low hit points

the future of AI/LLMs

There is a significant debate about whether "AI" (or LLMs) as of today will replace a lot of jobs.

Prerequisites for Building with AI

If you are building something that uses or leverages AI, there are a few realities to consider:

  1. You need to know what, fundamentally, an LLM (or AI) is doing (when solving that task you expect it to solve), for which you should have done that task by yourself many times to build a deep intuition. a lot of people are thinking that by not knowing this, you can build something great, I don't think so. often the devil lies in the details and i don't believe you can build something great by being abstract from the process.
  2. You need to have done that task many times to be able to instruct LLMs or write agents in the best possible way to get a great output.
  3. LLMs as of today can only implement or do what you ask them to do. To a certain extent, they can be your thought partner, go a step forward, and correct you. However, LLMs have yet to reach a state where they can figure out what has to be done by themselves.

The Calculator Analogy

Imagine an LLM as a calculator:

  1. You still need to learn addition to instruct what the calculator has to do.
  2. You still need to instruct the calculator on what it is supposed to add (the inputs?). Although, in many cases, LLMs are able to figure this out, but yet, you need to ask LLM to figure it out.
  3. Calculators are 100% accurate, but with LLMs, this is also an issue. LLMs still can't be trusted with 100% accuracy in most cases.

Who and what are LLMs going to replace then?

They will likely replace jobs where workers have to follow a given set of pre-defined instructions with 0 or minimal ambiguity. LLMs are almost at a point where they can simply follow a flowchart and get the results for intermediate steps within it.

In fact, LLMs are going to speed up the jobs of people who are great at coming up with execution plans—short, tactical but detailed plans woven into long-term strategies. The more detailed the plan, the better the execution by LLMs will be.

This suggests a few ways to remain irreplaceable in this world:

  1. Be someone who can handle ambiguities and uncertainties, and who can come up with experiments or ideas for problems that have not been solved before, because LLMs still can't do that.
  2. Be someone who is so good at something that they have an intuitive understanding of the niche. In other words, this can be called having a "deep taste" for something. An example would be world-class chess players who, even in difficult positions, can most of the time intuitively come up with the best or good moves without having to calculate.
  3. Maybe more... not sure.

Probably, these are some reasons why LLMs are still failing at computational/deep logic/time series activities like chess and other games?