Most AI approaches rely heavily on the program developed by humans. They can only do what they were programmed to do. They can only learn what they are taught. When faced with new environments, these systems get stuck. But this is slowly changing with the rise of AI agents. But is that enough?
In a recent interview, British neuroscientist Karl Friston underscored a transformative potential in current AI agents: the integration of curiosity.
“Their inability to independently select training data limits their capacity for genuine intelligence. While they excel at data processing and prediction, they lack the curiosity and independent thought essential for true scientific advancement and problem-solving,” he said.
Why is curious AI even necessary? To answer this, look at these four research papers to understand the need for curiosity.
A “linear vs. loopy maps” study found that while LLMs do well on simple, linear tasks, they have trouble on complex ones that involve cycles or dead ends.
This restriction was investigated in further detail in the “TravelPlanner” study, which showed that LLMs frequently perform poorly in complicated decision-making scenarios that call for taking into account a variety of limitations and possible outcomes.
An additional significant vulnerability is to an LLM’s capacity to efficiently retrieve and apply the knowledge it has received training on. Studies on “LLM lookup capabilities” have revealed that these models’ performance in locating pertinent data might vary, which can result in errors and imprecise results.
How Can Being Curious Solve the AI Problem
Nick Clegg, the president of global affairs at Meta, has called AI models stupid, similar to how Yann LeCun calls them ‘dumb’. “They can process information and identify patterns incredibly fast, but they don’t truly understand the world in the same way humans do. They’re essentially sophisticated pattern-matching machines,” he said.
The idea is to compound curiosity with intelligence. By combining human and AI curiosity, we can leverage their unique strengths to compound our creative potential. The intuitive nature of human curiosity can work in tandem with the computative power of AI curiosity to accelerate discoveries and drive innovation.
Putting it in the words of Sonya Huang, partner at Sequoia: “As the models get bigger and bigger, they begin to deliver human-level, and then superhuman results.”
This proved true in the new algorithm developed by researchers at MIT’s Improbable AI Laboratory and CSAIL. The paper highlighted that the algorithm automatically piques interest when needed, but when not, it stifles it until the agent gathers enough information from its surroundings to decide what to do.
When evaluated on more than 60 video games, the method performed well on both hard and easy exploration tasks, whereas prior algorithms could only handle a hard or easy domain on their own.
“Previously what took, for instance, a week to successfully solve the problem, with this new algorithm, we can get satisfactory results in a few hours,” co-author Zhang-Wei Hong said. This efficiency is crucial in real-world applications where time and resources are limited.
The study expands on past research by OpenAI that showed how AI agents driven by curiosity could succeed in challenging gaming environments such as Montezuma’s Revenge.
Making AI Agents More Curious
Over the years, scientists have worked on algorithms for curiosity, but copying human inquisitiveness has been tricky. For example, most methods aren’t capable of assessing AI agents’ gaps in knowledge to predict what will be interesting before they see it.
For example, TEXPLORE-VENIR is a reinforcement learning method that was created by Todd Hester and Peter Stone. It incentivises programs to find new knowledge and reduce uncertainty. It offers intrinsic rewards for comprehending novel concepts, such places or recipes, in contrast to traditional approaches that only concentrate on reaching predetermined goals.
Another example is IBM’s Project Debater AI which aims to engage in competitive debates rather than genuinely exploring the nuances of the topics discussed. IBM claims that Debater AI is the first-ever AI system designed to meaningfully engage with humans in a debate.
This improves curiosity-driven exploration and learning efficiency.
Take chatbots, for example—it’s common to see chatbots that can answer frequently asked questions (FAQs). On the other hand, customer service quality can significantly improve if chatbots have a certain level of perceived emotional intelligence that can be achieved by injecting curiosity-driven behaviours.
In healthcare, more curious models could accelerate drug discovery by exploring vast chemical spaces with greater efficiency. In robotics, curious AI could enable robots to adapt to new environments and tasks more rapidly.
Definitely, making AI agents curious is something to look forward to.
