The latest research from Meta’s AI scientist Yann LeCun shows that AI is unlikely to take over our jobs any time soon. In a recently published article, LeCun argues that artificial intelligence is much stupider than humans on the most important issues.
This paper, co-authored by a number of other scientists, including researchers from other AI startups such as Hugging Face and AutoGPT, examines how AI’s general purpose reasoning stacks up against the average human. To measure this, the research team created its own list of questions that would be, as the study describes, “conceptually simple for humans but challenging for the most advanced AIs.” The questions were fed to a group of people and a plugin-equipped version of GPT-4, OpenAI’s latest large language model. The research, which has not yet been peer-reviewed, tested AI programs on how they would answer “real-world questions that require a range of core competencies, such as reasoning, multimodality use, web browsing, and general tool use proficiency.”
The questions asked by the researchers required the LLM to take a series of steps to obtain information to answer. For example, one question asked the LLM to visit a specific website and answer a question specific to the information on that site; in others, the program was supposed to perform a general web search for information associated with a person in a photo. But the bottom line is that LLMs were not very successful. Indeed, research results show that large language models are typically outperformed by humans when it comes to more complex real-world problem-solving scenarios.
In his recent tweets, LeCun criticizes the industry’s current technological capacity, openly arguing that artificial intelligence is not close to human capacity.