Can generative AI help build a global hive mind?

On this planet of science fiction, the notion that humanity will someday join our brains collectively into a world “hive thoughts” dates again to a 1930 novel entitled “Final and First Males” by Olaf Stapledon. It recounts a fictional “future historical past” through which humanity evolves biologically right into a species that may hyperlink telepathically and type collective minds of utmost intelligence.

In the actual world this pursuit is named Collective Superintelligence, and it’ll not require telepathy or different fictional units. As a substitute, it would use rising generative AI applied sciences to attach massive human teams into real-time deliberative techniques, enabling us to solve difficult problems by harnessing our mixed data, knowledge and perception in highly effective new methods.

This pursuit has been my private focus as an AI researcher for the final decade and I consider it has the potential to supply superintelligent systems that preserve human values, morals and pursuits on the core of each perception, evaluation or resolution. In fact, for many individuals the thought of huge human teams considering collectively in real-time techniques appears unnatural, and even creepy, however mom nature would disagree.

In actual fact, many social species have advanced naturally on this route, growing the power to make speedy choices in massive teams that vastly exceed the brainpower of particular person members. Biologists name this Swarm Intelligence, and it allows colleges of fish, swarms of bees and flocks of birds to shortly resolve life-and-death issues at intelligence ranges that far exceed the psychological capability of their particular person minds.

One exceptional type of Swarm Intelligence that impressed my very own academic research over the past decade is the common-or-garden fish faculty. Though it appears easy on the floor, the underlying dynamics allow 1000’s of members to make advanced choices with no particular person in cost. In actual fact, fish colleges could make good choices even when no single member has adequate data to deal with the issue at hand. Contemplate the picture beneath: 

The fish faculty above faces a hypothetical drawback of utmost significance: Three predators method from three totally different instructions. For the time being proven, no particular person is conscious of all three threats. In actual fact, most should not conscious of any threats. There’s a small group of fish within the higher left that’s conscious of the primary predator, a small group within the decrease left that’s conscious of the second predator, and a small group within the higher proper that’s conscious of the third predator. A lot of the different fish are unaware of the hazards.

So, how can this massive group, through which each member has restricted data (and no members possess all the knowledge) resolve this life and dying drawback and shortly transfer in the appropriate route? The very first thing to know is that fish have a particular organ on the perimeters of their physique known as the lateral line that permits them to trace the velocity and route of neighboring fish primarily based on strain and vibrations within the water round them. Utilizing this organ, they’ll understand the intent of their neighbors (that’s, the velocity and route that these fish consider the college ought to go). This communication is bidirectional so we will consider it as a local deliberation through which small teams resolve on the movement of their small portion of the college.

That is attention-grabbing, but it surely doesn’t clarify how world choices are made. In any case, the group on the appropriate that sees a predator approaching, seemingly decides the college ought to transfer left. On the identical time, the 2 teams on the left seemingly resolve the college ought to transfer proper. And, teams within the center, which haven’t any data of the predators, are more likely to preserve shifting within the route they have been already going. So, how does this get resolved right into a single speedy resolution that evades the pressing threats?

The magic occurs as a result of each fish within the faculty “deliberates” with a distinct group of close by neighbors. This implies there are a lot of “overlapping conversations” occurring without delay which allows data to shortly propagate throughout the complete faculty. As proven beneath, the result’s a swift and decisive collective resolution to the issue.

On this approach, education fish could make rapid and effective decisions throughout massive populations, even when all members have restricted data. Such a ability could be much more highly effective for big human teams.  In any case, the issues confronted by teams of persons are considerably extra advanced and contain much more views. This begs the query: Might massive human teams deliberate in real-time with the effectivity of fish colleges and shortly attain optimized choices?

For years this aim appeared inconceivable. That’s as a result of human conversations have been proven to be most efficient in small groups of four to seven folks and shortly degrade as teams develop bigger. It is because the “airtime per individual” will get progressively squeezed and the wait-time to answer others steadily will increase. By 12 to fifteen folks, the conversational dynamics change from considerate debate to a collection of monologues that turn into more and more disjointed. By 20 folks, the dialog ceases to be a dialog in any respect. This drawback appeared impenetrable till advances in generative AI opened up new options.

The ensuing expertise is named Conversational Swarm Intelligence (CSI) and it guarantees to permit teams of probably any dimension (200, 2000 and even 2 million folks) to debate advanced issues in real-time and shortly converge on options with considerably amplified intelligence. Step one is to divide the inhabitants into small subgroups, every sized for considerate dialog. For instance, a 1,000-person group may very well be divided into 200 subgroups of 5, every routed into their very own chat room or video conferencing session. In fact, this doesn’t create a single unified dialog — it creates 200 parallel conversations.   

As described above, fish colleges resolve this drawback by having native teams overlap, permitting data to shortly propagate throughout the complete inhabitants. Sadly, we people didn’t evolve with the power to be in a number of conversations without delay. In actual fact, if we attempt to concentrate to 2 conversations, we instantly get confused and may’t give attention to both. That is generally known as the “cocktail party problem” as a result of it occurs typically when small teams collect inside earshot of one another. When you attempt to concentrate to a neighboring dialog you instantly lose monitor of the dialogue you might be in. 

So how can we overcome this human limitation?

CSI expertise solves this drawback by inserting LLM-powered “conversational surrogates” into every subgroup. These AI brokers are tasked with distilling the real-time human insights inside its assigned group and sharing these insights with surrogate brokers in a number of different teams. The receiving brokers categorical the acquired insights in their very own teams as pure first-person dialog. On this approach, every subgroup is given a man-made member that participates seamlessly in overlapping conversations, guaranteeing data freely propagates throughout the complete inhabitants. 

Quite a lot of latest research counsel the method is efficient. For instance, a 2023 study performed at Carnegie Mellon College to check real-time deliberations amongst roughly 50 folks in conventional chatrooms versus conversational swarms. When utilizing the CSI construction, teams have been capable of maintain extra coherent conversations that shortly converged on options. As well as, every particular person was discovered to contribute 50% extra content material (on common) than members utilizing conventional strategies.

However does this amplify group intelligence?

To discover this, a follow-up 2024 study by researchers at Carnegie Mellon and Unanimous AI examined the power of networked human teams to take IQ exams as a real-time “hive thoughts.” Outcomes confirmed that teams of 35 individuals who averaged an IQ of 100 (the fiftieth percentile), might rating an efficient IQ of 128 (the 97th percentile) when utilizing a web-based CSI platform known as Thinkscape. Though this examine used conversational teams of solely 35 members, different latest research have examined teams up to 250 with success.

Whereas the above research used textual content conversations, the core strategies of CSI may be deployed for teleconferencing, videoconferencing and even VR conferences, enabling massive teams of a whole lot and even 1000’s of members to carry coherent real-time conversations that effectively resolve issues, prioritize choices, brainstorm concepts and attain choices, all with amplified group intelligence. This has the potential to reinforce a variety of fields from enterprise collaboration and market research to civic engagement and deliberative democracy. 

In the long run, this method may very well be used to construct superintelligent techniques which are inherently aligned with human values, morals, knowledge and sensibilities. In principle, we might use CSI expertise to allow hundreds of thousands of people all over the world to “assume collectively” as a world brain-of-brains to unravel our most troublesome issues. For me, this can be a safer path than counting on a purely synthetic superintelligence, as AI techniques could not preserve human values or pursuits over time. That’s why I consider we’d like applied sciences like Conversational Swarm Intelligence and instruments like Thinkscape that leverage gen AI — however preserve people within the loop. 

Louis Rosenberg is a longtime researcher within the fields of AI, collective intelligence and combined actuality. He’s CEO and chief scientist of Unanimous AI.