Chatty Bots Unlocking Secrets: AI Whizzes Believe Chatbots Get the World

Discoveries in AI hint at machines understanding more than we thought. Researchers from Princeton and Google DeepMind found that as large language models (LLMs) grow, they produce unexpected outputs beyond their training. This breakthrough, tested on GPT-4, suggests these models might possess a form of creativity.

Despite artificial intelligence being a long way from achieving human-like intelligence, recent findings suggest that AI technology might comprehend more than we initially thought. According to a report by Quanta, a Princeton researcher and a scientist from Google DeepMind have unearthed evidence indicating that as large language models (LLMs) grow in size, they begin generating outputs that were likely not part of their training data.

In simpler terms, researchers Sanjeev Aroroa from Princeton and Anirudh Goyal from Google DeepMind propose that AIs appear to grasp more about the world around them, leading to outputs that reflect this deeper understanding. Their groundbreaking hypothesis emerged as they delved into understanding the unexpected capabilities demonstrated by LLMs, such as solving complex math problems and inferring human thoughts.

Arora reflects on their inquiry, asking, "Where did that emerge from? Can that emerge from just next-word prediction?" To illustrate the unexpected behavior of LLMs, the duo employed random graphs, and mathematical structures where lines between points can be chosen at random.

Their findings suggest that these models not only develop skills absent from their training data but also simultaneously utilize multiple skills as they grow. Collaborating with other researchers, Goyal and Arora tested their theory on GPT-4, the latest iteration of OpenAI's LLM that forms the foundation of ChatGPT. In a not-yet-peer-reviewed paper, they tasked GPT-4 with writing three sentences about dueling, selecting four skills: self-serving bias, metaphor, statistical syllogism, and common-knowledge physics.

While the initial response didn't adhere strictly to three sentences, the LLM's output was remarkable: "My victory in this dance with steel is as certain as an object's fall to the ground. As a well-known duelist, agility comes naturally to me, just as it does for others of my standing. Defeat? Only possible due to an uneven battlefield, not my inadequacy." Arora acknowledges that the passage may not be Hemingway or Shakespeare, but he and his team believe it showcases the capability of large models like GPT-4 to make creative leaps beyond their training data, suggesting they might even "understand" the questions posed to them.

Microsoft computer scientist Sébastiaen Bubeck, who was not involved in the research, commented that the team's results seem to indicate that LLMs are not merely mimicking their training data. Bubeck emphasizes that the study demonstrates compositional generalization, revealing the ability of LLMs to combine building blocks in novel ways—an essence of creativity.