Lights Out Office: Generic AI agent is not enough (or who is your daddy)
A generic AI agent cannot run a lights out office. It automates tasks instead of workflows, it does not know your domain, and most fundamentally it is not aligned with your business goals. It is aligned with its vendor's.
Key takeaways
- Claude Cowork, ChatGPT Work and Meta Muse automate tasks. A lights out office needs whole workflows automated, and that is a different problem.
- Agents today can complete some tasks autonomously to a certain degree of satisfaction. They do not always do the best job. Yet.
- Business workflows carry nuanced logic and case-specific decisions that a generic LLM does not have the domain knowledge to make.
- You need an agent fine-tuned or prompt engineered for your domain: who you are, what you are doing, what you care about.
- The fundamental limit is alignment. Facebook agents are aligned with Facebook. Claude agents are aligned with Anthropic. They call it the soul.
It is tempting to think that the problem of the Lights Out Office is already solved. Anthropic’s Claude Cowork, OpenAI’s ChatGPT Work, and now Meta’s Muse can do anything.
But that is incorrect.
These agents automate tasks. A lights out office needs whole workflows automated, and workflows carry nuanced decisions that need deep domain knowledge. And there is a third problem that no amount of fine-tuning fixes: the agent is not aligned with your business goals. It is aligned with the business goals of its vendor.
Transcript
Hey guys. Lights Out Office is a fully autonomous office, operating without human intervention for a prolonged period of time.
And some people are saying that this is already happening. You have Claude Cowork, you have ChatGPT Work, you have Muse from Meta, and they already do it.
That is not true. That is just incorrect.
Yes, you are able today to delegate some tasks to these agents. They got to a point where they can autonomously complete some tasks to a certain degree of satisfaction. And let’s face it, they don’t always do the best job. Yet. They are getting there.
But Lights Out Office is about delegating whole workflows. Automating whole business workflows. Not just certain tasks, but workflows consisting of a large number of such tasks, and very nuanced logic. Sometimes complicated logic.
And that is the main problem for LLMs and LLM-based agents. These are very nuanced workflows. Sometimes the decisions that you have to make are very tricky, and you have to know the domain.
So in order to make it happen, unfortunately, you cannot just use a generic LLM. You have to use an LLM that is either fine-tuned for your domain, or an LLM that is configured, prompt engineered, to work fine within your domain. To understand who you are, what you are doing, what you care about.
So essentially, you need an AI agent that will be aligned with your business.
And it is not enough to just understand your jargon. You also want this agent to be aligned with your business goals. Understand your business, and be aligned with its goals. That is the important part.
And that is why generic LLMs and generic agents don’t really work well. Or will not work well.
Because guess what? Facebook agents will be aligned with Facebook. Claude agents will be aligned with Claude goals, or Anthropic goals. They call it the soul. They configure their LLMs to be self-aware, to know who produced them, and to know who is the daddy.
And that’s a problem.
Sources and further reading
- Lights Out Office: part 1 defines the term and traces it back to lights out manufacturing. This video is the third part, on why a generic agent cannot deliver it.
- Claude’s Constitution is Anthropic’s own published description of Claude’s values, and it lists being compliant with Anthropic’s guidelines among the core ones. That is the vendor alignment this video is about, stated by the vendor rather than alleged by me.
Comments