LLMs are tools maybe

Rich Hickey has a post arguing that LLMs are not tools. He defines tools as something you wield. I'll buy that. It's Rich Hickey, I'll buy most anything he says. I'm not entirely sure that discounts LLMs from being tools though.

Here is my proposed tool definition: "An item that affects a subject's ability to enact change through leverage."

If we consider the teleological meaning of tool, it would revolve around the intended function of a tool - what a tool is designed to accomplish. In this way, a tool is evaluated by its specific nature, the purpose it was built for, and that it were to be 'employed' towards.

Since everything is teleologically defined in the context of purpose, the distinguishing aspect of a tool appears to be that it is employed. Which is quite close to being wielded. I see where this is headed.

Screw it, here's my revised definition: "An item that, when wielded, affects a subject's ability to enact change through leverage."

So, we can agree that a tool is wielded. This does not, however, exclude LLMs from being tools. LLMs can be wielded. Right? Let's check in with Rich:

"wield - to use (as a tool or instrument) especially with full command or power : handle, manipulate, control" - Rich Hickey, either quoting somewhere else or coining.

"Handle, manipulate, control" point towards the ability to command. Degree of command is a spectrum, and LLMs land at different points on it depending on how they’re used. "I used an LLM to write my article" is acceptable and coherent expression. "I wielded an LLM to write my article" is not. And yet, wielding is defined as a type of use, only "especially" with full command or power.

Can you "use" a search engine? Yes. Can you "wield" one? Arguably. If your experience and ability with using a search engine to achieve an aim enables higher performance compared to someone with less experience and ability, an argument could be made that you wielded it.

LLMs can certainly function as search engines. This could imply there is a spectrum, if one wants to put searching as the lower end of capabilities offered by an LLM.

As we climb higher in the spectrum, we somewhere would hit the act of learning. Can a learning service be considered a tool? Can you wield it? You can certainly use it.

I like the leverage part of my definition for tool because it feels true. But we can't even get to that. There's no way past the current impasse of verb invocation. There is no way to say whether you can or can't definitively "use" an LLM, let alone "wield" one.

But let's skip to leverage.

If we accept leverage in the tool definition, then it can be gained through the wielding of tools. It can also be gained through the delegation of tasks.

Let's take this from another angle: can you engage with a tool, or a tool's domain?

If we consider a programming language as the substrate upon which the tool of the text editor rests, would we say we engage with either? I would say not.

Do you engage with an LLM? Yes, clearly.

Is there any scenario where you could be said to wield an LLM, vs engaging with it?

As long as interfacing is done in natural language, it would seem not. Which is quite strange - you could input the same query into a search engine as an LLM, get quite similar answers (discounting search engine AI summaries), but see it as "wielding" the search engine while "engaging" with the LLM.

What if you include expert level shortcuts and a dsl in the search query? That could be seen as wielding.

But what if we consider employing skills, harnesses, mcp servers, et al when engaging with an LLM as an agent? Certainly far more complex than learning a search engine DSL.

We engage with LLMs, and always will. We direct them. We delegate to them. But it doesn't seem that we can wield them.

An agent harness though, and its associated concerns..what do we say to that?