Vitalik Buterin: GPT-4 Passes Turing Test

OpenAI’s GPT-4, an advanced generative artificial intelligence (AI) model, has reportedly passed the Turing Test, according to Ethereum co-founder Vitalik Buterin. The Turing Test, a concept introduced by the renowned mathematician Alan Turing in 1950, serves as a benchmark to determine how human-like an AI’s conversational abilities are. Turing suggested that if an AI could produce text that fooled humans into thinking they were conversing with another person, it would demonstrate a form of “thought.”

Nearly 75 years later, the co-creator of Ethereum, the world’s second most popular cryptocurrency, believes that recent research from the University of California San Diego indicates that GPT-4 has achieved this milestone. A preprint paper from the university titled “People cannot distinguish GPT-4 from a human in a Turing test” details an experiment involving around 500 human test subjects. These subjects engaged in blind tests with both humans and AI models to determine if they could tell which was which. The findings revealed that humans mistook GPT-4 for a human 56% of the time, meaning the AI fooled people more often than not.

Vitalik Buterin interprets these results to mean that GPT-4 has indeed passed the Turing Test. He elaborated on his perspective via Farcaster, stating, “It means people’s ability to tell if it’s a human or a bot is basically a coin flip!” Buterin added a caveat that humans were identified as humans 66% of the time compared to 54% for bots, but he considers the 12% difference negligible in real-world settings. He emphasized that this achievement marks an essential milestone, hinting that AI has now reached a level of sophistication previously contained to science fiction.

Alan Turing’s test does not directly correlate with artificial general intelligence (AGI), despite both terms often being conflated. Turing’s test focuses on whether AI can deceive humans into thinking it is also human through text-based conversation. Importantly, the Turing Test remains an ephemeral and subjective construct without a universally accepted benchmark or technical foundation. To date, there is no scientific consensus on whether AI can truly “think” like a human or how such a capability could be measured.

Turing’s conceptual predictions were made long before the advent of token-based AI systems and generative adversarial networks (GANs), which underpin today’s generative AI technologies. These modern systems have pushed the boundaries of what AI can accomplish, making the prospect of fooling humans more achievable than ever before.

The notion of artificial general intelligence complicates the discussion further. In scientific terms, “general intelligence” refers to an entity capable of performing any intellectual task that a human can. Given that no human demonstrates expertise across all intellectual domains, a “general artificial intelligence” would theoretically possess cognitive abilities far surpassing those of any known human.

While GPT-4 does not meet the stringent criteria for AGI as defined by the scientific community, it has not prevented enthusiasts from using the term “AGI” more loosely. Within today’s AI discourse, terms like “AGI,” “human-like,” and “passes the Turing test” frequently describe any AI that can produce content akin to human output.

The recent achievements of GPT-4 as indicated by the UC San Diego research and embraced by figures like Vitalik Buterin, signal a profound leap in AI capabilities. These advancements compel us to continually redefine and reassess our benchmarks for measuring AI’s progress. Whether or not the GPT-4 represents true thought or general intelligence, it is undeniable that the landscape of AI has dramatically evolved, bringing us ever closer to technologies that challenge our understanding of intelligence and human likeness.