Large Language Models and Our Use of Them
The usage of Large Language Models or LLMs has become ever more prevalent over the past few years. Whether you like them or not, feel they're unethical or not, they're here, and for better or worse, they're not going anywhere. This leads to an ever growing problem: offloading tasks to these seemingly magic “do anything” services. The concept of offloading tasks to an LLM in itself is not a problem, but the way that we as a society are integrating these services into our lives is.
The most obvious way that we are failing to integrate them correctly is in the business space: attempting to replace employees with these LLMs. Some companies are already realizing that this simply does not work, and more will follow.
“But I thought LLMs were replacing junior engineers?” you might ask. Well, that's great in theory, but really doesn't work in practice. Let's assume a company can hire a junior engineer for $80,000/yr. This is probably on the low end, but just bear with me here. That junior engineer is mostly autonomous, able to work off of a backlog of issues or with pretty minimal instruction. Now, for the cost of an LLM: looking at ChatGPT, you can get a 150-seat (the minimum) enterprise subscription for about $108,000/yr. You might be saying “wow, that's 150 junior engineers for only $108,000/yr! That's huge savings!”. However, you're forgetting the other cost to an LLM, the engineer who needs to drive it. If a company is no longer hiring junior engineers, we can assume they only have mid, or in many cases, senior engineers to drive these LLMs. That's an engineer who is costing a company anywhere from $150,000 to $250,000 per year. Due to how error-prone and how much hand holding an LLM needs, that senior engineer is now spending their much more expensive and valuable time driving the LLM to do the work of the not-hired junior engineer when they could be spending their time working on tasks that actually require their senior level experience. I acknowledge that any good team will require senior engineers to do some amount of mentoring of junior engineers, and that's fine, but the amount of their time that will be required for mentoring purposes will be much less than driving an LLM, and in my opinion, will be more impactful for the organization in the long run.
So where do LLMs fit? Well, they fit as a tool. I'm not saying the cost of an LLM isn't worth it to the company, it can certainly be useful to use it as a tool to augment the workflow of an engineer with any amount of experience. LLMs are tools, just like IDEs, language servers, docker, or any number of other improvements that have been introduced to programmers over the years. The only thing that really sets LLMs apart from any other tool is their use across many, many domains, and not just software development. It is irresponsible to try to replace humans with the LLMs we have available to use today, and given the slow-down in the advancement in this technology, it is very likely that will be true for a while to come.