A few related threads converged for me this morning:
- Some recent thinking about Objective Function Gaming.
- A brief conversation I had with my brother-in-law about the role that transformers and related technology will be playing in the network security space, and the ways in which the usual cat-and-mouse adversarial interactions will evolve. I started wondering if we might see a proliferation of ‘war gaming’ sort of exercises in which the white hats and black hats play out different scenarios in real-but-sandboxed or entirely simulated network environments.
- This video on the really cool DrEureka work in robotics.
There’s kind of an underlying theme in these threads about extending the realm of simulation / automation from only part of the system to the entire system—doing everything in simulation. Put another way, taking the slow-thinking, easily exhausted human out of the loop such that the entire design / gaming environment is simulation + AI agents.
There’s nothing new here (think of the original War Games movie!), but I’m thinking that the tools we now have available—LLMs capable of rapid and fairly creative idea generation, simple ‘reasoning’, and rudimentary agentic behavior, together with simulated environments—mean that we’re about to see an explosion of this type of work.
Here’s a vizcue (visual mnemonic) for this concept:
We move from a slow, plodding scenario of humans interacting with a simulated environment, twiddling the knobs and seeing results, then iterating (top of the figure) to one in which the agentic component is rolled into the simulation space, resulting in rapid iteration and improvement.
Zoom!