How Does ChatGPT Really Work? | Issue #23
+ how it does more than it is designed for
5 MINS READ
Hello,
The courteous thing to do is to apologize profusely whenever it’s sundown and I am not in your inbox.
Because it’s a DAILY newsletter for chrissakes.
Well, thank God 99% of my readers (I guess) are Nigerian and we know how power rolls in this black nation.
Hate to say it plain-speak,
Power played a fast one on me. Again.
Beating this would require I write twice as much and schedule ahead of time.
When you realize I also have to stalk Elon Musk and Sam Altman on X, you might have pity on me.
I had my power-hungry devices fed last night and pieced this issue 80% to completion.
Then I had to perpetually scroll to get all the gist about the coup at OpenAI.
Sorry. But we have enough gossip material for another issue.
For today, I’ll let you in on the open secret of how ChatGPT manages to write your assignments better than you do + how it is WAY simpler than you imagine + a lil’ bit more.
Have you heard of the AI that best the world’s best chess player in 1990?
That’s a perfect example of narrow AI.
If you asked that AI to play Scrabble, you’d unmask the dumb-ass.
That’s because narrow AI is very good at accomplishing one task or a set of tasks incredibly well. And nothing more.
For a long time, these AIs were the only things we could build.
General AI, on the other hand, can carry out a wide range of tasks as well as, or sometimes, even better than some humans.
We’re seeing a lot of these nowadays.
Over the past few years, some AI labs have successfully built AI models that are increasingly generic.
For example, you know ChatGPT can ace a math test and write a job application.
AI that creates stuff …
Generative AI is an example of general-purpose or general AI.
These AI systems can carry out content creation tasks and there are different types.
Text generators, like Bard and ChatGPT. Image generators like Midjourney and DALL-E. Audio generators like Elevenlabs and Jukebox. Video generators like RunwayML .
Now, let’s talk about the most famous of the generative AI tools:
Chat + GPT = omg
The name is a portmanteau1. A made-up word.
It comes from “chat” (because, well, it behaves like a chatbot) and GPT, which means Generative Pre-trained Transformer
GPT is the name of the AI model that ChatGPT uses.
Now, let’s take an even closer look at what GPT means:
G - Generative means this is a generative AI model that aims to well, generate content
This content could be text, images, video or audio as we’ve earlier described.
If all these things are generated simultaneously, it’s called a multimodal generative AI.
P - Pre-trained means that the AI system has already received training.
Training means it has been given millions of websites, books, and Wikipedia pages to read.
All this training data helps the model have some kind of understanding of the world and the connection between the parts of the content.
Well, you can’t feed an AI model to infinity. The training data will have to start from a particular time and end at a “cut-off” date.
T - Transformer is the name of the algorithm that GPT is based on.
This algorithm was invented by Google researchers and it was revolutionary when it was published.
Why?
When AI people feed computers with data, the model must figure out which part of that massive data is useful to each task.
That means it scans the whole of its dataset.
As you can imagine, this takes a lot of time and computing power.
With this algorithm, it means that a computer could focus on the most relevant information very quickly.
It could also process many tasks simultaneously to make the best use of computing power to perform its tasks.
Okay, but does ChatGPT really do?
There’s one thing a GPT is incredibly good at.
Guessing.
Really?
Well, the fundamental principle to how it works is to guess the next relevant word.
For example, if we ask ChatGPT what the capital of Nigeria is, it’d answer Abuja.
Smart-ass.
But how did it come about that answer?
It looks at it’s vast database and searches for the word that is most likely to come after. And inserts it.
At the risk of oversimplifying GPTs, they are auto-complete om steroids.
Wow!
Okay, how have things changed?
AI scientists made ChatGPT good at predicting what text was most likely to come next.
I don’t know how well they knew that this could unlock a few more superpowers.
ChatGPT can do some stuff which it wasn’t specifically coded to do as a result of it’s predictive accuracy:
Summarizing text.
Correcting spelling and grammar mistakes.
Rewriting text in a different tone of voice.
Translating a text into any language.
Brainstorming ideas.
Solving mathematical problems.
The appearance of these new capabilities came as a surprise even to specialist AI researchers.
So, in recap,
What we’ve discovered is, that when we:
feed an AI model with more and more data and,
increase its capacity to analyse the fed data
The model gets better and can perform some tasks that where not explicitly designed into it.
The raw power of well-utilized, qualitative data!
Tech like this is not showing any signs of slowing down. The (now-ousted) CEO of OpenAI, Sam Altman had announced that the tech lab was seeking for funds while it worked on GPT-5, a new, more powerful GPT model which he said will make the current leader GPT-4 look “quaint” in comparison.
What can I say?
Isn’t it a wondrous moment in history to be alive?
Prompt of the Day: Cover Letter Writer
Copy and paste this into ChatGPT:
Write a personalized cover letter for this job [JOB TITLE] at [COMPANY].Here’s the job description: [JOB DESCRIPTION]Here’s my CV or resume: [CV/RESUME]Tool of the Day
Use Guidde to create how-to guides, FAQs, SOPs, etc for your team 11x faster.
Did You Know?
Sam Altman (CEO) and Greg Brockman (Co-founder and President) of OpenAI were sacked by the company board in a surprise ousting on Friday. In an interesting twist, both have been appointed to spearhead an advanced AI research lab at Microsoft today by CEO Satya Nadella.
Image of the Day
Thank you for reading today’s issue.
I’m betting tomorrow’s gonna be a gossip feature. Lol
.
We’ll do more every day, innit?
With love and ink,
Emmanuel
A portmanteau is word that blends the sounds and combines the meanings of two words, for example the word podcast is a portmanteau of iPod and broadcast