If ants, birds, fish, and bees can pool their collective intelligence to make decisions, why can’t humans? The collective intelligence manifests itself in colonies, flocks, schools, and swarms, something that biologists refer to as “swarm intelligence.” So, why can’t humans form their swarms? That’s the question that guides Louis Rosenberg, founder and CEO of Unanimous A.I.
His company has developed an easy-to-use technology that allows humans across the world to amplify their collective intelligence, often with startling accuracy. The basic idea is that no single person has all the information to predict the future, but if a group of people pool their intelligence then they can come up with a prediction that’s significantly closer to accuracy.
“We’ve done a lot of work that shows that humans, when you amplify their intelligence as a swarm, can get significantly smarter,” Rosenberg told Inverse.
The company does this with its platform, called UNU, which allows users around the world to log in from their own computers and weigh in on questions until a consensus is reached. Rosernberg spoke with Inverse about how and why UNU works.
How would you explain the way UNU works to someone who hasn’t used it?
The basic technology that underlies the platform is what we call swarm A.I., because it’s modeled after swarms in nature. There are two forms of intelligence in nature: There’s neurological intelligence, and there’s swarm intelligence. Neurological intelligence is a system of neurons where an intelligence emerges that’s way smarter than any of the single neurons, and a swarm intelligence is one level up. It’s a system of brains that work together and are smarter than any of the individual participants. We model systems after how swarms work in nature and we enable groups of people to amplify their intelligence by thinking together as a system.
authentic smile unu
Unanimous A.I.'s UNU technology can be used for analytical tasks like predicting sporting events or for social tasks like determining whether someone's smile is authentic.
So how does a swarm’s prediction accuracy compare to an individual’s?
When we form a swarm we’re combining the knowledge and wisdom and insight and intuition of the group in an optimal way so they can combine these things and converge on the best possible answer for all of their insights. And when we do that the swarm is significantly smarter than the individuals who participate.
To give you an example, we were challenged to predict the Oscars by Newsweek. We took 50 regular movie fans, just average people, and we had them first predict as individuals who they thought would win each of the categories, and then we had those same 50 people work together as a swarm to predict the same set of Oscars.
Hepburn's Four Oscars
Unanimous A.I.'s users correctly predicted the Oscars with greater accuracy than the average movie critic, even though none of the users had seen all the movies.
And what was remarkable is that individuals were on average 44 percent accurate, which doesn’t sound great, but it’s actually really hard to predict the Oscars, there’s a lot of categories. But when those same 50 people worked together as a swarm they jumped all the way up to 76 percent accurate, which was almost double the accuracy when they were working together as a system. And what’s even more interesting is that we can also look at professional movie critics, because all that data exists. The average professional movie critic was 64% accurate. So what we saw is that these 50 average people were able to amplify their intelligence to the high end of expert level performance.
How much do individual participants need to know to come up with an accurate prediction in a swarm?
We asked the individuals in that Oscar example how many of them had actually seen all the movies. Not a single one of those 50 people had seen all the movies. They had all seen some of the movies, but that’s the magic of a swarm: As long as there is knowledge out there in the population it doesn’t need to be complete knowledge. These 50 people who all saw some of the movies were able to fill in the gaps in each other’s information and converge on the best answer. And part of the reason they can do that is that a swarm is not a poll, it’s not a survey, it’s a system where people are not just expressing their opinion but expressing their level of conviction, their level of confidence.
unu world series swarm
Unanimous AI correctly predicted the Chicago Cubs would win the 2016 World Series.
How are the results of a swarm different from the results of a focus group or a poll?
A poll or a survey that counts up votes treats the participants, the people, as data points. What a swarm is doing is it’s actually treating the participants as data processes. It’s not asking them for a single piece of data, it’s asking them to participate in a system. So what happens is in a swarm everybody is pushing and pulling against each other in real time. So they’re all present at the same time, and so we could have 100 people in the swarm at the same time and they are interacting with each other, allowing them all to find that one solution that they can best agree upon. So everybody’s varying their opinions continuously, as opposed to just providing a single data point in a poll and then somebody statistically computing an answer after the fact.
Unanimous A.I. correctly predicted not only that Nyquist would win the 2016 Kentucky Derby but also that Exaggerator, Gun Runner, and Mohaymen would follow, in that order. This order, called the superfecta, paid off 542:1.
This might sound like kind of a silly question, but at a glance UNU looks kind of like a Ouija Board. How is it different?
A Ouija Board is enabling a few people to work together as a system and if they’re really doing it right, meaning if they don’t believe that they’re really pushing on the puck then they might be tapping into their subconscious feelings. That’s the place where a Ouija Board gets a little different in that with a Ouija Board they’re not supposed to be expressing a view, they’re supposed to be a channel for supernatural feelings. But if we assume, for example, that what we really have is their subconscious opinions coming out then a Ouija Board is enabling a small number of people to combine those views in an interesting way.
What an online swarm does is enable not three or four people but 100 people, 200 people, to combine their views in real time and they can be located anywhere. They don’t need to be in the same room. They could be all across the world. You could have a swarm of people on every continent, and 200 people can combine their views in real time, and they’re not channeling the supernatural, they are expressing their actual knowledge, their actual opinions, their actual insights, and they’re finding the solution together that’s really the best possible combination of the information from all the participants.
UNU is like a Ouija Board if a Ouija Board was powered by A.I. algorithms and had 100 people gathered around it.
Ultimately a swarm hones in on a single answer. Do swarms cause participants to change their minds to go along with the group?
Swarms encourage participants to be flexible. That’s one of the other big differences between a swarm and a poll: A poll by its very nature is polarizing. A poll will cause people to entrench on a single view and it’s actually telling people to not be flexible, and that’s why polling leads to these very entrenched positions in a population. A swarm, which is really nature’s way of doing this, encourages groups to be flexible and to find solutions that they can best agree upon. Now that doesn’t mean that it’s encouraging people to go along with the group, it’s just encouraging people to be flexible, and by flexible it might mean if somebody starts out with a certain opinion and believes that’s the best possible solution, but then varies their opinion as they realize that there’s more support for another option, but it still doesn’t mean that they’re just going along with the majority.
How specific can UNU get in terms of coming up with predictions about dates, numerical values, et cetera?
So there’s two different basic styles of questions that we ask inside the UNU platform. We either ask a question where there’s a set of choices, so I could say, “Who’s going to win the Super Bowl this year.” And I could provide six teams, and the swarm could converge on the most likely team that’s going to win the Super Bowl this year. We never ask a single question with more than six choices because there’s also a lot of social science research that shows beyond six choices people get overwhelmed. It’s called choice overload.
The other types of questions is where we have a continuous value, and so instead of the group picking among said choices they’re actually moving along a number line, so we can say, “What are the odds that a particular candidate is going to get elected?” And then the swarm can consider all the values from zero percent to 100 percent and converge on 65.5 percent and get a very accurate number that way. We can do that with any kind of numerical value.
Sometimes we might not know the choices in advance. We might have the question to a swarm that says, “What’s the most important new technology that’s going to emerge in the next five years.” And we might want to actually collect choices from the group. And so we can set up the questions so that a dialogue box pops up on everybody’s screen in the swarm, and they can type in their own suggestion. So we collect suggestions from the swarm, which then populate around the swarm and they can choose among those suggestions, and if there’s more than six suggestions then they’ll do it in groups of six, but it allows us to have the swarm essentially engage in brainstorming, coming up with ideas, and then swarm to converge on the best possible answer among all the ideas that were surfaced.
unu campaign issues
Unanimous A.I.'s UNU platform can find answers to divisive questions by building consensus among users.
**We’ve been talking a lot about how humans are making decisions in a swarm. Where does A.I. come into play?
There’s two different ways to look at the artificial intelligence aspects of what we do. There’s the low level and there’s the high level, and I’ll talk about the low level first, which is that we use intelligence algorithms that work behind the scenes to combine the knowledge and wisdom and insight and intuition that is expressed by the groups of people. And our algorithms are looking not just at what option that they’re selecting, but it’s looking at their behavior. It’s looking at how are people behaving over time so it can determine their varying levels of confidence or conviction, because that’s really what a swarm is. It’s not just a group of data points, it’s a group of people who are behaving and varying their confidence and conviction and allowing us to find the best possible solution. So there are a whole bunch of algorithms that run behind the scenes to enable the swarm to react to the behaviors of the participants in real time.
The other level of the A.I. is the bigger picture esoteric view, which is that ultimately artificial intelligence is about building a replica of some other type of natural intelligence. And most A.I. research that’s been going on for the past 50 years has been focused on one type of natural intelligence, which is neurological intelligence, so there’s lots and lots of research out there on how to build artificial neurons and replicate a neurological intelligence, in other words replicating a brain.
What we do is we say, “Well nature actually has two ways of building intelligence. Yes there’s neurological intelligence but there’s also swarm intelligence.” Natural systems have evolved the ability to combine the intelligence of large populations in optimal ways, and so what we do is we build the algorithms and interfaces that replicate swarm intelligence. Humans cannot naturally form a swarm. We didn’t evolve the same abilities that birds have to form flocks and fish have to form schools, or bees have to form swarms. But with the right algorithms and the right interfaces we can enable these artificial swarm intelligence systems that involve people and algorithms.
So at the highest level we’re building artificial swarms in the same way that other types of A.I. systems are building artificial brains, and these things are related because ultimately what a swarm is is a system of brains. It’s a brain of brains. And what we do is basically enable these artificial hive minds to form and to answer questions and to ideally reach the best possible answers for the set of information that the participants have.
What uses do you see for UNU in the near future?
There are lots and lots of application areas that are interesting for UNU and for our swarm AI technology. We currently have a variety of business customers who are giving us a project to tap the intelligence of consumers to generate optimized insights about new products, new services, and new features, to predict how a new advertising campaign might succeed. We have customers who ask us to have swarms watch movie trailers and predict if those movie trailers will drive people to the theaters. We have swarms that can watch a TV ad and predict the impact that that ad will have on consumers. So by building swarms of consumers there’s all kinds of really interesting insights that can be generated that we would say falls into this area of business intelligence.
The other thing we’ve done for businesses is build swarms not of average consumers, but of their team, their staff. Inside of any organization there’s a lot of knowledge and insight and intuition. We did a project where we could tap the intelligence of a sales organization to make optimized predictions about sales forecasting. And again each individual could make their own prediction, but when they think together as a system they can be more accurate.
We recently did a study with some researchers at Oxford looking at predicting financial markets using swarm intelligence, and we find that a group, financial traders can be more accurate by predicting together as a swarm than they could be on their own.
The other thing that’s interesting is not predictions but reaching decisions when you have a diverse group.
And we’ve been looking at, for example, medical applications. For example sometimes treatment planning could involve a team of doctors that all have different specialties and different expertise, and what a swarm could do is allow a group of doctors that include a radiologist, an oncologist, an internist, people with different backgrounds and different perspectives to combine their different views in the most efficient way and converge on a decision that they can best agree upon.
So in the decision making area there’s a lot of interest as well, and you could imagine long term that going to political decisions or governmental decisions, where traditional votes and polls cause groups to polarize, swarming could bring groups together to find solutions that are just best for the group as a whole.
This interview has been edited for brevity and clarity.