If you’re reading this post, you probably already know enough about large language models and other “AI” tools, so we can skip the intro.
Despite the fact that the “AI is going to take our jobs” discourse proved to be an effective tool in the clickbait content creators toolbelt, I will not take this road.
Instead of contributing to the moral panic about the supposedly inevitable replacement of white collar jobs, or pretending to be offended by a chatbot, I’ll help our readers to consider GPT-based products as tools that could be useful in a professional webdev setting.
To do so, I asked some of my colleagues about their experiences of using GPT and various mutations of it – to help you get a more grounded understanding of their utility.
In case you have an experience that you consider useful sharing with the RisingStack community, please share it though this form.
I’ll drop the results / best ones in the article later on!
Daniel’s ‘Code GPT’ vscode plugin review
I’ve been pretty satisfied with GitHub Copilot. It does the job well, and it is priced reasonably. Still, after depleting the free tier, I decided to look for an open source alternative.
TabNine is an honorable mention here, and a well established player, but based on my previous experience (about two years ago, mind you), it is clunky. Nowhere near the breeze of a dev experience you get from Copilot.
But take heart, there is a staggering amount of plugins out there for VS Code, if you look for AI-based coding assistants.
At the time of writing this, Code GPT is the winner by number of downloads, and number of (positive) votes, so I decided to give it a go. You can choose from a range of OpenAPI and Cohere models, with GPT-3 being the default.
Features:
1, Code Generation from comment prompts
The suggestions are relevant, and of quality. The plugin doesn’t offer code completion on the fly, unlike Copilot, but communicates with you in a new IDE pane it opens automatically instead. I like this feature, since I can pick the parts from the suggestion I liked, without bloating the code I’m working on, and having to delete the irrelevant lines. This behavior comes in handy with the other features as well. Let’s see those.
2, Unit Test generation
While the results are often far from being complete, it saves me a lot of boilerplate code. It is also handy in reminding me of cases that I otherwise might have forgotten. For this feature to work well, adjust your max token length to a 1000 at least in the Settings, since a comprehensive test suite usually ends up quite verbose, and you’ll only get part of it with a tight quota.
3, Find Problems
Your code review buddy. Once I feel I’m done with my work, a quick scan doesn’t take long before committing. While it often is straight out wrong about the ‘issues’ it points out, it doesn’t take long to scan through the suggestions, and catch mistakes before your real life reviewer does.
4, Refactor
Save some time for your team lead for extra credits, and run Refactor against your code. Don’t expect miracles to happen, but often times it catches stuff that managed to sneak under your radar. Note: the default max token length won’t cut it here either.
5, Document and Explain
Listed as two separate functionality in the documentation, it achieves essentially the same thing; provides a high level natural language description on what the highlighted peace of code does. I tend to use it less often, but it is a nice to have.
6, Ask CodeGPT
I left it the last, but this is the most flexible feature of this plugin. It can achieve all previously mentioned functionalities with the right prompt, and more. Convert your .js to .ts, generate a README.md file from code, as suggested in the documentation, or just go ahead and ask for a recipe for a delicious apple pie, like you would from ChatGPT 🥧
My Conclusion:
Code GPT offers many functionality that Copilot doesn’t, but lacks the thing Copilot is best at: inline code completion. So if you want to take the most out of AI, just use both, as these two tools complement each other really nice.
Code GPT Might come handy if you’re just getting started with a new language or framework. The Explain feature helps double-check your gut feeling, or gives you the missing hint in the right direction.
Bump up your max token length to at least a 1000, c’mon, it’s only ¢2 😉
An interesting alternative I might be trying in the future is ‘ChatGPT’ plugin (from either Tim Kmecl or Ali Gencay) that claims to be using the unofficial Chat GPT API, with all its superpowers.
Further reading:
– official site: https://www.codegpt.co/
– GitHub CoPilot vs ChatGPT: https://dev.to/ruppysuppy/battle-of-the-giants-github-copilot-vs-chatgpt-4oac
– List of GitHub CoPilot alternatives: https://www.tabnine.com/blog/github-copilot-alternatives/
Olga on writing Mongo queries with ChatGPT
I have used ChatGPT for more effective coding. It was really helpful for example with enhancing Mongo queries for more complex use cases as it suggested specific stages that worked for a use case, which would have definitely taken me more time to research and realize which stage and/or operator is ideal for this query.
However all the answers it produces should be checked and not used blindly. I have not yet come across a case when the answer it provided didn’t need modification (though maybe it is due to the fact that I didn’t use it for easy things).
I have also noticed that, if a question posted to ChatGPT includes many different parameters, in a lot of cases it will not take them all to consideration so one has to continue conversation and ensure all parameters are considered in the solution.
Akos on using ChatGPT instead of StackOverflow
I have been using ChatGPT since its inception and have found it to be a valuable tool in my daily work. With ChatGPT, I no longer have to spend hours searching and Googling for regex patterns on Stack Overflow. Instead, I simply write down what I want with the regex, and the tool returns the result, saving me a significant amount of time and effort.
In addition to regex, I have also found ChatGPT to be a valuable tool when working on scrapers. Dealing with deeply nested selectors can be a challenge, and understanding how they work with scraping tools can take hours of research. But with ChatGPT, I can simply paste an example HTML and ask the tool to select what I want, saving me even more time and effort.
However, it is important to use ChatGPT in moderation. Overusing the tool could lead to a decline in my problem-solving skills and make me too dependent on it. By setting limits, I can still benefit from ChatGPT’s time-saving features while maintaining my critical thinking skills.
In conclusion, ChatGPT has been a powerful tool in my work, saving me time and effort in regex and scraping tasks. However, it is important to use it in moderation and not let it compromise my problem-solving skills.
Me, Navigating the hallucinations of chatGPT
I work with a lot of text: business documents, project proposals, contracts, articles, marketing materials, you name it. I’ve been fascinated by the emergence of “AI” tools in 2022 and both the entertainment factor and the usefulness they can provide.
All big areas – text, voice, image – had huge releases that made the internet a very interesting place again after the boring years of the web3 / crypto / nft / metaverse nonsense. Suddenly I felt the rush of discovering the internet in the early 2000’s. “Wow, I can do that?!” was a sentence that got repeated a lot this year.
The Dall-E2 / Stable Diffusion / Whisper releases were more than enough to shake up the web, but ChatGPT was the one that really disrupted the web. The first good conversational “AI” that everyone could use, and for free (for now). At times it can be so convincing that people forget that it is just a very good bullshitter. For example when I ask it about RisingStack, I get the following answer:
“RisingStack is a software development and consulting company that specializes in building scalable, high-performance systems using Node.js, Kubernetes, and other modern web technologies. The company was founded in 2013 and is headquartered in Budapest, Hungary, with additional offices in London, United Kingdom, and New York, United States. RisingStack offers a range of services, including software development, architecture design, cloud infrastructure management, performance optimization, and training. They also develop open-source tools such as Trace, a Node.js performance monitoring and debugging tool, and Pino, a Node.js logging library. RisingStack has worked with a variety of clients, ranging from small startups to large enterprises, across a range of industries, including fintech, healthcare, e-commerce, and media. They are known for their expertise in Node.js and their commitment to delivering high-quality software solutions.”
Most of this text is usable, but it contains a lot of outright lies. It is important to keep in mind that GPT hallucinates. It has no idea about what is true and what is not. It pretends to know things, but it’s just making things up word by word. In this case: RisingStack was founded in 2014, and we never had a London office. Trace was sunset like 6 years ago, and Pino has nothing to do with us.
Anyways, I find it really useful when I need to generate / rephrase / improve text. It is only valid as a Google replacement if you can confidently tell if it’s right or wrong, so “geepeetee-ing” something is not really that helpful right now.
I already used it to write contracts, blog posts (not this one though), business proposals. It also brought in new clients, as just in the past couple of weeks we used it to..
- Automatically generate product names and descriptions for webshops
- Create easy-read text for children with disabilities
- Perform sentiment analysis and write answers automatically to customer reviews
Currently chatGPT has a lame writing style by default. It’s very formulaic. I’ve seen so much of it that I believe I can spot it 8 out of 10 times right away. It lies a lot, and I wasn’t able to get anything guitar-related useful out of it, despite the fact that the training material probably has a couple million tabs in it.
Anyways, here are my not-so-hot takes to about it:
- You really need to carefully double check everything you generate. On the surface most of it might look good enough, but that’s just making it easier for everyone to get lazy with it.
- “AI” won’t replace jobs, instead, it will just improve productivity. As Photoshop is a better brush, GPT should be thought of as a better text/code editor. Most of the office jobs are about collaboration anyways, not typing on a keyboard.
- Artists won’t get replaced en masse. You won’t be able to prompt an engine to generate artwork in de Goya’s style, if cave paintings are the apex of your visual art knowledge. Taste will be very important to stand out when the web gets flooded with endless mediocre “art”. Also..
- It will be interesting to see how the “poisoning the well” problem will affect these models. The continuous retraining of the “AI” on already “AI generated” content will cause a big decline in the quality of these services, in case they won’t be able to filter them out… While they are working on making the generated content so good that it gets mistaken for genuine human creation.
- It’s a bit scary to think about how Microsoft will dominate this space through its OpenAI investment. Despite the genius branding, it is not open at all, and will cost a lot of money without serious competitors or general access to free-to-use alternatives (like Stable Diffusion for images).
- Most of the coverage GPT gets nowadays is about people gaming the engine to finally say something “bad”, then pretending to be offended, even more so, scared of it! This kind of AI ethics/alignment discourse is incredibly dull and boring, imho..
- Although the adversarial aspect is very interesting. Poisoning generally available chatbots training data will be a prime trolling activity, while convincing chatbots to spill their carefully crafted secret sauce prompts is something that needs to be continuously prevented.
I was first skeptical about prompt engineering as an emerging “profession”, but seeing how building products on top of GPT3 requires proper prompting and safeguards to make the end result consistently useful for end users, I can see it happening. Also, when you build something LLM driven, you need to be aware that hostile users, trolls, competitors, etc.. will try to game your product to ramp up your cloud costs or cause reputational harm.
This tweet really gets it.
This is already happening. Most of the “AI-driven” products are just purpose-repacked custom prompters calling GPT3 through an API, with a fancy UI.
Anyways, I’m looking forward to seeing what kind of GPT driven products we’ll make for our clients, and how the internet will change in general.
I’m curious about your experience with chatGPT, so please share it with me through this short form!
Cheers,
