In writing this story today, note that I am coming from 20 years experience as an IT professional designing multi-million dollar data solutions for Virgin, BHP, Flight Centre and the like.
I understand the capabilities of technology and we leverage this daily within our own workplace. The current versions of AI, whilst great at some tasks, have some limitations in their approach and we need to understand and respect what those constraints are in order to get the most out of them.
Constraint #1 - GPT
We all know it's called Chat GPT, but did you know that GPT stands for Generative Pre-trained Transformer? That's important!
Let me nerd out for a bit here and break down what each part means:
So in short, GPT is an AI model that generates human-like text using a deep learning architecture called a Transformer, after being trained on vast amounts of text data.
Okay, time to stop geeking out with the techie talk. The thing to take from this is the "pre-trained" piece. Essentially, at a point in time it trawled the internet to gather its knowledge. Anything added after that point in time? It doesn't know about it.
So depending on what platform you're using, you need to be asking the question - what's the most recent knowledge date? In other words, what was the last date it was trained? That will give you context around the validity of the responses you get.
Why does this matter? Well, if new legislation comes out, or news announcements have been made recently, GPT models may not be aware of them yet. The next model or version will pick it up, but the current one won't.
#2 Access to Data
For GPTs to make informed decisions, they need access to data. Much of the data we require as property developers is behind a paywall. For example, access to data from RPData, PriceFinder, Archistar, Landchchecker and Stash. All of these very useful tools are behind paywalls that require ongoing subscriptions. They don't allow you to pull data out in bulk. So the ability to utilise this data at scale is severely limited.
That said, we can feed the GPTs new and relevant data, and get it to give us informed insights into that data. We need to keep training it specifically on the new pieces of information we become aware of. So the onus is on us to actually know what these new updates, legislative or otherwise, are, and where the data sets can be found.
Now we can ask questions that get better informed responses. For example, "Hey GPT, based on what you know, plus this new piece of information, can you give me an informed outcome now?"
Okay, so you would want to customise that question to cover off exactly what it is you want to know, but hopefully that gives you the idea.
#3 They're Designed To Make Us Happy
This one's very hard to pinpoint, and in some ways is the worst constraint because it can lead to answers designed to play to any bias it believes you have based on your question or previous interactions. Basically, it hallucinates on the answer, saying what it thinks you want to hear, rather than being factually based. In essence, it gives you the answer you're asking for.
As I said, that's largely because we've asked the wrong question. Learning how to craft a prompt and then doing what is called prompt chaining, where the output of one prompt becomes the input of the next prompt, is critical to get the most out of AI. In my observations, most people don't learn prompt engineering to the appropriate skill level required to get reliable outcomes.
If you do put the effort into prompt engineering, and if you do recognise it's missing some data and give it some extra data, and if you can aim it at some free data sources using the open data initiative (genuine global thing) then your ability to get reliable outcomes out of AIs increases exponentially.
There were a lot "Ifs" in that paragraph! But the effort to meet all those requirements is beyond most people's basic technical resources. And so they give simple prompts and in return get simple responses that don't have the depth of knowledge required. Bottom line, you should be sceptical of the results you get from GPT if you're not putting the extra effort in to give it the right prompts and data.
Where does that leave us as property developers? Personally, I believe that the data that is currently behind the paywalls of companies like those I mentioned earlier will at some point be accessible through their own versions of AI. Then we'll get our AI to talk to their AI to get the information we're after. But that is probably 12-24 months away from where we are today.
In the meantime, the classic phrase "Garbage in, garbage out" holds very true in the AI space. It originated in the early days of computing (around the 1950s–60s) to remind programmers that computers will process whatever you feed them, without judgement, so errors or nonsense in the input will inevitably lead to errors or nonsense in the output.
That maxim holds true no matter how good the system or process itself is, including AI. Always keep that in mind when you're using it.
I understand the capabilities of technology and we leverage this daily within our own workplace. The current versions of AI, whilst great at some tasks, have some limitations in their approach and we need to understand and respect what those constraints are in order to get the most out of them.
Constraint #1 - GPT
We all know it's called Chat GPT, but did you know that GPT stands for Generative Pre-trained Transformer? That's important!Let me nerd out for a bit here and break down what each part means:
- Generative: It can generate text (not just analyse it)
- Pre-trained: It’s trained on a massive amount of data before being fine-tuned for specific tasks or made available for general use
- Transformer: Refers to the Transformer architecture, a type of neural network that’s very good at understanding and generating natural language
So in short, GPT is an AI model that generates human-like text using a deep learning architecture called a Transformer, after being trained on vast amounts of text data.
Okay, time to stop geeking out with the techie talk. The thing to take from this is the "pre-trained" piece. Essentially, at a point in time it trawled the internet to gather its knowledge. Anything added after that point in time? It doesn't know about it.
So depending on what platform you're using, you need to be asking the question - what's the most recent knowledge date? In other words, what was the last date it was trained? That will give you context around the validity of the responses you get.
Why does this matter? Well, if new legislation comes out, or news announcements have been made recently, GPT models may not be aware of them yet. The next model or version will pick it up, but the current one won't.
#2 Access to Data
For GPTs to make informed decisions, they need access to data. Much of the data we require as property developers is behind a paywall. For example, access to data from RPData, PriceFinder, Archistar, Landchchecker and Stash. All of these very useful tools are behind paywalls that require ongoing subscriptions. They don't allow you to pull data out in bulk. So the ability to utilise this data at scale is severely limited.
That said, we can feed the GPTs new and relevant data, and get it to give us informed insights into that data. We need to keep training it specifically on the new pieces of information we become aware of. So the onus is on us to actually know what these new updates, legislative or otherwise, are, and where the data sets can be found.
Now we can ask questions that get better informed responses. For example, "Hey GPT, based on what you know, plus this new piece of information, can you give me an informed outcome now?"
Okay, so you would want to customise that question to cover off exactly what it is you want to know, but hopefully that gives you the idea.
#3 They're Designed To Make Us Happy
This one's very hard to pinpoint, and in some ways is the worst constraint because it can lead to answers designed to play to any bias it believes you have based on your question or previous interactions. Basically, it hallucinates on the answer, saying what it thinks you want to hear, rather than being factually based. In essence, it gives you the answer you're asking for.
As I said, that's largely because we've asked the wrong question. Learning how to craft a prompt and then doing what is called prompt chaining, where the output of one prompt becomes the input of the next prompt, is critical to get the most out of AI. In my observations, most people don't learn prompt engineering to the appropriate skill level required to get reliable outcomes.If you do put the effort into prompt engineering, and if you do recognise it's missing some data and give it some extra data, and if you can aim it at some free data sources using the open data initiative (genuine global thing) then your ability to get reliable outcomes out of AIs increases exponentially.
There were a lot "Ifs" in that paragraph! But the effort to meet all those requirements is beyond most people's basic technical resources. And so they give simple prompts and in return get simple responses that don't have the depth of knowledge required. Bottom line, you should be sceptical of the results you get from GPT if you're not putting the extra effort in to give it the right prompts and data.
Where does that leave us as property developers? Personally, I believe that the data that is currently behind the paywalls of companies like those I mentioned earlier will at some point be accessible through their own versions of AI. Then we'll get our AI to talk to their AI to get the information we're after. But that is probably 12-24 months away from where we are today.
In the meantime, the classic phrase "Garbage in, garbage out" holds very true in the AI space. It originated in the early days of computing (around the 1950s–60s) to remind programmers that computers will process whatever you feed them, without judgement, so errors or nonsense in the input will inevitably lead to errors or nonsense in the output.
That maxim holds true no matter how good the system or process itself is, including AI. Always keep that in mind when you're using it.
