In the world of science fiction, the concept humanity will in the future connect its brains into a worldwide “hive mind” dates back to a 1930 novel called “Last and First Men” by Olaf Stapledon. It tells a fictional “future story” through which humanity biologically evolves right into a species able to telepathic communication and forming collective minds of utmost intelligence.
In the true world, this pursuit is named Collective superintelligenceand it’s going to not require telepathy or other fictitious devices. Instead, emerging generative AI technologies shall be used to attach large groups of individuals into real-time deliberative systems, allowing us to solve difficult problems by utilizing our combined knowledge, wisdom and insights in powerful recent ways.
This pursuit has been my personal focus as an AI researcher for the past decade and I imagine it has the potential to Superintelligent systems that put human values, morals and interests on the centre of each insight, assessment or decision. Of course, the thought of ​​large groups of individuals considering together in real-time systems seems unnatural and even scary to many individuals, but Mother Nature thinks otherwise.
In fact, many social species have naturally evolved on this direction, developing the flexibility to make rapid decisions in large groups that far exceed the intelligence of any individual member. Biologists call this swarm intelligence, and it allows schools of fish, swarms of bees, and flocks of birds to quickly solve life-or-death problems at a level of intelligence that far exceeds the mental capability of their individual brains.
A remarkable type of swarm intelligence that made me my very own Scientific Research over the past decade is the standard school of fish. Although it looks easy on the surface, the underlying dynamics allow 1000’s of members to make complex decisions without anybody individual being in charge. In fact, schools of fish could make good decisions even when no single member has sufficient information to resolve the issue at hand. Consider the image below:
The school of fish above faces a hypothetical problem of utmost importance: three predators are approaching from three different directions. At the moment shown, no individual is aware of all three threats. In fact, most are usually not aware of any threat. There is a small group of fish within the upper left that’s aware of the primary predator, a small group within the lower left that’s aware of the second predator, and a small group within the upper right that’s aware of the third predator. Most of the opposite fish are unaware of the hazards.
So how can this massive group, where each member has limited information (and no member has all the knowledge), solve this life-or-death problem and move quickly in the fitting direction? First of all, you’ve to know that fish have a special organ on the perimeters of their body that Sideline This allows them to trace the speed and direction of neighboring fish based on the pressure and vibrations within the water around them. This organ allows them to sense the intent of their neighbors (that’s, the speed and direction through which they imagine the college should move). This communication is bidirectional, so we will consider it as a local advicethrough which small groups settle on the applications of their small a part of the college.
This is interesting, but it surely doesn't explain how global decisions are made. After all, the group on the fitting that sees a predator coming is probably going to make your mind up that the swarm should go left. At the identical time, the 2 groups on the left are likely to make your mind up that the swarm should go right. And the groups in the center, with no knowledge of the predators, are more likely to proceed within the direction they were already going. So how does this translate right into a single quick decision that avoids the pressing threats?
The magic happens because each fish in the college “consults” with a special group of neighbors. This implies that many “overlapping conversations” are happening concurrently, allowing information to spread quickly throughout the college. As shown below, the result’s a fast and decisive collective solution to the issue.
In this fashion, education fish quick and effective decisions across large populations, even when all members have limited information. Such a capability could be much more powerful for big groups of individuals. After all, the issues facing groups of persons are much more complex and require much more perspectives. This raises the query: could large groups of individuals deliberate in real time with the efficiency of faculties of fish and quickly arrive at optimized decisions?
For years, this goal seemed inconceivable. Human conversations have been proven to be the best in small groups of 4 to seven people and drops off quickly as groups get larger. This is because “airtime per person” becomes increasingly scarce and the wait time to reply to others steadily increases. At 12 to fifteen people, the conversation dynamic changes from a thoughtful debate to a series of monologues that turn out to be increasingly disjointed. At 20 people, the dialogue ceases to be a conversation in any respect. This problem seemed insoluble until advances in generative AI opened up recent solutions.
The resulting technology is named Conversational swarm intelligence (CSI) and guarantees that groups of any size (200, 2000, and even 2 million people) will give you the chance to debate complex problems in real time and quickly find solutions with greatly increased intelligence. The first step is to divide the population into small subgroups, each large enough to facilitate thoughtful dialogue. For example, you may divide a gaggle of 1000 people into 200 subgroups of 5 people each, each directed into its own chatroom or video conference session. Of course, this doesn't create a single unified conversation, but fairly 200 parallel conversations.
As described above, schools of fish solve this problem by overlapping local groups, allowing information to spread quickly throughout the population. Unfortunately, we humans didn’t evolve with the flexibility to take part in multiple conversations without delay. In fact, after we attempt to listen to 2 conversations, we turn out to be immediately confused and unable to give attention to either one. This is often known as “Cocktail party problem“because that's what often happens when small groups gather inside earshot of one another. If you are trying to listen to a conversation round the corner, you'll immediately lose track of the particular discussion.
So how can we overcome this human limitation?
CSI technology solves this problem by inserting LLM-based “Conversation surrogates” into each subgroup. These AI agents are tasked with distilling the human insights in real time inside their assigned group and sharing those insights with surrogate agents in a number of other groups. The receiving agents express the insights they receive inside their very own groups as a natural, first-hand dialogue. This way, each subgroup gets a synthetic member who seamlessly participates in overlapping conversations, ensuring that information spreads freely throughout the population.
Quite a lot of recent studies suggest that this approach is effective. Study 2023 carried out at Carnegie Mellon University to match real-time deliberations of about 50 people in traditional chat rooms with discussion groups. When using the CSI structure, groups were in a position to have more cohesive conversations that quickly led to solutions. In addition, each individual was found to contribute (on average) 50% more content than participants using traditional methods.
But does this increase group intelligence?
To investigate this, a follow-up study in 2024 by researchers at Carnegie Mellon University and Unanimous AI tested the flexibility of networked human groups to perform IQ tests as a real-time “hive mind.” The results showed that groups of 35 individuals who had a median IQ of 100 (the fiftieth percentile) were in a position to achieve an efficient IQ of 128 (the 97th percentile) when using an internet CSI platform called Thinking landscapeAlthough this study used discussion groups with only 35 participants, other recent studies have tested groups as much as 250 with success.
While the above studies used text conversations, CSI's core methods will also be applied to teleconferencing, videoconferencing, and even VR meetings. This enables large groups of a whole lot and even 1000’s of members to have cohesive real-time conversations through which problems are efficiently solved, options prioritized, ideas collected, and decisions made – all with enhanced group intelligence. This has the potential to enhance a wide selection of areas, from corporate collaboration to Market research on civic engagement and deliberative democracy.
In the long run, this approach could possibly be used to construct superintelligent systems which can be inherently in tune with human values, morals, wisdom, and sensibility. In theory, we could use CSI technology to empower thousands and thousands of individuals all over the world to “think together” as a worldwide brain of brains to resolve our hardest problems. To me, this can be a safer path than counting on purely artificial superintelligence, as AI systems may not maintain human values ​​or interests over time. That's why I imagine we’d like technologies like Conversational Swarm Intelligence and tools like Thinkscape that leverage generational AI – but keep humans within the loop.
