人工智能让世界变得更愚蠢？专家警告AI炒作很快就会破灭

On August 14, a report emerged that sheds light on Gary Marcus, a 54-year-old cognitive scientist and one of the most controversial figures in the artificial intelligence (AI) community. His ongoing debates with Sam Altman, CEO of OpenAI, and Elon Musk center around the capabilities of large language models. While Altman and Musk predict that AI will soon develop human-like abilities, Marcus warns of potential pitfalls, arguing that the current hype surrounding AI models is likely to "burst soon."

When asked about his criticisms of OpenAI, Marcus clarified that his opposition lies not with the organization or Altman personally, but rather with what he describes as "false promises." He criticizes OpenAI's claim of developing general artificial intelligence for the benefit of humanity, stating that their software is neither safe nor reliable and often neglects public interest.

Despite Altman's assertions that OpenAI is on the brink of achieving general AI, Marcus dismissed this notion as exaggerated. He pointed to GPT-4 as an example of a failed attempt at building GPT-5, emphasizing that the pace of AI development has been slowing down, indicating a misdirection by OpenAI.

While recognizing the popularity of conversational AI like ChatGPT, Marcus contends that these models are not trustworthy. He acknowledges the quality of responses but emphasizes the significant risks they carry, particularly regarding misinformation and hallucinations—instances where the models generate incorrect or fabricated information.

He explained the mechanics of large language models, which function as probabilistic systems predicting the next word in a sentence. Although they can handle a majority of queries, their lack of understanding means they often falter in scenarios outside their training scope. This unpredictability poses a substantial issue for users.

Marcus differentiated human errors from those made by machines, stressing that humans bring innate reasoning to certain tasks that AI simply lacks. He cited the risks associated with AI-assisted driving, where overconfidence in the technology can lead to dangerous errors.

Addressing the reliance on vast data sets to improve AI performance, Marcus raised concerns about diminishing returns from newer model variants and the misconception that simply increasing data volume will lead to significant advancements. He highlighted the inadequacy of the data being used and pointed out that human intelligence, which harnesses a variety of inputs, is not being replicated by machines.

Lastly, he criticized the trend of using synthetic data—AI-generated data—for training models, viewing it as a misguided approach that only exacerbates the proliferation of misinformation online. While many celebrate advancements like ChatGPT passing the Turing Test, Marcus argues that such tests do not truly measure intelligence and merely expose how easily humans can be deceived.

His reflections emphasize a call for a more cautious and discerning approach to AI development, advocating for a focus on building reliable and intelligent systems rather than succumbing to the allure of groundbreaking claims.