Generative AI is already changing the way software engineers do their jobs.
GitHub Copilot, Amazon CodeWhisperer, ChatGPT, Tabnine, and various other AI coding tools are quickly gaining traction, helping developers automate mundane tasks and freeing them up to work on more challenging problems.
A new survey from GitHub found that 92% of U.S.-based developers are using AI coding tools both in and outside of work, and 70% say the tools will give them an advantage at work. A majority also believe AI tools will lead to better team collaboration and help prevent burnout.
Another survey from Stack Overflow showed similar adoption of AI coding tools, with 77% saying they feel favorable about using AI in their development workflow.
In an interview with Wired published Tuesday, Microsoft CEO Satya Nadella said he realized AI was going to be transformative when he saw its ability to code, which led to the company developing GitHub Copilot.
We caught up with engineering leaders at six Seattle tech companies to learn about how they’re using generative AI and how it’s changing their jobs.
Diamond Bishop, CEO and co-founder at Augmend, a Seattle collaboration software startup
“AI is making it so small startups like ours can accelerate all aspects of the software development lifecycle. We’re a team of five devs, and we estimate productivity impact of almost 2X.
Each of us leverage both Copilot and ChatGPT in day-to-day development, debugging, and learning. Knowing the right way to leverage these tools and how best to be an AI-augmented cybernetic developer is a skill in itself, as there are pitfalls (hallucinated APIs/functions) and prompt incantations to learn. But it’s quite valuable in areas you aren’t already an expert, which is many in a jack-of-all-trades situation at a new startup.
Going back to coding without this augmentation feels a lot like working without the internet at your fingertips.”
Bridget Frey, CTO at Redfin, a Seattle-based real estate company
“We’ve already found a number of places where AI tools are making our engineers more efficient. For instance, we’ve used LLM models, including ChatGPT, with a fair amount of success to assist with internal tasks like migrating from one programming language to another, helping developers understand legacy code written by other colleagues, or writing functions for converting data formats. These are good examples of tasks that our engineers can do without the assistance of LLMs, but with these models, they can move much faster. Something that used to take an engineer 30 minutes to do can now be done by AI in one minute.
“Something that used to take an engineer 30 minutes to do can now be done by AI in one minute.”
We’ve also found ways to use these tools to help us serve customers more efficiently. For example, it’s not unusual for homeowners to contact Redfin’s customer service team and ask why the Redfin Estimate has priced their home the way it has. So we taught a LLM to explain to us in plain language why the Redfin Estimate may have priced a specific home in a particular way, and then we can pass those insights via our customer service team back to the customer to help them understand what’s going on.
In other cases, we’re still too early in the experimentation phase to see a significant impact but are running tests to see what benefit we might gain. One such example is with Copilot, from GitHub. We’re trying to determine if its ability to suggest predictive code might help make some of our more rote coding tasks a little faster and easier by helping with simple scripts.”
Jonathan Wiggs, CTO and co-founder at Outbound AI, a Seattle-based conversational AI startup
“We’ve recently leveraged GPT to write narrative summaries of phone calls made by our AI-powered virtual agents, and what we’ve found is that GPT is able to do this effectively the same, if not better, than an experienced human worker. It’s one of the things GPT does extremely well. We’re also using some generative processes to write code snippets, specifically for well-known API calls. Large language models are great at this kind of focused, pattern-based code building.
This said, it’s important to emphasize that we still need experienced and skilled engineers for 90% of our work. I think the companies that remain flexible and adopt these tools in a practical and thoughtful way will win the day for themselves and their customers.”
Laura Butler, CTO at Armoire, a Seattle clothing rental startup
“We recently started using GitHub Copilot. It’s helpful with generating much of the boilerplate for unit tests. The auto-complete and auto-suggestions in Visual Studio Code are pretty good, too, without being annoying. Intellisense and language plugins like Pylance have been around for a while. GitHub Copilot just takes it to the next level.
The biggest problem with generative AI and coding is that it works best with patterns and as a fancier copy-and-paste. But great engineering isn’t pumping out tons of code. It’s about less code, shared components, using APIs and services that already exist, quality, and maintainability. The world doesn’t need more Javascript files without comments. It doesn’t need more buggy code that wasn’t thought through, reviewed, or tested properly.
I think AI tools would be of real use for engineering in two areas: documentation and brainstorming.
Documentation and comments in code get out of date so easily or they are useless. Making it easy for busy programmers to write good docs and to keep them up to date would be wonderful.
Brainstorming assistance, generating lots of images and options from words to speed up ideation — that would be awesome, too. Those of us who aren’t graphic artists need a lot of assistance to visualize and share what’s in our heads before we start prototyping.”
John Zhang, vice president of engineering at Highspot, a Seattle sales software startup
“We’ve seen equal parts benefits and limitations using generative AI as a ‘co-pilot’ for engineering. It’s saving us time, especially for generic implementations or generating code for some testing scenarios.
Of course, we always check what it provides and haven’t seen value for complex implementations yet. We have begun scaling out its use across a broader range of engineers and expect greater value as we use it more.”
Kevin Leneway, principal software engineer at Pioneer Square Labs, a Seattle startup studio
“We started building apps using GPT during the GPT-3 private beta in 2020, and I was an early beta user of GitHub Copilot last year. Copilot was particularly helpful for many PSL tasks like prototyping around an unfamiliar API, or swiftly writing out boilerplate code or simple tests.
Once I got access to GPT-4 in March, I really started to explore how far I could push the AI to help me to build real working code.
Just for fun, I started building an AI coding “intern” named Otto. The goal is to have Otto interact with my existing dev and communication workflows just like any other dev on my team. I can assign issues in GitHub, give feedback via Slack, and do an interactive code review via pull requests. Otto can take on various roles; for instance, acting as a project manager by converting project briefs into a detailed list of GitHub issues, and even as a designer through a Figma plugin that turns designs into working components in my app.
When I first conceived the idea for Otto, I recognized that one of the initial steps would be to write all the code to interface with the GitHub and Slack APIs. Ordinarily this would have been a lot of tedious work to learn these APIs and get the basics working, but with GPT-4, I managed to get the core code written in just a few hours instead of days.
“It reinvigorated my joy of coding in a way that I haven’t felt since I was a kid tinkering with my Commodore 64.”
I’ve found that GPT-4 can efficiently handle the mundane parts, allowing me to focus on the higher-level planning and prompt engineering to get the whole project up and running. It reinvigorated my joy of coding in a way that I haven’t felt since I was a kid tinkering with my Commodore 64.
I should note, though, despite being a huge AI enthusiast, I also have some reservations about the future. I’ve dedicated the past 20 years of my career to software development and there’s definitely some fear that my amassed knowledge will become less relevant within the next 5-to-10 years. I’ve been building out my auto-coding bot slowly over the past few months and now that it’s mostly working, I’m realizing that I’m probably one of the first people in the world to truly understand how it feels when you have a customized, aligned AI agent who is capable of doing much of my day-to-day work. It’s thrilling, but also there’s a bit of sadness knowing that the skillset I’ve cultivated is gradually declining in usefulness.”
