On 12 August 2026, the French developer podcast If This Then Dev (IFTTD) published episode 369, “Intelligence(s): Collective”. Its guest was Émile Servan-Schreiber, co-founder and Chief Scientific Advisor of The Forecasting Machine and CEO of Hypermind. Host Bruno Soulez framed it as part of a six-episode summer series with six researchers. He explained why: of the show’s last 52 episodes, 38 had been about artificial intelligence, while intelligence itself had barely come up. The hour-long conversation ranged from a country fair in 1906 to the inner workings of today’s neural networks.
Groups are not smart by default
Servan-Schreiber, a cognitive psychologist trained at Carnegie Mellon, made one point early that shaped the rest of the episode. Left to itself, a group tends to become collectively foolish, because members’ natural biases reinforce one another. Collective intelligence only appears when somebody adds rules, algorithms or incentives that work against those instincts. In that sense, he argued, collective intelligence is itself a kind of artificial intelligence. To illustrate why organisation matters as much as numbers, he compared a termite colony with a human brain. A million termites with about 250,000 neurons each add up to roughly three times the 80 billion neurons in one human head. Yet the colony cannot design a cathedral, because our fewer neurons are much better organised.
Galton’s ox and the diversity theorem
The episode’s anchor example was the ox-weighing contest that Francis Galton described in Nature. Servan-Schreiber said he had reworked Galton’s 787 tickets for his 2018 book Supercollectif (Fayard). Individual guesses were off by about 4.5% on average, while the average of the whole crowd was almost exactly right. Accuracy also improved with crowd size, though with diminishing returns: going from one to ten people cut the error by about three, and going from ten to a hundred cut it by three again.
He explained this with the “diversity theorem”, a result associated with the complexity scientist Scott Page. The crowd’s error equals the average individual error minus the diversity of opinions, meaning how far individual guesses spread around the group average. Put positively, he said, performance equals expertise plus diversity. Less expertise can therefore be offset by more varied opinions, because independent errors tend to cancel out, much as noise-cancelling headphones play an inverted wave to cancel ambient noise. The episode page took its tagline from his summary:
On utilise la bêtise des autres pour annuler sa bêtise à soi.
Émile Servan-Schreiber, If This Then Dev, 12 August 2026
In our translation: “We use other people’s foolishness to cancel out our own.”
How many novices make an expert?
Asked to put a ratio on expertise versus diversity, Servan-Schreiber said it depends on the task. He cited a study on estimating civil jury verdicts that compared experienced trial attorneys with Stanford law students, published in 2011 in the Journal of Empirical Legal Studies. In his account, a single expert made about half the error of a single student. Averaging about 14 students matched one expert, and matching two experts took about 42. He stressed that this does not make experts useless, since experts also gain from pooling their views. These ratios are his reading of the data; the paper itself focuses on how much weight participants gave to a partner’s second opinion.
Rules for juries, markets and meetings
Soulez brought up 12 Angry Men. Servan-Schreiber read the film as a lesson in rules: the unanimity requirement gives a lone dissenter room to resist the pressure to conform. He then described a 2014 study in PNAS that ran simulated stock markets in South-East Asia and Texas. Ethnically diverse groups of traders priced assets more accurately than homogeneous ones. His explanation was that among people who look like us, we tend to assume they think like us and stop thinking for ourselves. Among people who look different, we instinctively think more independently.
For meetings, he drew practical advice from the same theorem:
- experts should listen to less expert colleagues, because their input adds diversity to the group’s judgement;
- newcomers should speak up when they see something the group has missed, as long as it has not already been said;
- the boss or the expert should not speak first, because that discourages disagreement;
- the goal is a decision, not a consensus: diverge as much as possible, then combine views through a vote, an average or a bet, much as an election does.
Language models as collectives
In the last part of the conversation, Servan-Schreiber described large language models as collective at every level. They are trained on the combined output of many different people. Their attention “heads” each read the input in a different way. And very deep “residual” networks, which add shortcut connections that skip layers, behave like ensembles of many shallower networks. This last point was set out in a 2016 paper by Andreas Veit, Michael Wilber and Serge Belongie, which found that performance degrades gradually as layers are removed. He linked that curve to the one seen when Galton’s crowd is made smaller. Even crowds and models hit limits, he added: well-run forecasting crowds are well calibrated, so events given a 25% chance happen about a quarter of the time, but past a certain horizon the world becomes too chaotic to forecast.
A long-form precursor from 2022
The IFTTD conversation was not Servan-Schreiber’s first extended podcast treatment of these ideas. In November 2022 he shared an 84-minute episode of Can I get that software in blue?, hosted by Steve Mayzak and Chad Tindel. The official description says the discussion went into the science of collective intelligence, how people can borrow information from others in their networks, and how Hypermind’s prediction-market platforms were being used by companies and governments. That earlier episode supplies a broader business and organizational counterpart to IFTTD’s later focus on the mechanisms that make groups intelligent.
His recommendation for listeners, beyond his own book, was Max Bennett’s A Brief History of Intelligence. Earlier in the year, on 18 March, he had covered similar ground on Nicolas Bassan’s OrgaNova podcast, in an episode titled “Prédire l’avenir en collectif : quand les foules battent les experts” (“Forecasting together: when crowds beat experts”).
