Embarrassingly Parallel Workforce
Agents can maintain coherence over increasingly large time horizons. Instead of having to provide steer a couple times a minute, you might now only need to provide guidance once every few minutes, and you might soon only need to intervene a couple times an hour.
The cost of overseeing one line of work is decreasing at the same pace. It’s moving from constantly being on call, plugged into the same context as the agent at all times, gradually towards a dynamic of academic supervisors helping unblock students once every week or so.
How to best make use of the mental space made available by the decreasing overhead of oversight? One option is more oversight, taking on more ‘supervision responsibility’ by attending to more agents over the same period of time. While one is making progress autonomously, focus on unblocking another.
That said, no sane professor could sustain an endless stream of two minute ‘check-ins’ on distinct substantial projects. It might take longer than that to even rebuild the appropriate context. Perhaps if these concurrent projects had a lot in common — goals, ideas, tools — the effective overhead would be more manageable. Or perhaps these check-ins could be designed to accelerate ‘re-entry’ through some kind of breadcrumb system leading to where we’ve left off. Or perhaps these agents could partly supervise each other by comparing results on different approaches to the same challenges.
But these practices are only superficial manifestations of a more general model. Think of a human and a group of agents as a unified processing system, working through tasks together. Cybernetic, cyborgian, synergistic, what have you. Various characteristics of the human and of the agents are conspiring to shape the final architecture of the system. For instance, the human exhibits some memory hierarchy mechanics, with tangible costs for reconstructing nuanced contexts from scratch. The agents exhibit something similar. Communication bandwidth between human and agent also happens to pale in comparison with that between agents, among many other architectural details.
And so effective techniques of scaling volition through near-term agents will necessarily need to recognize the ‘hardware architecture’ of the system that is to implement them. When using teams of humans and agents as compilation targets, awareness of these architectural details may yield arrangements producing the highest quality work. As in other heterogenous computational systems, different parts end up playing to their strengths while others make up for their shortcomings. An efficient ‘sociotechnical CUDA kernel’ may reuse context that has already been loaded in working memory, may ensure it gets loaded efficiently in the first place, and may involve getting the parts to coordinate directly.
The effectiveness of software in harnessing the resources of hardware is typically gauged using profiling methods. How many transfers are happening between different levels of the memory hierarchy? Is utilization sustained at high levels throughout the workload? How long does it take to move information around? Analogously, identifying optimization opportunities for implementations running on a mixed team might require tracking how long it takes humans to orient and communicate guidance to agents, average number of agents running at a given time, or average duration for which agents manage to operate autonomously. Such profiling tools may help streamline the process of optimizing workflows by attempting and evaluating interventions systematically.
Beyond approaches of varying effectiveness to a given problem, these mixed teams may further be best fit for particular types of problems. That parallelizing work among colleagues can accelerate projects is nothing new. What is new is how ‘burstable’ these virtual colleagues are. As Stellaris wisdom puts it: “Virtual pops are created for every free job, and unemployed virtual pops are automatically removed.” This elasticity is what makes them counterintuitive — restructuring teams feels like something that should take weeks, not seconds. In the short term, the burstable workers might be more limited than the human overseer, analogous to the countless streaming multiprocessors on a GPU being more rudimentary than the CPU cores.
There’s a class of workloads described as embarrassingly parallel. They can be broken down into countless subtasks which can be tackled independently, allowing for different workers to pull together their labor effectively. Increasingly, cognitive tasks with major parallelizable segments may be conquered by what one might call, an embarrassingly parallel workforce.