If AI, as I previously claimed, is just a very clever predictive text, then why does it sound so human like and smart?
The answer is because it’s trained really well. You might have heard about companies like Meta and Google investing billions of dollars in AI farms, that use as much power as a small town. In those AI training centres the AI program is ‘learning’ to recognise patterns, and then creating rules based around those patterns.
Smart Washing Machines
Let me explain how an AI washing machine works and then you will see how this carries over to language models.
There are many different loads that you could put into washing machine. It could contain a sturdy linen or fine cotton. The clothes could be clean or dirty. There could be a lot of detergent or not much. With an old-fashioned washing machine you would need to select which wash you want to use based on what you put in it. But how does an AI washing machine work or as some manufacturers call them a smart washing machine machine?
Firstly there are various inputs that these washing machine machines can use. They can measure the weight of the clothes when the clothes are put in dry. They can detect how much water needs to be added to those clothes until all the water is absorbed and there is excess water in the bowl. Once the bowl starts filling they can agitate and stop and measure the resistance or the kinetic energy of the water and the clothes. All these things go together to give a sense of what is in the load. But how does the machine then know how to wash?
The answer is that the machine is trained in the factory. Lots of different loads are put into the machine, and for each load the machine measures the things it is able to. A human could help train it by telling it how to wash, or the machine could train itself by trying various things and seeing which worked best. Once all this data is gathered, the washing machine can make its own ‘rules’ about how to wash. This learning was done on one machine in the factory but it can can now be sent out into all the machines as they are sold.
If for example in the factory there was one pair of jeans in the machine, the machine knows the best way to wash these jeans. When you put one pair of jeans in your machine at home, the smart logic realises the kind of wash that it should do and adapts it for that pair of jeans.
The important thing here is that the washing machine is not thinking as it’s doing the wash at your house, it’s just reacting to rules that it learned back in the factory.
Language Models
An LLM (LArge Language Model) in an AI chat is very similar. It has done a lot of learning “in the factory” so to speak and then it uses that learning to know how to reply to you when you ask a question.
For example, if I fed an AI model every single Jane Austen novel that existed, it would be able to find patterns in the way they were written. It would notice what words Jane Austen uses together. If it is fed enough Jane Austen novels, it can make up rules about how Jane Austin writes. When I give it a word, it can use those rules to predict which words Jane Austen is most likely to use after that word. In it’s output it will mimic Jane Austen and attempt to write its own Jane Austen novel.
AI it’s not just trained on Jane Austen, it’s trained on every single piece of literature and every website that we have access to!
Once the AI training centre analyses lots and lots of writing, it creates some rules, as to how to write, and then it releases this set of rules and this is the AI model that you use. This is what a model is (e.g. ChatGPT 3.2). This is why early AI models did not have up to date data from the internet, the models were using data that was weeks or months old.
Now AI has improved to use live data as well as its models. It can do tricks like run a few models at once and then pick the best result. And different AI tools have been written to deal with specific data sets. You can train a model using legal documents or medical documents or any other area.
As AIs develop, each one becomes more specialised in a certain field. Chat GPT is a bit of an all rounder. Claude is known for the ability to handle more complex problems and good for coders. Claude has a specific model for Excel. Notebook LM is better for digesting information and giving summaries. Claude co-work has access to your local files (use with care!). We are moving on from the basic language models to more complex AI models that sit on top of the language models and use other systems as well to create the most realistic response that they can.
At the moment there are a lot of new AI tools being released, especially for specific applications.
Here’s a good article on which AI to use: https://www.oneusefulthing.org/p/a-guide-to-which-ai-to-use-in-the
My next article will be on some of the dangers of AI.
Leave a Reply