Should you use an AI or your stupid monkey brain?

None of my articles are written using LLMs.

This article is to show how LLMs can be used and is designed so that non-technical people can follow. It argues against the narrative that all LLM usage is bad and encourages thoughtful usage.

A while back, one of my friends said they asked ChatGPT to multiply two numbers and it got it wrong. He said, "why would anyone use AI if it can't do basic math? It's stupid." This opened my eyes to how big of a disconnect there is between people with some technical knowledge on how AI and LLMs work and how they can be used effectively, and those who don't when talking about the environmental impact.

The problem computers help us solve in general is automating intellectual labor.

I think intellectual labor can be solved using 3 strategies: Human labor, traditional computing, and AI.

Traditional Computing: Calculators

Let's first compare human labor and traditional computing. We already know what humans are. But first, what's traditional computing?

At its core, a computer is a tool used to help us do calculations. To do math. Basic arithmetic. As we became more advanced as a society, we learned to stop using our fingers and brains and started creating tools to help store information. Our brains are spiking neural networks. (Very bad at math). We invented the abacus. It uses the position of its beads to help us remember the place values of digits and perform addition, subtraction, and multiplication.

Eventually, we needed to do more math. We started using electricity, and used vacuum tubes to store and process information rather than us doing it manually. We then moved to use smaller forms of storing information. We started using transistors and they've shrinking ever since, following an exponential decay in size, while computing power has grown exponentially. The observation that the number of transistors on a chip roughly doubles every two years is called Moore's Law. We're beginning to hit the floor as the size of atoms limit how small we can make transistors.

We can use traditional computing to handle tasks with low ambiguity. Let's refer to traditional computing as "calculators."

AI

If we want programs to 'learn' and 'think' on their own, we need AI. There are different types, ranging from simple machine learning models to neural networks to large language models (LLMs). At their core, they use traditional computing to recognize patterns in data and improve over time. Just like how simple neurons combine to produce something as complex as human thought, AI builds on basic computing to create sophisticated systems like LLMs.

AI: LLMs

Think of LLMs as text ingestors and text generators. They can take a small amount of text and produce lots of text as output and vice versa. The important thing here is that they can also solve problems requiring "intelligence." LLMs are able to solve problems outside of what training data they've been shown. This is exactly what makes humans smart. We're able to come up with creative solutions to problems we've never seen before. LLMs are getting better and better at this.

Keep in mind most LLMs are multimodal meaning they can take in and produce things like images, but let's keep things simple for now.

Humans

The set of tasks humans excel at has been very challenging for us to compete with in tech but the gap has been closing fast especially in recent years. Humans have an amazing ability to essentially learn any skill requiring dexterity, planning, and knowledge. Our image recognition (ability to look at something and instantly know what it is) is extremely impressive. We're able to convert our ideas into language and act on it. We can ponder our own existence and even create things smarter than us. There are definitely limits in our abilities: We're terrible at computational math (basic arithmetic), hence our reliance on computers to make up for it.

What's the Right Tool for the Job

Basic arithmetic

$graph showing math comparison$

LLMs and humans are essentially the same skill level here. They would either need to do the long tedious process of manually working through the computation by hand without help from a calculator.

LLMs and humans can excel at this task if they're able to use a calculator. An LLM, when given or prompted would be able to make a tool call to write a script or calculate a math question the same way a human would or pull out a calculator on their phone.

Coding

Software engineering has changed so much in just a few years. We're learning what software engineers are really supposed to do. Do they just write code? Are they problem solvers?

If you think about what LLMs are best at, being good at coding makes sense:

Input text → intellectual labor → output text

As a software engineer, what we do is take a task/problem and create a solution we can write in english. Then we convert that solution into code. Now, the job changes a little. We simply describe what solution we want at a high level and the LLM converts our english into code.

Using Tools

Humans have been using tools to solve problems for millions of years. If we want to cook food, we use our skills to create a fire or use a stove. Let's say we want to multiply two numbers. If it's something easy like 2x3 or 2x8 maybe we can do that in our head. But what if it's 123 x 214? Very few people can do that in their head very quickly so most people will decide to use a calculator. The same way we can use a tool like a calculator, we give LLMs tools to solve problems because even though it's a computer, it's a system that thinks like a human so it has trouble solving problems like this. If we ask it to solve a problem like this, we might get lucky and it'll call a tool to do this computation but sometimes it won't.

Summarization

The task of getting a large amount of text, using intellectual labor to compress the information into a summarization is something LLMs are best at. Humans are good at it but it takes us a really long time. Imagine you read a book and want to write a report on it. LLMs can do this task in less time and with much less energy.

What are LLMs the best for?

Are there uses for LLMs that are both more energy efficient and quicker than traditional computing and human labor?

Summarization
Translation
Sentiment analysis
Coding*
Pattern recognition
Classification

Coding — Really depends on the type of coding especially if it involves ideation. But at its core, using an LLM to convert a solution to a problem written in english into a programming language is a great use.

What are good and bad uses for AI?

Depending on the type of AI used, it's obvious there are good and bad uses for it. In terms of energy consumption, LLMs, image and video generation are extremely resource intensive. Using an LLM to summarize a legal document is different from generating graphic HD videos and deleting after use.

Image and Video Generation

Image and video generation are probably the most controversial areas in AI today. They're both very energy intensive (especially video generation). And not only that, they're very unpopular among consumers. Since it's so easy to produce AI images and videos, people spam it (even if its low quality) on their outlets and social media with low effort leading to people calling it AI slop.

Here's what I'll say about this: I think it's fine to use AI in these use cases, but people should just be mindful of its consequences. It's also wrong to claim AI-generated content is your own.

A gray area for whether to use AI or not

We know AI can speed up work, and we shouldn't use it excessively. But what about all the tasks it opens up to us to do? There are many tasks that we don't complete because it'll take too much intellectual labor that becomes doable due to AI. We'd need to decide what's valuable and what's not. What tasks are worth the extra energy usage and what tasks aren't. This is similar to commuting. Maybe it's worth it to travel on a jet to vacation once a year. But what about doing it excessively like getting a sandwich in the middle of the day?

Why are we scared of losing our jobs?

The range of problems AI can solve overlaps greatly with what we can accomplish with human labor. The set of problems humans can solve that can't be solved by AI is becoming smaller and smaller really fast.

Job Displacement

Job displacement due to AI, especially mass layoffs, is really damaging to our society.

It makes sense to introduce some type of automation tax where companies or individuals using AI to automate labor that was originally done by humans should be taxed extra similar to how taxes on carbon emissions work to protect our environment. We can also place a law that says companies have to put in "best effort" to let their workers increase their productivity with AI instead of laying off people and/or implement a cap or something where companies can't conduct mass layoffs due to AI or they have to do it slowly.

Energy Use

The total energy use will go up with AI, even if it makes us more efficient. Let's say someone's output without AI is 1x and their energy requirements are 1x. With AI, maybe they can make their output 10x and their energy requirement 2x. They're still increasing their energy requirement so total energy usage goes up but efficiency (output per unit of energy) increases.

What the government should do

Since frontier LLMs are a new technology, we don't have updated policies to defend against the problems they cause. Our policies will always be behind technology. Governments need to work overtime and stay ahead of the curve in the AI era. It's debatable whether or not AI is capable of causing mass layoff, but doing nothing is absolutely foolish. Whether the best solution is universal basic income or banning AI outright, we should start researching what the best solutions are. Humans evolved to spend time working. I don't think our brains are prepared for a consumption only lifestyle. We're programmed to do work and reap the rewards. Not sit and do nothing. We would get depressed due to lack of purpose.

We need the best and brightest minds in the government to create good policies to ensure we can smoothly transition technologically. I'm sure we're in good hands. 🇺🇸🦅🦅🦅🦅

What I didn't cover

The human brain uses roughly only 20 watts of energy, and yet, it's capable of incredible things.

I didn't go over any actual numbers of how much energy queries take. This is because it depends on the model, the effort level you give, the task, and even the hardware it's running on. AI does use a lot of water even though many data centers recycle their water. Some cool their centers through evaporation. These numbers are constantly changing and I wanted my article to be 1. Timeless in its content and 2. I think a general understanding is enough and 3. It's getting pretty long.

The human brain being so efficient and so capable tells me there's so much room for improvement. Also: AI doesn't have an energy limit like our brains did evolutionarily so we can theoretically make AI as intelligent as we give it more and more energy, which would allow us to solve harder and harder problems.

I will say it's pretty alarming that most frontier AI labs don't release their numbers on energy use and overall environmental impact. But why would they? I think the only one that does is Google. It's important to note that these companies are self interested and profit-motivated.