One question that Geoffrey Hinton, the Godfather of AI says OpenAI's GPT-4 still gets wrong

Geoffrey Hinton , the Godfather of AI has admitted that he sometimes trusts his favorite chatbot more than he should. In a recent interview with CBS, Hinton revealed that he uses OpenAI 's GPT-4 and he tends to believe what GPT-4 tells him, even though he knows he should be more careful. “I tend to believe what it says, even though I should probably be suspicious,” he said. Sharing an example from his own experience, he said that he once asked GPT-4 a simple riddle:

"Sally has three brothers. Each of her brothers has two sisters. How many sisters does Sally have?"

The correct answer is one as Sally herself is one of the two sisters, and each brother has the same two sisters. But GPT-4 answered two, which surprised Hinton.

"It surprises me. It surprises me it still screws up on that," he said during the interview.

Geoffrey Hinton on the limits of AI
Stating that while GPT-4 is very advanced, it’s not perfect, he said “It's an expert at everything. It's not a very good expert at everything.”

Still, Hinton believes future AI models will improve. When asked if GPT-5 would get the riddle right, he replied, "Yeah, I suspect."

OpenAI launched GPT-4 in 2023 and became popular for clearing difficult exams like the SAT, GRE, and bar exam. In May 2024, OpenAI released GPT-4o, which became the default model for ChatGPT.

The company recently unveiled new models— GPT-4.1 , GPT-4.1 Mini, and GPT-4.1 Nano that are designed to outperform previous iterations, offering developers more efficient and cost-effective AI solutions.

One of the standout features of GPT-4.1 is its expanded context window, supporting up to 1 million tokens, significantly surpassing the 128,000-token limit of its predecessor, GPT-4o. This enhancement allows the model to process and understand larger datasets, making it particularly effective for complex tasks such as analysing extensive codebases or lengthy documents.
Additionally, GPT-4.1 delivers notable gains in coding performance, showing a 21% improvement over GPT-4o and a 27% increase compared to GPT-4.5.