Tommy's Notes Repository

The Problem

6 min read

I cannot name a single job that I believe it would be unrealistic for AI to be able to do, outside of entertainment sectors.

Sure, when you’re talking to chatgpt, it just makes stuff up, but once you incorporate Multi Agent Concensus Seeking , and you look at how fast we are making progress on topics like that, it is easy to see how a AI models could be used to automate the entirety of the government and the entirety of each corporation.

Vocab

Here is the necessary vocab for the rest of this article:

  • LLM - Large Language Model. A model which is trained on text from all around the internet, and once it has read enough text, it can start generating text of its own, just by predicting the next “token” in a phrase, from the statistics it has on already written text. This is a flawed analogy for a number of reasons, but it is good enough for the writing of this article: think about your phone’s autocorrect. You type in hello and it will suggest ” There ” and “Mom ” and “Jake “. That’s a very dumbed down version of the idea behind an LLM.
  • Token - a segment of a word or whole word. Just splitting off whats common. Think: Revolutionized ? Revolution ized Language Learning Models. Language Learning Models .
  • AGI - Artificial General Intelligence. An AI which is good enough to do any intellectual task a human could do. We are not here yet.
  • model - gpt 3.5 is a model, gpt 4 is a model, an AI model is an AI approximating reality.
  • multi-agent - When you have multiple AI models working together to achieve one goal. Think about, one model is the idea guy, one model is the implementation guy, one model is the safety guy, and one model is the testing guy. The Idea guy purely generates ideas, and feeds them to the implementation guy. The implementation guy purely takes ideas and ouputs code. The safety guy purely looks over that code and sees if any of this stuff is morally unacceptable ( are we making code for malware, are we making code for baby head seeking bullets, etc… ). The testing guy takes code output, runs through a code sandbox, installing any required dependencies, if any errors occur, logs them, logs why they might be happening, and sends back to implementation guy. If sees core problem with the entire idea, sends back to idea guy.

So what if AI takes some jobs?

It would not just be some jobs. Capitalism incentivizes companies to take the cheapest route possible to get the best output possible. This is how you get a product to the consumer that beats the competition in price and in quality, and it is how you get the consumer to come back to you looking for more of that product. I am going to go through a few example jobs now, what they currently do, and how in my mind they are replaceable.

These are some jobs picked at random off this list of jobs: https://www.indeed.com/career-advice/finding-a-job/most-common-jobs-in-america which was the first result on my duckduckgo search for ” list of common jobs ”

  1. Cashier

I do not think this one takes much convincing. This doesn’t even need AI. Self Checkout machines already exist and are widespread. The reason we still have cashiers is partially that people steal. The Theft detection system needs to be automated, and really, it has been by certain companies. Look at the amazon fresh stores. The other reason we have cashiers is that people want to talk to people as they check out. Just wanting human interaction. Obviously im not suggesting AI can replace that ( really it can, but it SHOULD NOT ), but I think this want for human interaction isn’t enough to maintain cashier as a job.

  1. Food Preparation Worker

Robots. Legitimately. Obviously hugely expensive upfront, but long term way cheaper than employing an actual employee. Have a look at this article: https://www.maginative.com/article/google-researchers-can-teach-robots-new-skills-with-just-natural-language/

You can teach a robot to make a sandwich. You can then teach a robot to make pasta. You can slowly build out your arsenal of menu items from there.

  1. Stocking associate

https://www.businessinsider.com/walmart-adding-robots-help-stock-shelves-to-650-more-stores-2020-1?op=1

already being done.

  1. Bookkeeper Large language model takes input of expenses and income. Generates emails for customers for payment. Generates queries for APIs to accounting software and uses python with matplotlib to build reports, and then a seperate model does statistical analysis of those reports.

These would all have multi-agent verification of their outputs.

  1. Bartender https://www.culinaryunion226.org/news/clips/robot-bartenders-ces-2024-showcases-future-of-hospitality-sparking-job-security-worries

Already being done.

  1. Administrative Assistant A large language model can manage an administrator’s calendar and travel arrangements. All it would need is an integration with google calendar or apple calendar. Simple stuff. Also can do industry research and order office supplies, given proper integrations with the web and with a checkbook ( again, multi-agent verification of outputs, so it does not buy 9999999999999999 staplers. )

  2. Office clerk Large language models answering phone calls is already a thing. the google screen calls function built into my pixel is already wonderful. Large language models sorting through mail and filing confidential records would also be easy, if locally hosted. Context windows might be annoying, but it would just be an integration with paperless ngx hopefully. Compiling reports would also be simple for an LLM. Greeting visitors would be motion detection / pressure detection and a speaker.

  3. Operations manager A large language model can manage employees better than most managers any day. A large language model has all of the resources on management to ever exist at it’s fingertips. It will not micro manage, it will not lose respect by trying to be too chill, it will just manage. A multi agent setup would have one agent writing requirements for a employee to go under review, one agent writing protocols for how to respond to employee incidents / performance, one agent monitoring the employee performance and making graphs via matplotlib, and one agent dedicated to trying to figure out moral implications of each decision, researching what the people have going on in their lives, and deciding if their lowered performance is acceptable given the situation they are in ( If you have been employed here for 30 years, it is ok if your performance goes to a record low after your wife dies. Willing to accept that for as much as a year and a half, and then you’re gonna need to get it together. )

I got tired of doing these. If you have a job question let me know. the only ones in my mind that i cannot answer are entertainment related.

You get the idea though - huge sections of the economy are automatable and there are huge amounts of people working jobs that i do not consider sustainable for a company to continue paying for under a normal model of capitalism.