Artificial intelligence (AI) has been a topic of fascination for generations with its potential to revolutionize our world. We dreamed of a future where robots did the work of humans and while this dream is not yet a reality, we’ve come a long way in terms of AI advancements. But have we really thought out how AI would work in practice? In this blog post, we will explore the history of AI and how our expectations have changed over time. Prepare to take a deep dive into the world of AI and see what the future holds for you. Keep reading to find out how AI works and how our expectations have evolved!Read more: How AI is Going to Change The World As We Know It!
How we thought AI would work
Up until now, everyone has really thought of a machine as something that sees an input and then responds in a pre-trained fashion. This is possibly due to the simplicity in which humans have perceived the mechanical world as being all 1s and 0s, and expecting a very binary behavior. Yet as technology is starting to become more mainstream we realize that how we are teaching the current stream of technological advances, we are allowing them to be much more creative. No longer do they behave in a standard and fixed way, and now they have learned to adapt, just as humans have, to present the same information in new and more advanced ways each and every time.
What is an LLM
So what is a Large Language Model? An LLM uses previous experience to try and guess the next word (or token) based on the previous tokens in the sentence. For example, if we were to feed it ‘the weather is’, it knows that the next word may be ‘nice’ with a 74% probability of appearing. A large language model speaks to a piece of machine learning that has been taught in a similar way that we might teach any other bit of math. Ultimately it is linear regression that goes on behind the scenes, which is just trying to solve an equation to pull out the next ‘number’ on a graph. We have been doing this for many years, with clever equations on how to solve for smaller numbers of parameters. In excel you might have put in a trend line. That’s linear regression when you only have 1 input.
Think of how we do sales forecasting, where we treat the amount of foot traffic, time of year, amount of staff working, and the weather as input parameters in an effort to guess sales. Each of these affects the predicted value. Large language models scale this up ridiculously. With GPT-3 using 175 BILLION different parameters, all ultimately trying to guess the next ‘output’, ie the next word (encoded as a number/vector/token).
Because there are so many parameters, this takes a HUGE amount of computation, which is why these models can require months of training with hundreds of high-powered processing units, however once trained, such a wide assortment of data, allow them to be incredibly clever. Interestingly when they were first devised, the idea was that they would be able to write the next word in a sentence, ultimately building up full paragraphs to text, but where they have become so context-aware, they have proven to be great at other tasks, for example, zero-shot classification. As an example of this, you could feed one the sentence “I am feeling very happy. The previous sentence has the sentiment of “, and you will likely gain ‘positive’ as a result. An unintentional but fantastic extra result.
How do LLMs mean we can do things differently?
When we think about how chatbots of old work, or even how your amazon Alexa might work, it’s usually listening for keywords. The example I like to give is ‘the weather is’. There are a number of ways we can use this sentence and have the resulting sentence be completely different based on context. Here are some examples:
- The weather is – nice today
- When talking about a biome, the weather is – only one aspect of this.
- I think we will have a great snowboarding trip as the weather is – snowing.
- Can you cancel my trip scuba diving as the weather is – causing terrible visibility?
Previously when we think about chatbots, they like to respond to keywords. This is where they can fall over and become very obviously a bot. For example:
- You have completely ruined my holiday, ‘thanks’ for nothing – You’re welcome
- I want you to cancel my jetski, the ‘weather’ is awful today – The weather today is raining
- I never want to speak to that lady in the ‘spa’ ever again – Connecting you to the spa
Obviously, all these examples highlight where problems can arise when we look at keywords alone. Using the new LLMs also allows us to handle situations that we might not have previously expected, but the chatbots can interpret in clever ways, and handle them correctly.
Why Creativity was the first thing they learned
Because these models are not just playing pattern matching, they are very context aware. This means that whilst we might have taught them how to continue writing, they’ve also learned what to do in similar situations. This is how a human learns language. To begin our lives, we probably have no idea how to tell a story. But after we hear them more and more, we figure out the patterns, what works, and what doesn’t. It’s a similar thing with art. Give a child a pen and paper and they will make a horrible mess, but the more they are exposed to ‘what’s good’, and practice, the better and better their art will become. The machines have learned in a very similar way, and due to this its allowed them to be creative, emulating what they have seen before as being correct, whilst still adding their own flair to the final output. It’s incredible, and just think this is only the very first round o their creations. I’m incredibly excited about what this might mean for the future!
Why it isn’t going to take your job, only make it better
So the common question I hear is, it’s coming for my job. It’s not coming to replace you, it’s coming to make your life easier. When the first mobile phone got used by the public sales people were worried it would mean they would be less required as rather than needing to travel around to deal with customers, they could quickly speak to them where ever they were. This didn’t mean fewer salespeople, it simply meant one salesperson could speak to more customers and ultimately be more productive.
When photoshop came out, it could suddenly do things that most artists and photographers could only dream of. It could remove people entirely from pictures and leave no evidence, it can do incredible things! Did that mean those guys no longer have a job? Nope, it actually meant they can do even more incredible and amazing things.
Now out comes technology like Article Fiesta. Does that mean copywriters no longer have a job? No absolutely not, it just means that rather than having to spend all day only being able to earn money for one client, you can bust out enough content for 10 and just spend that time editing, ultimately allowing you to do your job even better still. Reduce the amount of time needed for research, and power through those contracts!