My first encounter with GPT and how it might be a life changing one
TL;DR: I share why I think GPT-4 (and tools like it) can be life-changing, explaining how it helped me build simple apps and websites quickly. I found GPT-4 to be a powerful educational tool that can improve productivity, creativity, and maybe even transform society. I briefly acknowledge potential risks and ethical concerns — although in this article I focus on the positives. I am starting a newsletter where I will touch on these topics in more detail. Sign up for my newsletter to get a short practical guides on building prototypes with ChatGPT and other AI systems, explore the potential impact of AI on the workforce, economy and day to day life.
For the past few weekends, I dove into the ChatGPT craze. I never really got into Crypto or the Metaverse, as it didn’t really make sense why I’d use them, I couldn’t really apply to a problem or need in my life where they were the obvious solution.
However, one ambition I’ve had for nearly 10 years, that I haven’t really been able to satisfy yet, has been to build apps or websites for ideas I happen to have, and there is always an idea that I am mulling over. Typical cliché millennial “idea guy” you could argue.
This has been further fuelled by a person I have been following on Twitter for a few years, that does exactly this. See Here: https://twitter.com/levelsio — he travels around the world, building websites which created an income and lifestyle based on it. This showed me how it’s possible to quickly build on an idea, validate it and move on if it’s not the one.
A key issue has always been building out the developer skill set to do this. There’s a long list of half-finished programming courses and tutorials that I never really got round to finishing, and even if I did, with a full-time job and life, there is little time to learn and build something meaningful.
Rather than the lack of skills, time is my main issue, how do I learn fast, and learn something specific to what I need to build, and then quickly validate whether to put in more effort or not. Towards the end of last summer, I took a short career break to do a data science bootcamp for 3 months, which was great in terms of learning and focus time, but also required taking the risk of not having an income for a while, so not something easily repeatable.
A few weeks before GPT-4 launched, I saw this tweet, someone built a fully functional web-tool using ChatGPT — it’s a simple tool, but one that does a simple job end to end. And this got me excited, unlike Crypto and Metaverse hypes before, ChatGPT felt like a perfect solution for a long-standing ambition.
I quickly then went over to chat.openai.com, paid the $20 subscription and I set off trying to build a Chrome Extension. Why a Chrome extension:
Perceived them as simpler to build
Wanted to create something that generates data to practice real world machine learning on (completed a Data Science bootcamp last November and wanted to generate some data to keep practicing on — seemed my browsing data and scraping websites I visit, would be a good start)
And then for the extension to serve as something to deploy trained models on for whatever uses I end up discovering.
I initially started a few weeks before GPT-4 was due to be released, using GPT 3.5. It felt like a game changer, but not without its issues and need for troubleshooting. In the 2 weekends I was using GPT 3.5, I could easily build an extension that simply downloads my browsing history to a CSV. However, I found on more complex tasks, like building a job board type website, with log-in and multiple accounts, the need for troubleshooting increased significantly. Insert Tweet meme here.
I eventually resorted to use ChatGPT as a learning tool, I’d take the output of a more complex answer, and ask it to explain it further. Thinking it might not be able to generate a full project, but I could use it to guide myself to learn enough to be more self-reliant, and importantly learn only what I need to complete a specific side project, to keep to my time constraints.
Based on this first encounter with GPT, I took away to following points:
1. It works, it really works, but it’s best when you already have some knowledge and skills in an area to enable you to ask specific questions and then deal with any necessary trouble shooting.
2. Break stuff up & keep your questions closed ended. Meaning writing your prompts incrementally and being specific as to what you want to get back.
3. Tooling and interfacing can be improved — we could really use an AI Powered IDE and a change management tool to track changes back to specific prompts. Having to copy code out of the Chat GPT interface and into a code editor became quite messy after the initial set-up, where I started losing track of changes made, which changes I made related to which prompt and so on. Eventually felt quite overwhelming. Thankfully tools such as
https://www.cursor.so/
are already springing up to help with this, see demo here and Github’s Copilot is integrating GPT 4 next — announcement here.
4. Could be a great educational tool, but within a structured flow. Imagine a tutorial where you build-along but you also get to go off script when you don’t understand something. If there is a general structure to get back to, I found that it was incredibly useful as an educator.
5. I can see how it could turn into a big opportunity for companies to give their current workforce specialised AI tools — and tech consultancies to build bespoke solutions for it. I applied it to a coding task for simple side projects, but I can see how giving employees AI tools to help them in their roles, could improve productivity, output, but also unlock people’s creativity as they can now try things they couldn’t before as they may not have had certain skills or knowledge, where now the AI can step in. I can also see how tech companies and consultancies can help business, identify which roles and use cases they want to build bespoke tools for and build them (although also soon, I can imagine the employees can build their own tools with the help of AI systems)
Note on text interface
What makes this so easy to use for me is the text interface. A text interface is so familiar to all of us and has such a near flat learning curve making it highly accessible. Although perhaps not in all cases or for all audiences, especially where fine tuning may be a factor. Insert tweet about prompt vs. technical interface. In this example, the users of a AI photo app, prefer sliders and buttons over text prompts, but I do wonder if that due to the audience of this specific use case.
As you might imagine, I’ve been quite insufferable, talking non-stop at anyone around me about this thing since I started playing with it. Including my parents, my mum has a small Etsy shop, but as English isn’t her first language, writing exciting product descriptions is quite difficult for her. I told her she could try ChatGPT for it, and she was excited but was asking me to do it for her, thinking it was something complex, until I showed her how it works. And this is where I see it being very power for non-technical people to enhance their skillset and output quality near instantly.
GPT-4 came along
When GPT-4 was released, I achieved a better output, in minutes, for the same tasks. Mind Blown.
First thing I did was to ask it to build the same chrome extension as before, but with a few more features for bookmarking which I couldn’t quite get right with GPT-3.5 a few days before (thinking it’s an improved version wanted to see if it can handle the things its predecessor struggled with).
The output worked first time as intended without needing any manual changes, while I’m not trying to hype this thing, but it’s difficult not to. In the previous attempt the UI was not good and many things, such as buttons required manual interventions to get them to work. Not this time, the UI was half decent, and everything worked as expected.
Out of the 5 points above, I think they still all hold true for GPT 4, even if it can give you more quality code output, it still needs you to know how to then structure a project, get your dependencies sorted out and running the app you’re trying to build. However, for each of those things you can ask it to give you step by step instructions, so you can still expect and assist from it.
In the short it gives me time, I can do so much more with my time outside of work, without needing to completely sacrifice my personal life or other interests. I can learn and build at the same time — big win.
I am excited and quite hopeful of what this can do for us in terms of transforming society, however, I can also see that there are huge risks around people’s jobs, at least in the short term before people can fully retrain. And I have also seen more philosophical arguments around the nature of being human and what gives us purpose, which AI might threaten. So, there is an important discussion to be had around how we use it responsibly.
This feels like a steam engine moment, of my lifetime at least, but I think more so the digital age (although we can say the same about the internet). But unlike the steam engine, maybe we can figure out the equivalent pollution question earlier.
I personally see this as a potential tool to unlock people’s creativity, and productivity, allowing us to focus on tasks that give us more enjoyment and a better balance to life.
I have many more questions and thoughts I want to explore further:
Can we increase individual productivity, as one person can access a broader set of skills, therefore with fewer handovers between specialisms, can they move quicker at higher quality — Highly proficient generalists?
“Enhanced” humans — one person “armies”
Some research on labour market impacts https://arxiv.org/abs/2303.10130 — not all positive short term, but opportunities for new better jobs categories
Humans might still need to be QAs for quality and appropriateness of AI output
Is there a role for highly specialised individuals identifying and training/fine-tuning AI for specific use cases?
A.I. And economic growth — are we going to get more of it?
Equitability! Teach everyone to use it and for free — those who use it will have a huge advantage, so we should get at many people knowing about it, and show how it could help them and how to practically use it
Giving people access to skills that were previously prohibitive either due to cost or time — you don’t need a powerful laptop, you do need to be able to pay for GPT subscription, you need a bit of a time commitment, but you can do more in the same amount of time and we as a community need to build the content for others.
Are we going to work more or less — to match current output or exceed it by keeping to same work schedules? — do we continue with 20th century work culture or find a new way. Can we use it to maintain the economic growth agenda of the past decades, while also developing new economic approaches, as we might have more time to devote to it, without affecting our current system until we have a good alternative.
UIs and interfaces as a growing area of investment — Will there a be a rush of new start-ups built to address specific use cases for AI, each with their own specific interface suited to the job they’re doing.
What happens when quantum computers start running AI models?
What are the real costs of running AI services, is it viable at scale. Monetary costs are already dropping, see here — but what is the limit?
Is the work less produce more scenario likely, can it happen? What’s the value to economy and society of giving back people’s time and hopefully unlocking their creativity. Can there be a positive impact on health and quality of life without sacrificing economic growth. What are the pitfalls?
I’m planning to write short practical guides on how I built my prototypes with ChatGPT, alongside exploring the questions above — let me know if there’s something specific, you’d like to see. To get a notification when I put out new content please sign-up to my newsletter.