With all the rage about letting AIs automate important business tasks, I thought of a fun way to illustrate why this could be a serious problem.
Introducing the Stochastic Rocket Mini-Game 🚀
Just like LLMs, the rocket randomly deviates from its path and I've simplified the source of deviation into 3 levels, each level adding on to the other.
Level 1️⃣ Output Noise: Even if the direction is clear, the stochastic (non-deterministic) nature of LLMs will add some noise to the result
↳ 🚀 moves randomly around the center path.
Level 2️⃣ Objective Drift: In the real world, a direction is generally never 100% clear and may even change as things progress
↳ 🚀 path tends towards a moving target, straying from the center (correct) path.
Level 3️⃣ — Interpretation Noise: Although the stated direction is clearly defined, AI can misinterpret it (lack of context, mental model, bad judgment, etc.) leading to undesired and even potentially dangerous results.
↳ 🚀 path correction varies for same the same input, weak one moment and stronger the next.
Now imagine MORE dimensions, MORE uncertainty and MORE noise in a real world context with REAL consequences and you can see why letting AIs run aloof is a questionable proposition.
The good news is, there is a solution to this: Human intervention 🙋♀️🙋♂️
Smart human intervention into AI processes means we can have AI propulse 🚀 ourselves without arriving lightyears from our intended destination and also avoiding the asteroid fields on the way.
Hope you guys have fun playing the game, it also plays very well on mobile.
For the true gamers out there, would be interested to hear your high scores!