AI's Secret: It Thinks Less Than You Think

Prominent figures in artificial intelligence from organizations such as OpenAI, Anthropic, Google, and more continue to assert with confidence that AI achieving human-like intelligence is imminent. However, critics are becoming increasingly numerous and vocal. They argue that AI does not actually think in the same way humans do.

The research conducted by these scientists indicates that there might be inherent limitations tied to the basic design of current AI models. Present-day artificial intelligences can mimic intelligence primarily through learning an immense quantity of general guidelines, which they then choose to apply based on the various data points they come across.

This differs from the numerous methods through which humans and even animals can comprehend their surroundings and anticipate what’s ahead. As living organisms, we create "mental frameworks" that explain how various phenomena operate, encompassing causality and effects.

Numerous AI engineers assert that their models also possess internal representations of these world models within their intricate networks of artificial neurons. This is demonstrated through their capability to generate coherent writing that suggests logical thinking. Advances in what are known as "reasoning models" have led certain observers to believe even more strongly that systems like ChatGPT indeed exhibit this capacity. already attained human-level capability, referred to in the sector as AGI, which stands for artificial general intelligence.

For most of their existence, ChatGPT and its competitors have been mysterious black boxes .

There was no insight into the methods they used to generate their impressive outcomes, as they were educated instead of coded, with an enormous quantity of parameters within their AI "minds" capturing data and reasoning in complex ways that eluded even those who created them. However, scientists are crafting innovative instruments that enable examination of these systems from the inside. These findings have led numerous experts to doubt whether we are truly nearing Artificial General Intelligence (AGI).

"There's disagreement over what these models are truly accomplishing, and some debate surrounding the human-like terminology used to describe them," according to Melanie Mitchell a professor at the Santa Fe Institute specializing in AI research.

‘Bag of heuristics’

New methods for investigating large language models—a part of an expanding area referred to as " mechanistic interpretability "—demonstrate to researchers how these AIs perform mathematics, learn to play games, or navigate through environments. In a series of recent essays Mitchell contended that an increasing amount of research indicates models might be forming large collections of "problem-solving shortcuts" instead of developing more effective cognitive frameworks to understand scenarios and tackle the tasks they face. ("Shortcut" here refers to what experts call a heuristic—a method used to solve problems quickly.)

When Keyon Vafa An AI researcher at Harvard University was introduced to the "bag of heuristics" theory and remarked, "It felt like a revelation." He adds, "This precisely captures what we’ve been aiming to explain."

Vafa’s investigation aimed to uncover the type of mental map an artificial intelligence constructs when it is trained using millions of turn-by-turn instructions similar to those provided by Google Maps. Vafa and his team utilized New York City’s intricate web of streets and avenues as their primary data set.

The outcome didn’t resemble a street map of Manhattan at all. Upon closer examination, it became clear that the AI had deduced numerous unrealistic movements—paths that bypassed Central Park entirely or moved diagonally across multiple city blocks. Despite this, the final model was still able to provide accurate turn-by-turn instructions for navigating between any two locations within the borough with an impressive 99% precision rate.

Despite its upside-down layout potentially confusing any driver, the model has effectively mastered distinct guidelines for various navigation scenarios, beginning from all potential start points, according to Vafa.

The immense computational capabilities of AI, combined with extraordinary processing power, enable them to tackle problem-solving in complex ways that would be unachievable for humans.

Thinking or memorizing?

Other investigations focus on the unusual behaviors exhibited when large language models attempt mathematics—a task they have traditionally struggled with but are improving upon. Certain studies indicate that these models develop distinct sets of rules for multiplication within specific ranges, such as between 200 and 210, differing from those used for other numerical intervals. If this approach seems suboptimal to you, then you're correct.

This body of research indicates that modern artificial intelligences operate much like intricate, makeshift contraptions designed by Rube Goldberg—comprising numerous patchwork solutions aimed at addressing specific queries from us. Recognizing these systems as extensive collections of improvised guidelines might clarify why they falter when faced with tasks slightly beyond what they were trained to handle, according to Vafa. His group observed that merely obstructing 1% of the simulated Manhattan streets, thereby necessitating alternate routes, caused significant degradation in the AI's effectiveness.

He points out this shows a significant distinction between current AI systems and humans. While an individual may struggle to recall turn-by-turn directions for navigating New York City with 99% precision, they would possess the mental adaptability to navigate around construction work.

This study indicates why numerous models are so large: These systems must remember countless heuristics instead of condensing this information into a comprehensive understanding as humans do. This could clarify why these models require vast quantities of data for learning when people grasp concepts with only a handful of examples. For each specific heuristic rule, exposure to every conceivable arrangement of words, visuals, board-game scenarios, etc., is necessary. Moreover, effective training demands repeated encounters with these configurations multiple times.

This study could also shed light on why AI systems from various corporations appear so similar. “thinking” the same way , and they are even reaching the same level of performance— performance that could be leveling off .

AI researchers have overestimated their progress previously. Back in 1970, Marvin Minsky, a professor at MIT, claimed to Life magazine that within "three to eight years," computers would possess the cognitive abilities akin to those of an ordinary person.

Last year, Elon Musk claimed That AI will surpass human intelligence by 2026 was mentioned in February by Sam Altman. wrote on his blog that “systems that start to point to AGI are coming into view,” and that this moment in history represents “the beginning of something for which it’s hard not to say, ‘This time it’s different.’” On Tuesday, Anthropic’s chief security officer warned That "remote workers" will be employed by U.S. companies within one year.

Even if these predictions turn out to be too early, AI is here to remain and transform our lives. Developers are still discovering ways to work with it. Utilize these unquestionably remarkable systems to help us all be more productive And despite their natural intelligence perhaps leveling off , efforts to refine them continue.

In contrast, exploring the constraints of how AI "thinks" might play a crucial role in improving their capabilities. Recently, this area of study has gained importance. essay , MIT AI researcher Jacob Andreas indicated that gaining deeper insights into the difficulties faced by language models opens up innovative approaches for their training: "As we begin to tackle these limitations, we have the potential to enhance LMs, making them more precise, reliable, and manageable."

Send your message to Christopher Mims at christopher.mims@wsj.com