Every type of artificial intelligence has a set of specific capabilities and a set of specific limitations. Neuro-semantic artificial intelligence with human teaming takes the best capabilities of each type of machine and combines that with the best capabilities of humans. The strengths of each is used to remedy the weaknesses of each and the sum of what results is better than any of the individual approaches used individually. The complete quadrant. The four agent workflow model. A multi-agent workflow.
Within teams, agents cooperate to improve the overall performance of the group. Teaming is the only way to safely deploy certain systems in complex, high-stakes, high-risk environments where there is low or even zero tolerance for error.
And this includes human intelligence, human agents, as part of the team. Human intelligence is an integrated component, not just a passive user. Besides, human intelligence and machine intelligence are different. My working definition of intelligence is "Human task performance."
There is one additional moving piece to this puzzle which is complexity and computability. Not everything is computable (e.g. there is certain work that a computer simply cannot perform). And so, when performing work which is typically part of a process, project, or workflow of some type; the process, project, workflow is described by some "algorithm" be that algorithm description manual or in machine readable form; it will always be the case that "partially algorithmic processes" exist or said another way; sometimes this will be automated but other times there will be manual steps to a process and a human needs to be involved.
And so, this is my personal working model of how I believe a collaborative work system will leverage "intelligence"; be that intelligence that comes from a human or intelligence that comes from a machine:
Basically, as I see it, work will be performed by leveraging the strengths of each provider of "intelligence" and leveraging the ability of one type of intelligence to overcome the limitations of the other types of intelligence. This will yield a "hybrid" that takes advantage of the best each contributor to this system. Further, the
better and more deliberate the boundaries of the system are defined; the better the system will operate.
This synergy between the multiple different intelligent agents is designed to capitalize on the strengths of each approach and to overcome their respective weaknesses.
The first I heard of
hybrid artificial intelligence was from Alan Morrison then of PWC. Alan had
this post on QUORA which led to
this slideshare presentation which mentioned
this video in which DARPA explains their philosophy of artificial intelligence. This was prior to ChatGPT and transformers and LLMs. DARPA basically suggest a hybrid of "hand crafted knowledge" via symbolic artificial intelligence and "machine crafted knowledge" via machine learning. DARPA did not add the human into the equation. The thing is that this DARPA thinking did not include humans contributing intelligence.
AllegroGraph explains neuro-symbolic artificial intelligence in this article, What is Neuro-Symbolic AI, with this graphic and they are not factoring in human intelligence either:
Perhaps both DARPA and AllegroGraph are factoring in the human intelligence via the "hand crafted knowledge" that humans instantiate in the form of machine readable knowledge.
To be clear, by "collaborative work system" I do not mean a "
collaborative work management" system. A collaborative work management system allows you to manage a project, not perform work for a project itself.
A collaborative work system is a system that is used to actually perform the work you need to perform. Part of such a collaborative work system is a collaborative work management system which can be uses to plan and manage the work. And the work I am contemplating is the work performed by accountants, auditors, and analysts. The
work relates to accounting, which is a deterministic system, and the collaborative work system is
industrial strength.
An artificial intelligence system is aligned when its goals, decisions, outputs, and actions (e.g. intentions) remain consistent with the legitimate human intentions, values, constraints, and authority applicable to the enterprise in which it operates (e.g.
governance), including under conditions which were not explicitly anticipated.
Epistemic risk is the risk of misalignment. Human first systems empower a human to manage alignment as contrast to artificial intelligence being responsible for system alignment.
Explainability is the "epistemic bridge" that lets governance verify whether an artificial intelligence system is actually acting in alignment with legitimate human intention. Without explainability, governance cannot perform its core function: validating, constraining, and authorizing artificial intelligence and human behavior.
Additional Information:
Comments
Post a Comment