A suite of programming tools powered by Large Language Models such as Claude, Cursor, and Gemini have nearly overtaken the software industry. These tools enable programmers to provide informal specifications for programs and will generate source code and documentation as output. The programmer can answer prompts the model creates to try and resolve uncertainty in the output. And these tools, some programmers claim, are our power tools.
The analogy goes like this: we used to cut boards by hand until we invented the table saw. If you’re a commercial wood-worker you use a table saw. It makes you more productive.
The Problem with Weak Analogies
Let’s imagine a table-saw is like an LLM. It doesn’t make a cut in a piece of wood. It generates one or two pieces of wood. You describe, in plain English, the cut you were going to make to the machine. It then uses it’s training data set, context window, reinforcement and alignment rules to generate the pieces of wood that most likely match those described by the input.
As a wood-worker using this LLM-table-saw… would you be more productive? You can’t trust that the pieces generated by the table are actually what you requested. Occasionally pieces come out as different species of wood. Or the cuts are not within tolerance of what was requested. When you use this tool you have to inspect every piece of wood it generates.
And then what if table saws were like LLM coding agents? Now a single wood worker can describe what they want a whole fleet of LLM-powered table saws to do. Only now the table saws have to be told how to inspect their inputs and verify their outputs as well unless you want to end up with a pile of unusable saw dust. The alternative is forcing the wood-worker to become the bottle neck to this whole system by having them inspect and every each piece of wood before it proceeds from the output of one LLM table saw to the input of another.
In other words, we exhaust the wood-worker or we force them to trust the tool and not be involved in the wood working.
That doesn’t seem like a good tool for a wood-worker. The mechanical table saw only required them to set the table once, check the cut once, and make many repeated cuts quickly. The LLM table saw required them to check each output piece methodically after every, “cut.”
The problem with weak analogies is that they seek to be persuasive through obfuscation. They rely on the assessor agreeing to the plausibility of the premise and not scrutinizing the argument. A strong analogy isn’t infallible to analysis either but the point doesn’t get lost upon examination like it does with a weak analogy.
Are LLMs Like Power Tools?
Use of power-tools in wood working is not as universal as most programmers using LLMs seem to think. They surely did transform the industry and the craft as they arrived. However we still make and use hand tools and use techniques and skills we learned before we started using power tools.
LLM-based programming tools are too different to draw any useful analogies from power tools. The work a mechanical device does for us is something we can inspect and intuit. Not so with an LLM: the output is generated by a black-box model and we can’t inspect the process (yet) that generated it. It can’t be deterministic, like a machine, because we would risk “over fitting” the output to the model inputs and have succeeded in making a really expensive copy-paste system. The system has to generate plausible output in order to fit into the category it’s been given: artificial intelligence. Companies building such tools go to great lengths to make these systems appear intelligent.
Such systems could be more “tool like,” than I am making them out to be. I don’t think it’s a simple matter of re-framing how we think about them… only because it’s hard to remove the other problems these tools come with: the political, ecological, cognitive, and social damage they inflict on society. But if we do change our expectations for what these systems are it might lead us to making better choices for the implementation of such a technology. And that might make them more useful tools in the future.