AI vs. Amdahl
LLMs aren't coming for your whole job. Just parts of it.
Ever since the debut of ChatGPT in 2022, AI has persistently been the biggest story in the tech world. The market still seems to be in a very frothy state, with fortunes being made and destroyed on promises of revolutionizing this or that industry. Promises of massive increases in economic productivity have been predicted, as well as concerns about AI-driven spikes in unemployment. Amusingly, the first class of jobs to be obsoleted by the boom in large language models seems to have been prompt engineer, AKA professional chatbot chatter.
The current generation of AI integration is so far fairly surface level and simplistic. You’ve likely noticed that seemingly every website has added some sort of chatbot companion or AI driven search feature. Oftentimes these feel as if they are awkwardly bolted on features that don’t integrate well with the existing site, often because they are basically wrappers around some commercially available Large Language Model such as ChatGPT or Gemini with some rudimentary domain-specific training.
But as systems improve and companies develop a better understanding of the capabilities of large language models, more sophisticated integrations of LLMs into the general operation of business are going to become more prevalent. These are the places where AI has the potential to cause major displacement in the workforce. However, while AI may eliminate the need for certain kinds of human work, it isn’t necessarily going to enable the one-for-one replacement of human workers with soulless machines.
To understand why, it may help to learn about a concept in computer architecture called Amdahl’s Law. This is a rule that there is a mathematical cap on improvements to performance that one can achieve based on throwing more resources at a task.
Say you have a program that automatically color corrects a video. The naive way of using it would be to have one computer go through the whole file and apply the color correction algorithm to each frame. But if you have lots of computer cores available, you can speed this process up by parallelizing computation. So rather than having a single computer color correct the whole video, you could tell one computer to correct the first half of the video, and then have another computer correct the second half at the same time, which should significantly reduce the time the color correction task takes. You might then suppose that you could make the whole process virtually instantaneous by infinitely subdividing the task such that you might have an array of thousands of computers which will each color correct one frame in parallel.
But this is where Amdahl’s law comes in to rain on your parade. While you may be able to get a significant amount of performance gains by running a bunch of tasks in parallel, at some point you hit a limit on the benefit you gain because there are certain bottlenecks in the process which can’t be efficiently parallelized. In the color correction case, this bottleneck comes in the form of computer memory. Even if you have enough computers that you can have each one correct just one frame, you still need to load that frame into memory and then write the corrected frame back from where it was pulled. Thus, no matter how many resources you throw at a problem, your performance gains are capped by the fact that you’re working off a single memory source. Thus, your ability to improve performance is fundamentally constrained by tasks that can’t be eliminated by the addition of more computing power.
This bottleneck situation is somewhat analogous to the question of AI-worker substitutions. With the current generation of large-language models, there are certainly aspects of jobs that can be done by LLMs without any major issue. But once you’ve AI-ified those aspects, you’ll be left with a collection of tasks that can’t be easily offloaded onto LLMs, which will still need to be performed by humans.
An instructive example might be something like a hospital. With the existing generation of AI models, you could find ways to automate tasks like appointment scheduling, health record generation, routine diagnosis, and inventory tracking. But there are other parts of working in a hospital where there’s no way to replace a human with an LLM. Even if ChatGPT or Llama are virtuosos at reading MRIs, they’re not going to be able to put in an IV or replace a bedpan.
But beyond the obvious issues of interacting with physical space, LLMs are also likely going to struggle to supplant humans in spaces that require the regular handling of irregularities or in places where a question is being asked and it’s really, really important that the right answer was given. Take something like software development. There are a lot of AI tools, like Cursor that can generate code based on prompts. But there’s a number of steps besides writing code needed to actually push a usable change. There’s integration testing to make sure that the program in question actually produces the expected output within its own context, as well as testing to make sure that it functions properly within the broader software ecosystem into which it is being deployed. If the AI generated code is going to be used in a context where performance is a major concern, performance benchmarking and optimization will also be needed. Finally, the code needs to be deployed and launched to users, which can be a difficult decision on its own depending on context. “Never deploy on a Friday” is an aphorism amongst developers for a reason.
AI is going to be risky to employ on its own in high stakes situations due to challenges regarding amalgamating information and incentive structures. Consider the case of a regional manager of a chain of stores deciding whether and when to purchase more stock of a particular item. This is already a task that many companies have semi-automated via the use of data models to predict price trends. But it is not fully automated for good reason. A hypothetical AI inventory purchasing agent may have greater insight and ability to predict price trends than our human manager can. But there are many other factors that go into making an informed decision that a naively implemented AI agent isn’t going to be programmed to consider. Say the AI predicts that, based purely on historical market data, the ideal time to place an order for more inventory will be at a time such that it arrives at the company logistics center next Wednesday. But imagine if, at our imaginary logistics center, there are two employees that usually handle processing that particular category of items. One of them is on vacation through Friday and another is coming back from parental leave next Monday. This isn’t the sort of knowledge that an AI is likely to have access to. In fact it may be information that is fundamentally impossible to present to an AI. It may well be feasible to plug things like employee schedules into an AI model, but is it possible to plug in every employee specialization and relationship with different aspects of the business?
Consider your own job for a moment. Whatever your formal job description may be, there are any number of impactful micro-tasks that you take responsibility for, just because historically you just happened to gravitate towards doing that task. Everyone knows the state of play, but it isn’t documented in the company org chart or anywhere that could be accessed and used by an AI to make critical decisions.
Besides questions of information, decision-making AI agents can also be a risky proposition due to questions of motivation. To some extent, all human workers are motivated to do their jobs well and make good decisions by the threat of firing for screwing up. AIs don’t possess that same gut instinct. We have all seen amusing stories of AIs hallucinating that Elvis is still alive or that glue is an acceptable pizza ingredient. Developers will try to use these failures to better train their models in the future. But when real money is on the line, it is going to be preferable to have humans that are inclined to proactively check their decisions for potential mistakes.
None of this is to say that LLM based tools aren’t going to be a key player in various work environments going forward. There will certainly be places where they can perform tasks better and/or faster than a human can, as well as other places where they can supplement human decision making with a broader scope of knowledge. However, it’s important to keep in mind that, for the foreseeable future, there are going parts of every job that require skills that AI does not currently possess, be that manual dexterity, flexibility in changing tasks, or just intuition on how to handle something unexpected.