AI Agents: Human Synthesis vs Human Substitutes

Human synthesis vs human substitute, this appears to be the two core perspectives for accessing GenAI when it comes to Agentic execution. 

From a human synthesis perspective, you observe that for the first time we have broad computational access to what were previously uniquely human capabilities. At the moment this can be enumerated as the ability to speak, see, hear and understand information. When hominoid robotics are mature enough, we can add the ability to use limbs to this synthesis.

Prior to the emergence of LLMs, if you had a process that required any one of these capabilities, you needed to plug a human into the loop. It didn't matter whether the human had any particular expertise or not, if you needed to know what's in a picture or what someone said, you needed a person integrated into your process. 

LLM/GenAI have changed this, we now have computational access to what I would describe as human synthesis. This is different from human substitute, which is what folks pursuing AGI/Super-intelligence I presume are after.

I think this distinction needs to be made more salient because at the moment a lot of the conversation around AI capabilities tend to present them in terms of human substitutes as opposed to human synthesis.

The distinction is also very helpful to those building AI solutions. I think it is an effective way for builders to constraint themselves and gain clarity on what the nature of the AI capability we currently have access to is, and how to effectively leverage it.
 

Human synthesis means these capabilities are ultimately disjointed, meaning you can't treat current AI agents as human brains inside a computer as many are currently attempting to do. 

Instead you think of AI agents as providing access to those previously exclusive human abilities and you can now use them within computerized processes in a piecemeal manner without requiring a human being in the loop....A sort of synthetic human if you will.

Let's consider one common example for autonomous agent use cases, the Insurance Claims Adjuster.

It is a popular use case example for AI agent solutions. I don't work in insurance claims processing but I suspect a large fraction of the non-human part of the process can be satisfied with conventional automation. 

Two places where I can think of injecting AI capabilities, analyzing the claimants claim, ie their account of what happened, which is likely to be in a free form informational format, ie text, audio, video.

The second place being analyzing evidence, which would likely be media, though it could also be corroborating statements from a trustworthy entity. 

Both these cases fall under the human synthesis characterization, ie they are requirements that previously would have required a human in the loop because the ability to understand information, to see an image or video, to hear audio and understand it all were strictly human, with GenAI that's no longer necessarily the case.

It is however important to clearly demarcate and understand where AI fits in this solution, instead of the Agentic sales pitch which would rather talk about an Insurance Claims Adjuster as if she is a lady living inside the computer that you can now leverage as you please. 

Human synthesis in other words does not equal Human substitute.