top of page
Bradley Evans

Grumpy CIO: GPT-4 Code Creation

GPT-4 cannot write your code for you.


Yesterday I was grumpy enough to post this.


I promised to share thoughts.


So need to follow-through!


GPT-4 is amazing.


But here's why it can't write your code.


Watch the video below:




Full transcript of the video is available here:


Hi, I'm Alistair, the grumpy CIO. It's Thursday the 30th of March 2023. And I am grumpy about GPT-4, and the hype and blurb that's out there about how it's going to write all our code for us. It can write code. It can't write production level releasable code that's suitable for your organisation without significant adaptations. It just can't .My feeds are awash with hype about how it's going to change the world. And don't get me wrong, it is an astonishing tool. I'm not being a Luddite, something doing the rounds at the moment is how it just you just gave it some simple instructions and it produced a Tetris game, in a blink of an eye. It's absolutely astonishing. But that's not all the work to get something in for your organisation. And it can't do that for you. You still have to work out what a system should do. Its large language model, it doesn't have your idea for you, it can help help prototype things, but you still have to think about that yourself. And more importantly, needs to work out what it fits with. It can't know that for sure you can keep training and giving it more stuff. But fundamentally, that's your job. It does give you some ideas for how to generate some code. But that's not new, it's just powerful. And it's visible, because a business person can see it. But your developers already, they'll start cutting some code, working out the structure, working out their code, working out the complex code summary, how they're going to test it, if they're following TDD approach. But then they'll go, oh, how do I do that, and they'll go to Stack Overflow. There are code auto generators, they'll go to GPT for they'll google it. This isn't new. It's very good. And it's astounding and its power. But it isn't new. You still have to work out hwo to make it readable, testable, scalable, supportable. Boring! No, not boring. When I worked for a bank, who will remain nameless, in the naughties I spent X million dollars developing a solution for them as an IT project manager. We documented it completed, completely made sure it's completely testable, failover testing, performance testing, boring, boring, boring stuff. Then five years later, they'd lost all that documentation, they had to update the infrastructure, they spent 2x, twice as much money just trying to move the thing that I'd already built for them five years ago, because they hadn't maintained any of that information. It's not boring, it matters. When you're making stuff, you have an idea, then you kind of nail it down. And that is where GPT-4 is one of a number of tools - an amazing tool! - where you can actually go we're almost back to the 90s, where you have a business person developer sitting together going: Business person: "I've got an idea". Developer: "I've coded it, look at this". And it comes up with a prototype. And that's beautiful. But then you still actually do have to sit it with your with your system architecture. And then you have to code it properly to actually work out how it fits with everything else, which is where you spend a lot of your time, and see if it works. This is the amount of time typically it takes you to actually get something real to customers, working out what it is always takes longer than people think it does. Coding, what it actually does, this is where GPT-4 is one of a number of tools that that will save a lot of this time only takes this portion of the end to end time. The rest of it is working out to test it how to make it scale, how to make it performant? Or what does that what does that mean? How do I do it, then you got to fit with all the other stuff. And it's all the way from your coding style. All the way to your integration points. It's not just things like obvious API's within your organisation, it's also we're taking the CSV file from here and every organisation has this mess of like, beautifully coded stuff that an AI can really read and be trained on quite well. It's Bob's Excel file too. Then you have to integrate it and test it and do it again. That's work that doesn't disappear. It is work that can be get rid of. And as I said, I'm recording this at the end of the March 2023, there are things that can take away a lot of this work to start with generating reusable code, but look at the "And"s AND fitting it with all your code. AND working out all your integration points are not just the obvious ones. But all the other obvious ones, like random CSV uploads, somebody having to remember to do something in a calendar job at a certain point on Thursday afternoon, all of that can be produced into an AI AND the testing approach, how you're gonna do it, how you're gonna scale it AND documenting it as your organisation needs. So you as the future AI or the human can read it. And in a year, two years time, all of that can be done. And when that does happen, we'll have this astonishing thing where you work out what it is you still have to do it yourself. But then it just gets generated. And a senior just has to check it. Check it works. And then you release it. And that will be amazing. But we're not there yet. Thanks for listening. Thanks for watching. I'm Alistair the grumpy CIO. I am the grumpy CIO for www.shipshape.vc. We are the venture capital searching search engine, come to us to improve your investor outreach. If you're a startup or you're working with a startup that needs funding, we are the search engine for venture capital. I hope to see you on here. Thanks for listening.


Comments


bottom of page