questions to a computational system. The model turns the request into precise code, a deterministic system computes the answer, and a person reads the code.
What is the second fix, used by Aravind Srinivas at Perplexity? Show answer
Retrieve documents with a search, feed relevant paragraphs to the model, and force it to cite everything. It should not say anything it did not retrieve.
What is the third fix, associated with Jimmy Wales? Show answer
Anchor answers in sources that people have checked. Wikipedia does this through verifiability and attribution, and a model answering only from those articles inherits that work.
In what four ways can retrieval still hallucinate, according to Srinivas? Show answer
The model misunderstands the query or paragraphs, the snippets are stale or thin, too much irrelevant detail confuses it, or the retrieved documents are irrelevant.
How does Marc Andreessen reframe hallucination? Show answer
He says it is called hallucination when we don't like it and creativity when we do. A law firm wants creative models to explore legal theories and then checks the citations separately.
Does LeCun think the three fixes cure hallucination? Show answer
No. He holds that the weakness comes from token-by-token generation and wants a different design that plans its answer before turning it into text.
What do all three fixes have in common? Show answer
Each moves the factual claim out of the generator, to a computed result, a retrieved passage or a human-checked source.
What should you do when an error would be costly? Show answer
Check the primary source or ask a qualified professional, as in legal, medical or financial matters.