The intelligence explosion has a bottleneck nobody models: coordination

Most models of an intelligence explosion share one assumption: once AI can do AI research, you can add virtual researchers almost without limit, and progress compounds. A new Epoch AI report by Phil Trammell argues that this skips over a real bottleneck. More researchers only help if you can split the work among them, coordinate it, and combine the results, and that ability is itself a technology that has to keep improving.

Trammell lays out three cases. If this coordination technology improves as fast as the number of researchers grows, you get the explosive takeoff the standard models predict. If it improves faster than the research frontier but slower than the researcher count, an explosion still happens, just more gradually. If it only keeps pace with ordinary exponential growth, then automating research speeds things up but never tips into runaway acceleration. He uses a robot factory to make the point concrete: each improvement still has to be identified, built, and tested in sequence, so doubling the engineers cannot halve the time unless the tools for parallel work also get better.

The value of the report is that it names a variable the usual takeoff arguments leave out. Epoch's framing is that parallelization is one of the parameters that decides whether and when an explosion happens, yet it is barely studied in the economic models people cite when they forecast timelines.

Why it matters

If you weigh AI timeline forecasts for planning or policy, this gives you a concrete question to ask of any explosion argument: does it quietly assume coordination scales for free? Forecasts that ignore parallelization limits are likely too aggressive on how fast automated research can compound.