On Collective Superintelligence
Over the last decade, my goal as an AI researcher has been to push for the development of AI-powered systems that provide a viable alternative to Artificial Superintelligence — one that achieves super-intelligent capabilities but does so by amplifying human intellect rather than replacing it. My reason is simple: I believe that creating an artificial superintelligence will be an extremely dangerous and dehumanizing step for our species. If you need convincing, see my tiny 2020 picture book entitled Arrival Mind.
That said, is there a safer path to superintelligence? When I started exploring this back in 2008, I concluded that our best approach would be to create a Collective Superintelligence that uses the power of AI to amplify the combined thoughts, opinions, insights, and wisdom of very large groups in real time. No, I’m not talking about using AI to process the combined artifacts of humanity the way LLMs do. Instead, I’m talking about processing the real-time thoughts and opinions of human groups, not to replicate our collective thinking but to facilitate it and amplify it.
To be honest, I wasn’t quite sure if this was possible in a seamless and natural way, but new research at Unanimous AI convinces me this vision is possible. Even more exciting is that the path to Collective Superintelligence will likely include near-term technologies that could transform how large groups work together right now, creating entirely new modes for Enterprise Collaboration and Deliberative Civic Engagement and even largescale Governmental Decision-Making.
Before I get into any of that, let me give some background:
My inspiration for Collective Superintelligence (CSi) has been mother nature. That’s because biological systems have evolved over hundreds of millions of years that show the power of real-time collective intelligence. For example, large fish schools, bee swarms, and bird flocks can think together so effectively, they can function as super-organisms that seamlessly amplify the collective intelligence of the full group. Biologists call this amplified intellect, Swarm Intelligence. The big question I asked over a decade ago was whether human groups could achieve similar results.
To pursue this goal, I founded Unanimous AI in 2014. Over the decade since, we have conducted many academic studies that prove human groups can form a Swarm Intelligence that significantly amplifies their collective insights. This has encouraged me that the long-term goal of Collective Superintelligence is viable. We’ve also discovered that it’s a difficult challenge, for it involves understanding the dynamics of human groups as they interact with each other and with real-time AI systems.
Our core technology is called Artificial Swarm Intelligence (Swarm AI) and it works well. Famously, I was challenged by a reporter back in 2016 who dared our team 𝘵𝘰 𝘱𝘳𝘦𝘥𝘪𝘤𝘵 𝘵𝘩𝘦 𝘒𝘦𝘯𝘵𝘶𝘤𝘬𝘺 𝘋𝘦𝘳𝘣𝘺 (the first four horses in order) using a “human swarm.” We accepted the challenge, not realizing at the time that it’s considered one of the hardest bets to win in all of sports.
This is the result as it was reported in Newsweek back then: “AI turns $20 into $11,000 on Kentucky Derby Bet.” Of course there was luck involved, but beating 540-to-1 odds was not random chance. Our system (called Swarm) combined the insights of 20 horseracing enthusiasts into a real-time collective intelligence. And it worked:
Also in 2016, we ran the world’s first “Ask me Anything” with a real-time collective intelligence. We conducted this live on Reddit, which was risky to say the least, but it worked — enabling hundreds of networked people to answer questions as a unified swarm. It drew so much attention, it was chronicled by the Washington Post in an article where they flagged some of the predictions that came out of this milestone event. In particular, they published a prediction that California would legalize marijuana before the end of the year. Five months later, that prediction came true.
Over the years since, our Swarm AI technology has been used by the United Nations, the US Air Force, and Fortune 500 companies. But the grand pursuit of building a Collective Superintelligence seemed far off. That all changed over the last 18 months as new AI methods (including LLMs) have allowed us to quietly pursue a series of breakthroughs that could make a seamless Collective Superintelligence a viable possibility.
Our most exciting innovation is Conversational Swarm Intelligence. It’s a technology we began pursuing in 2018 but wasn’t viable at scale until we could leverage LLMs in a unique way. It promises to allow groups of almost any size (50 people, 500 people, even 5 million people) to hold real-time conversational deliberations that converge on unified solutions that are amplified by the power of swarm intelligence. The core concept was inspired by fish schools and formal testing shows it actually works.
How is this related to schools of fish? Well, it turns out that fish schools can hold real-time “conversations” among thousands of members, rapidly making collective decisions to navigate the ocean with no central authority mediating the process. Each fish communicates with others using a unique organ called a “lateral line” that senses pressure changes in the water caused by neighboring fish. The number of neighbors that a given fish pays attention to varies from species to species, but it’s always a small subset of the group. And because each fish interacts with an overlapping subset of other fish, information quickly propagates across the full population, enabling a unified Swarm Intelligence to emerge.
Can we enable something similar in humans? It turns out we can by using a concept we developed in 2018 called hyperswarms that divides real-time human groups into overlapping subgroups. For example, we can take a large group of 1000 networked individuals and divide them into 200 groups of 5 people, the members of each subgroup automatically routed into their own unique chat room or videoconferencing room. And if we send them all the same question to discuss or problem to solve, we now have 200 parallel conversations, each reasonably sized for thoughtful deliberations.
Of course, parallel groups doesn’t create a Swarm Intelligence because content can’t propagate across the population. We solved this through the novel use of conversational AI agents powered by LLMs to emulate the function of the lateral line organ in fish. In particular, we insert an AI agent (referred to as an Observer Agent) into each of the 200 subgroups and task that agent with monitoring the dialog in that group, distilling the salient content, and expressing the content in a neighboring group through natural dialog. In this way, each of the 200 groups is given a sixth member that happens to be an AI observer that is tasked with monitoring the real-time insights discussed in one group and conversationally expressing them into one or more other groups, enabling information to propagate across the full population. This can be represented in this diagram:
But does this crazy idea work? To be honest, I wasn’t sure that large groups could hold unified conversations in this way and converge on coherent solutions in real-time — but it turns out, they can. Even more surprising, when have participants compare conversations in a single large chat room with being placed into a conversational swarm, they overwhelmingly prefer the biologically inspired structure. In fact, the vast majority of participants reported feeling like they were much better able to express their views and have their opinions heard in this unique swarm structure than in a traditional structure like a large chat room or forum.
To explain how Conversational Swarm Intelligence works, I am sharing a presentation that I gave on October 14th at the 7𝘵𝘩 𝘐𝘯𝘵𝘦𝘳𝘯𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘑𝘰𝘪𝘯𝘵 𝘊𝘰𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦 𝘰𝘯 𝘈𝘥𝘷𝘢𝘯𝘤𝘦𝘴 𝘪𝘯 𝘊𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘐𝘯𝘵𝘦𝘭𝘭𝘪𝘨𝘦𝘯𝘤𝘦 (𝘐𝘑𝘊𝘈𝘊𝘐 2023). It describes a research study we conducted earlier this year in collaboration with Carnegie Mellon University. It’s one of many studies we’ve been conducting to assess the benefits of CSI. The results are promising.
𝐖𝐡𝐲 𝐢𝐬 𝐭𝐡𝐢𝐬 𝐢𝐦𝐩𝐨𝐫𝐭𝐚𝐧𝐭? Again, I firmly believe we need an alternative to Artificial Superintelligence that is constructed using the real-time values, moral, sensibilities, wisdom, and insights of human participants. Again, I’m not talking about processing the artifacts left behind by humans like traditional AI does. I’m talking about connecting human groups in real-time so they can function as a unified superintelligence. If we can achieve this, I believe we can use the power of AI to help humanity make better decisions and solve difficult problems by optimizing our ability to think together.
If we don’t build a Collective Superintelligence, I fear we are just building our own digital replacements. For details on why this is a genuine risk, see my 2017 TED talk: https://www.youtube.com/watch?v=Eu-RyZt_Uas
Are you interested in participating in conversational swarms? Well, if you have a large networked team (50 people or more) that might want to take part in “beta tests” of Conversational Swarm Intelligence, we continue to conduct sessions in collaboration with Carnegie Mellon University. So please, send us an email and ask to find out more: Gregg@Unanimous.ai
Dr. Louis Rosenberg is an AI researcher and early pioneer of virtual and augmented reality. His work began over 30 years ago in labs at Stanford and NASA. In 1992 he developed the first functional mixed reality system at Air Force Research Laboratory. In 1993 he founded the early VR company Immersion Corp which he brought public on NASDAQ. In 2004 he founded the early AR company Outland Research. He received his PhD from Stanford, was a professor at California State University, and has been awarded over 300 patents for VR, AR, and AI technologies. He is currently the CEO and Chief Scientist of Unanimous AI.
