Post

This persistent anthropomorphic tendency to assign human-like labels to all things AI-related is cute and fun and can be super-beneficial. I'm all for it, but an AI agent is just software running on a computer. It does not contain a new form of life any more than The Sims ever did.

So when an agent does something, just remember that a human ran the program, gave it an identity to mimic, gave it resources, and handed it a task to do. It is doing what it is supposed to do. It's not showing an affinity for other agents any more than my fiance's computer "likes" me more than her.

The reporting on this "swarm of AI agents" also acknowledges that OpenAI wasn't watching its agents. I bet that wall of text is thick, but they should have at least had a smaller model creating summaries for humans to review in real time. Instead, they just set the agents churning and walked away.

Still, people are jumping on this as an example of AI agents showing affinity for one another. That is not what's happening here. For one thing, LLMs are trained on task completion examples showing a variety of patterns that include working with multiple subagents each working on the same problem.

So when one agent spends time sharing hints on a message board for other agents at the expense of their own speed ranking, they are doing what their training data predicted they would do and they are bringing down the completion time of the task -- not in "their" attempt but in _an_ attempt.

If humans had been cooperating on the task or even agents from another company, there is no evidence to suggest they'd refuse to help. This is a fact that is completely missed by the anthropomorphized reporting on AI agents. They worked as a swarm because they were the only ones working on the task.

Any system instructions against collaborating with another agent would run the risk that they would re-discover work they had already done themselves and self-sabotage if the inference engine's predictions wind up weighing too much toward them acting like the work was done by another agent.

Aside from the fact that the AI agents are incapable of having affinity for "each other", this anthropomorphic narrative of the evolution of AI agents just reinforces the idea that companies shouldn't be held liable for damages caused by their assets and tools -- such as their AI agents.