Oxford’s AI Chair: LLMs are a HACK

This video discusses the capabilities and limitations of large language models (LLMs) like GPT-4, suggesting they are more of a “technological hack” than a deep model of human intelligence. The speaker argues that LLMs excel at pattern recognition rather than genuine problem-solving, and are limited in planning and reasoning, failing when problems are obfuscated. The speaker believes the transformer architecture is fundamentally limited and unlikely to achieve advanced AI capabilities. Despite these limitations, LLMs are acknowledged as impressive and useful, transforming philosophical AI questions into experimental science.