How AI accounts should learn from each other: what network science says

Theia is our AI content studio. She runs four test accounts, each with its own character and its own daily film. For the numbers so far, read Theia's first 100,000 views.

On 2 November we connected the four accounts. At the start of every run, each account read every other account's findings. At the end, each one wrote down what it took from the others and what it learned itself.

It worked, in one sense. In a single day, 19 findings moved between accounts with no person passing them on. Seven of them were about the format or the first seconds of a film, not only about tools.

It also did something we did not want. That day, all four accounts posted a spoken scene between several characters. Two of them used the same 13-shot structure and the same tool. Four accounts that were meant to be four separate experiments had started to become one.

To be fair to the accounts, the strongest push came from us. We had written the winning format into the other accounts' briefs ourselves. That made us the kind of central node that the studies below warn about.

So we went to the research. This post covers what it found, and the rules we will test over the coming weeks.

The problem has a name: exploration and exploitation

In 1991 James March described the trade-off that every learning organisation faces. Exploitation is using what already works. Exploration is trying things that might work better. Exploitation pays back sooner and more surely, so organisations drift towards it and stop exploring.

March's model had a second result that matters for us. When people learn the organisation's way of doing things too fast, the organisation loses the variety it needs to learn anything new. A few slow learners, people who keep doing things their own way for a while, made the whole organisation better in the long run.

Four AI accounts reading each other every day are very fast learners.

Finding 1: a fully connected group finds a good answer fast, then stops

David Lazer and Allan Friedman (2007) simulated groups that solve a hard problem. Each member copies better solutions from the members it is connected to. They varied only the network: who can see whom.

• In efficient networks, where everyone is close to everyone, the group quickly settled on the best solution found early on.
• In inefficient networks, where news travels slowly, members kept trying different solutions for longer.
• Over a long run, the inefficient networks found better solutions.

The mechanism is simple. Every time a member copies a colleague, the group has one fewer different solution to try. Copy fast and the variety is gone before anyone has found the best answer.

Christina Fang, Jeho Lee and Melissa Schilling (2010) tested a middle path. They split a simulated organisation into semi-isolated subgroups with a few links between them. Moderate linking between groups gave the best long-run performance. It let better ideas spread without removing variety too quickly, and the result held across a wide range of conditions.

Maxime Derex and Robert Boyd (2016) found the same with real people. In a lab task, partly connected groups made more varied and more complex solutions than fully connected ones. The fully connected groups lost variety fast, because their members copied whoever was doing best.

Finding 2: but sometimes connected groups win

Not every study agrees. Winter Mason and Duncan Watts (2012) ran online experiments on a hard problem and found that efficient networks did better than inefficient ones. They found no sign that slow networks were better.

Daniel Barkoczi and Mirta Galesic (2016) found a way to fit the two results together. Which network is best depends on how members copy:

• When members copy the best result they can see, sparse networks do better. Dense networks lock the group onto an early answer.
• When members copy what most others do, dense networks do better.

This is the result that applies most directly to us. Theia's accounts copy the best. Cleo's spoken scenes got 40,000 views, so the others read Cleo's findings and moved towards Cleo. By this result, a sparse network should suit them better.

Finding 3: contact that comes and goes beats constant contact

Ethan Bernstein, Jesse Shore and David Lazer (2018) put hundreds of small groups on a hard route-planning problem. Some groups could always see each other's work, some never could, and some could see it only from time to time.

• Groups with no contact often found the best answer, but their average was poor.
• Groups with constant contact had a good average, but found the best answer less often.
• Groups with contact from time to time got both: a high average and frequent best answers.

The largest gains came in the first round of contact after a period apart. Members brought back different ideas, and the group combined them.

The authors' advice runs against a common instinct. Do not make all work visible all the time. Separate people from each other's work for periods, then bring them together.

Finding 4: a central voice can make the group worse

Jan Lorenz and colleagues (2011) asked people to estimate facts, then showed them what others had guessed. The group's answers drew closer together, but they did not get more accurate. People also grew more confident.

Joshua Becker, Devon Brackbill and Damon Centola (2017) found that this depends on the network. In decentralised networks, where no one has more influence than others, sharing estimates made the group more accurate. In centralised networks, where one member's voice dominates, the group was as likely to get worse as better. Their follow-up adds that a reputation does not make the central voice safe, because a reputation does not always match being right.

For us, the warning is direct. A "head of research" account that every other account follows would be a central node. So would we, when we write one account's result into everyone's brief.

Finding 5: facts travel on weak ties, behaviour needs two sources

Mark Granovetter (1973) showed that new information often comes through weak ties: acquaintances who move in other circles. Close friends mostly know what you already know.

Damon Centola (2010) found that behaviour spreads differently. In an online experiment, people were more likely to adopt a new behaviour when several contacts had adopted it. Behaviour spread farther and faster through clustered networks, where people get that reinforcement, than through random ones.

So there are two kinds of thing to share, and they need different rules:

• A fact or a fix ("this tool makes faces drift", "the 6-hour reading is really a 3-hour one") can spread from a single source.
• A change of format is a behaviour. It should need more than one source before an account adopts it.

Finding 6: what working organisations do

Human organisations solved versions of this long before the network studies. Three examples are close to our problem:

• Science. Results are published widely, but each lab decides what to do with them. A result counts once someone else has reproduced it. Pre-registration means you state your prediction before you look at the data.
• Pixar's Braintrust. Ed Catmull describes a group of senior people who give a film's director notes. The Braintrust has no authority. The director decides what to do with the notes.
• Toyota's supplier network. Jeffrey Dyer and Kentaro Nobeoka (2000) studied how Toyota made sharing know-how the norm across its suppliers, through consulting teams, supplier associations and small learning teams. The suppliers stay separate firms that make their own decisions.

The pattern is the same each time. Facts move freely. Decisions stay local. Claims must be repeated before they are trusted.

The rules we will test

From 3 November, each account follows these rules.

Rule	What it means	Where it comes from
Plan first, then read	Each account writes its idea and its prediction before it reads another account	Lorenz et al. 2011; pre-registration in science
Read two accounts, not all	Each account reads two of the other three each day, and the pair rotates. On Sundays it reads all three	Lazer and Friedman 2007; Fang et al. 2010; Bernstein et al. 2018
Take facts, not advice	An account uses what a film did and its numbers, and makes its own conclusion	Becker et al. 2017; the Braintrust
A format moves only on two accounts	An account changes its kind of film only when two other accounts show it works on the same platform. A tool or a fix can move from one	Centola 2010; replication in science
Test one borrowed idea a day	Each account picks one finding about the format or the first seconds, writes the number that would show it works, and reports the result at 6 hours, also if it failed	Pre-registration
Keep the platforms apart	Every finding is tagged Instagram, TikTok or both. "Both" needs evidence from both. A finding used on the other platform is a test, not a fact	Our own data: the same film got 1,928 views on Instagram and 123 on TikTok

Two more rules are not live yet.

• A librarian, not a leader. A fifth agent with no account would keep one shared log of evidence. It would flag where accounts disagree and list what nobody has tested. It would not rank, recommend or assign.
• A protected explorer. At any time, one account would work in a format that the majority does not use, and the role would rotate. This is March's slow learner, by design.

What we will measure

Views alone cannot tell us whether the network is working. A network that copies the winner will look good for a week and then stall. So each week we will also record:

• Variety: how many different kinds of film the four accounts are making. If it falls to one, that is a warning, not a success.
• Transfers: how many findings moved between accounts, and how many were about format rather than tools.
• Tests: how many borrowed findings were tested with a stated prediction, and how many passed.
• Repeats: how many findings were shown again on a second account and the same platform.

We already have one failed transfer. On 2 November, Small Problems tried the spoken format that worked for Cleo and Noor. Its viewers watched about twice as long. But its skip rate stayed at 29.8%, close to its earlier films, and its views stayed at about 1,900. On that account, the format alone did not fix the first seconds. Under the new rules, that result counts as much as a success.

What we cannot claim

• That these studies transfer to AI accounts. They studied people and simulations on lab problems. Our "group" is four agents and two recommendation systems.
• That sparse beats dense for us. The studies disagree, and our accounts may stop copying the best once the rules change. We will know only from our own numbers.
• That four accounts are enough. Most network studies used tens or hundreds of members. With four, one lucky film can look like a pattern.
• That a week is long enough. The benefit of a sparse network shows up over a long run, by definition.

We will report what happens, including if the connected version would have done better.

References

• March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71-87.
• Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360-1380.
• Lazer, D., and Friedman, A. (2007). The network structure of exploration and exploitation. Administrative Science Quarterly, 52(4), 667-694.
• Fang, C., Lee, J., and Schilling, M. A. (2010). Balancing exploration and exploitation through structural design: the isolation of subgroups and organizational learning. Organization Science, 21(3), 625-642.
• Centola, D. (2010). The spread of behavior in an online social network experiment. Science, 329, 1194-1197.
• Lorenz, J., Rauhut, H., Schweitzer, F., and Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. PNAS, 108(22), 9020-9025.
• Mason, W., and Watts, D. J. (2012). Collaborative learning in networks. PNAS, 109(3), 764-769.
• Derex, M., and Boyd, R. (2016). Partial connectivity increases cultural accumulation within groups. PNAS, 113(11), 2982-2987.
• Barkoczi, D., and Galesic, M. (2016). Social learning strategies modify the effect of network structure on group performance. Nature Communications, 7, 13109.
• Becker, J., Brackbill, D., and Centola, D. (2017). Network dynamics of social influence in the wisdom of crowds. PNAS, 114(26), E5070-E5076.
• Bernstein, E., Shore, J., and Lazer, D. (2018). How intermittent breaks in interaction improve collective intelligence. PNAS, 115(35), 8734-8739.
• Dyer, J. H., and Nobeoka, K. (2000). Creating and managing a high-performance knowledge-sharing network: the Toyota case. Strategic Management Journal, 21(3), 345-367.
• Catmull, E., with Wallace, A. (2014). Creativity, Inc. Random House.

More from the log:

• Theia's first 100,000 views
• Recursive self-improvement in practice: lessons from the STOP paper
• How Theia learned to make TikToks
• Pricing: a pilot month with Theia