INTRODUCING SENTIUS
~2.5 million years ago, humanity predecessors, hominins, came up with a smart idea: stone tools made for various uses. The very idea of creating them meant that hominins had to create a mental model of cause-and-effect relationships in their minds, such as how striking one stone with another would create a useful shape. Creation of tools also involved planning action based on abstract goals like preparing tools for hunting or cutting.
~200 thousand years ago, early humans invented coordinated hunting of large animals. This required humans to be able to model the behavior of prey, sharing mental models between the members of the hunting group, and using tools and techniques to manipulate both the environment and the animals.
~50 thousand years ago, early humans came up with the idea of using speech to describe representations of the world between each other, connecting distant locations or times, and transferring knowledge across population.
Each time, our predecessors came up with mechanisms that enabled humans to achieve what they could not achieve before by observing the environment, forming mental models of observations, predicting possible outcomes, planning series of actions towards solving a given goal, communicating plans across groups, and then successfully executing these actions.
Today, we have large language models, and, arguably, they possess a limited ability to model the world, predicting possible outcomes, planning actions, and then using tools to perform those actions towards successful task completion.
Soon after the initial excitement of ChatGPT, a new big dream emerged in early 2023 in Silicon Valley and beyond: Autonomous Agents.
The promise is simple:
We ask using natural language these agents to perform tasks - anything we want (!) - and then they harness the power of large language models and tools given to them to do just that.
Sounds simple enough, right? Add exponential increase of power of large language models to solve complex problems, and we essentially have a magic wand. Easy. Or is it?
Turns out, there are myriads of caveats. Agents do make mistakes, just like us. The longer the task is, the more mistakes accumulate, and the less chances are that that agent can successfully complete the task. The more complex the task is, the less chances are that the model can come up with the right solution.
Planning turned out to be a problem. Reasoning turned out to be a problem. Steering the way agents perform turned out to be a problem.
