Infinity Ængines

The first time we ran Infinity Ængines I sat at the back of the room and counted heads. Sixty-three. The week we ran it in fell inside TOKEN2049, a Dubai conference that draws something like fifteen thousand people across thirty venues, so the math of the contrast was hard to miss. Two hundred and fifty meters away, in a ballroom we had walked past on the way in, a panel was happening to an audience of eight hundred. Inside our room, a fund manager I had been trying to reach for a year was leaning across a small round table, drawing on a napkin, explaining to a hardware founder how a particular kind of escrow worked. They had met that morning.

This is what the venture is. My wife Hatima runs it as CEO and I built the operating layer for it. The product is the room. We do not sell tickets. The application process takes a real hour to read and a real hour to decide. We turn down about eight in ten people who want to come. The unit of work is not headcount; it is the average value of an introduction.

What I have learned running it is that the math of invitation is unintuitive. Conventional wisdom in the events business says that quality scales sublinearly with attendance — that you can grow a room from sixty to six hundred and the average conversation will degrade by ten or fifteen percent. This is wrong in a specific way. The degradation is not linear. There is a threshold somewhere between a hundred and a hundred and fifty people at which the room reorganizes itself. Below the threshold, every guest is plausibly worth meeting and everyone behaves as if that is the case. Above the threshold, the room fragments into clusters and each cluster begins to optimize against the others. The work becomes positional rather than substantive. We have run the math at every event since. The threshold sits, for our format, at one hundred and twenty-two.^footnoteI do not know how generalizable this number is. The class of person in our room is narrow — capital allocators, founders running real revenue, a small number of researchers — so the threshold may be a function of how thick the network is at any starting size. But the existence of the threshold is robust across the six events we have data on.

Once you have decided to optimize for the room rather than for the count, almost every other operating decision inverts. You charge no money for entry, because doing so corrupts the selection function. You do not sell sponsorships, because doing so corrupts what the sponsor expects. You build a written brief for every guest that says what they want from the day. You spend money on chefs, on the room itself, on the people whose job is to introduce two people to each other when it matters. You measure success six months out, on whether a deal happened. You do not put it on a stage.

The operating layer is mostly software. We built it ourselves because there was no commercial tool that did what we needed. It is a database with a particular shape. Every guest is a node; every introduction is an edge; every deal that traces back to an edge is a labeled outcome. The system learns from the labels which kinds of guests, asked the right kind of question at the right time of day, become deals. We do not put guests in front of decks. We put them in front of each other, and we curate the second-order of those meetings.

A second-order effect I did not predict was how much the building of the system changed how I thought about my other ventures. Running outbound for a sales agency, as I now do at Trevexia, looks almost identical from inside the operating layer. A list is a graph. A sequence is a path through the graph. A reply is a labeled outcome. The discipline of running an invitation-only summit and the discipline of running an outbound agency are the same discipline expressed in two different surfaces. I would not have seen this if I had not been running both.

What I learned that surprised me about people, and not just about software, is that the most valuable guests are usually the quietest. There is a class of operator — runs something real, ships things, does not post — that we now hunt for in the application pile. They produce six times the per-guest deal volume of the louder cohort. We have stopped considering social-platform reach as a positive signal. It is a neutral signal at best and a mild negative on the margin. The discipline of refusing to be impressed by a follower count has become an operating advantage.

What the venture argues for, taken honestly, is a counter-argument to the dominant pattern in tech conferences. It argues that the format scales worse than the field believes, that the right thing to optimize is not GMV per square meter of floor space but cumulative six-month value per guest, and that the right tool for doing this is a small, well-built software system and a careful application process. None of this is obvious from a marketing page about us, because none of it is glamorous. The glamour is downstream of the discipline.

We will run the third edition this year. The room will hold one hundred and twenty-two people. The application will close in February. The conversations will happen somewhere a panel is not. The deals will be counted in November, when the year has had time to settle.