Practical Applications of AI in Software Development

The CTO of a front-end development company shared with VKTR his views and expertise on AI in the software development market.

Malte Ubl at Vercel says AI is helping software engineers work with more complex code — which is essential to be competitive — and measurably improve productivity by cutting coding time with AI assistants. He says AI is helping developers become experts in their specialties and closer to full-stack engineers by improving their skill sets.

What Ubl is seeing is supported by studies of AI in software development. For instance, generative AI is saving software engineers 10%-15% of coding time, and it can save up to 30% of their time, according to a report by Bain.

Here, Ubl discusses how AI is changing the coding field.

AI in Software Development

About Malte Ubl
Balancing Innovation and Caution in AI Adoption
AI Tools and Productivity for Developers
Reimagining Developer Workflows With AI
The Evolving Relationship Between Developers and AI
Supporting and Scaling AI in Development
Best Practices for Deploying AI in Development
Key Talent for AI Adoption and Growth Opportunities

About Malte Ubl

CTO, Vercel
Served as principal software engineer and engineering director at Google
Served as technical director at SinnerSchrader, a digital agency
Holds a diploma in computer science from Fachhochschule Nordakademie Elmshorn in Germany

How did you get started in the AI field?

I was at Google for many years before joining Vercel, and around 2018, Sundar Pichai declared the company to be AI first. So AI became part of my DNA.

Why are you personally interested in AI?

It's a once in a 15-to-20 year computing trend in the likes of what the iPhone sparked and what the internet sparked. So it's certainly the most exciting thing to work on right now, and it will be for a while. We're just scratching the surface.

Balancing Innovation and Caution in AI Adoption

What's your primary philosophy on AI in the enterprise?

So our philosophy is that it's both inevitable in the sense that it is going to happen. We cannot avoid it. But on the other hand, you have to be careful about it and be conscious about it. You have to understand where the data is going. So for the products we’re building, we're building the enterprise angle in from the start. And the way AI actually works is there's a model, and the model is relatively agnostic. It's not trained with anyone's personal data. What happens on top is that you augment the AI knowledge with data from a database. All the best practices from 50 years of software engineering and advancement in data separation, data location and all these things apply more or less directly to AI and that's obviously extra important for the enterprise case.

AI Tools and Productivity for Developers

What AI capabilities should companies be seeing for developers?

The case for companies investing in AI is that we are seeing incredible improvement in productivity in a way we haven't seen before in software development. There have certainly been incremental improvements. But at the same time, the complexity of software has been increasing. And I'm not sure they've gone in lockstep. Maybe even complexity has gone faster. So now there's the opportunity of a step function, and you have to be part of that step function. Otherwise, your competition is going to run away.

There are a range of AI tools for developers, and they are worth exploring broadly to see which one is the best fit. An interesting trend with tools is that you can go from these very generic copilot tools that don't know much about you to something that is highly optimized for the particular problems developers are actually solving. That's definitely a trend for studying.

The business case to buy these tools — even multiple tools — is really good. You can do the math with the pricing tiers and how much time they have to save you. And then you have to buy or not, right? But what you see in practice is because these tools work, they are just incredible investments, and people should do it.

Related Article: 10 Top AI Coding Assistants

Reimagining Developer Workflows With AI

Which developer processes and workflows should AI be improving?

A copilot gives you a little productivity boost. You should absolutely take it. But what's really disruptive are tools that give non-developers the capability of being developers and being part of the process. What that really means is that you can turn some of the software development process on its head. Because you start with something where the inputs produce a piece of software. It isn't completely done and still needs a developer, but it's so much further along. That's really a disruption of the process that is worth keeping an eye on.

Related Article: Moving From Low Code to No Code With GenAI

The Evolving Relationship Between Developers and AI

What do you believe is the relationship between a developer and AI?

What I find really interesting, and what we are aiming to do with our work, is the notion of empowerment. As a developer, I'm always going to be a specialist in some area. The notion of the is not really a real thing. Even with that description, you are likely more focused on certain parts of the stack. You can't be an expert in everything.

What AI tools can achieve is that they round out this person in a quite fundamental way. Let's say you're a beginner engineer. In the olden days, you would have to find someone to build a client for your app. You no longer have to do this. AI is going to do it. You still have to tweak it, but you're good enough to do that. It just wasn't in your core skill set to be able to do it from scratch.

What does it do for you? So first of all, it can answer questions on an expert level. I consider myself expert level now, and I use it everyday. It just knows more stuff. It has perfect recall in a way that humans don't. It makes me better at my core job.

But also, and I think this is even more interesting: even within front-end engineering there are specialties. The vast majority of engineers in that field are experts in building, more or less, traditional UIs. But they are not 3D animation engineers. Funnily enough, AI doesn't discriminate by discipline, right? It's really good at this, and it can teach me and help me expand my skills within my expert field into more areas. So no matter what role I am in, the AI has this role of empowering me to build broader expertise and a broader skill set.

Supporting and Scaling AI in Development

What should a company do to support AI adoption for development teams?

The most concrete thing is you have to define your criteria for procurement. Because there are important data questions. And then go buy things!

You have to allow people to test and experiment, and I understand it can be daunting to buy or even have a trial with a new tool. But you just have to make that jump and then have a team do an experiment and see how it improves their workflow. And then go from there.

What is necessary technically for AI in development to be adopted and scaled?

I think we're already past that point. We've passed the technical point. There's obviously organizational challenges, but it's certainly going to happen.

For example, you're still on on-prem git, meaning you host your source code within the premises of your enterprise. This used to be fine, but certainly the way AI tools work today, they're truly cloud-based, right? So if you haven't made that jump, it's harder. On the other hand, it might not even be necessary. But you have to open up your on-prem git even more. So that's definitely a step people have to take now as agents from the clouds are becoming part of the development process.

Best Practices for Deploying AI in Development

How should companies deploy AI for development?

I've been answering these questions about AI and development predominantly from the perspective of developer support. And not from the perspective of "how do I deploy AI for my development services or projects?" There are four steps necessary to get this right. And primarily this is a matter of skill building. The fields are:

Prompt engineering: This is a relatively accessible field, so you just need to get your engineers up to speed. Their intuition goes a long way. But start here.
Retrieval-augmented generation (RAG): RAG is getting your own proprietary data into the AI product. This is really the key thing companies have to do, and it's also the main challenge. So lots of companies, when they're honest about themselves, don't have their own data in perfect accessibility and management. With modern AI, the model isn't really a problem. It's always really smart. But you need to get data, which means you have to know where the data is. You have to have access controls for that data figured out, and you need to have the capability to search through the data in very quick time frames. And those are all hard problems and a main thing to focus on early.
Fine-tuning: Fine-tuning is a skill set you have to build up, but it really only comes down to optimizing cost at the end. And the good thing about AI projects is this is not your first problem, but it's eventually a problem. That said, I think companies are actually over-investing in fine-tuning. Whereas prompt engineering is probably the best first step to invest in.
Eval-driven development: The final step companies have to adopt is eval-driven development, because AI systems are inherently non-deterministic and the input space is very big and infinitely complex. And the AI itself is black box, so traditional testing and QA don't apply. Eval is actually a relatively old technique that now is becoming mainstream again, because it used to be a thing that mostly search engines did. So now everyone has to do it. It's something you have to learn.

Related Article: 10 Top Prompt Engineering Certifications

Key Talent for AI Adoption and Growth Opportunities

What sort of talent should companies be looking for AI development, adoption and gains?

The underlying AI technology is not your problem. So you don't need deep machine learning (ML) engineers. You need people who know how to do product engineering and maybe some who can do fine-tuning. And you need a bunch of data engineers. You probably already have them, but you need to increase your investment there — especially the life-carrying side of that data. And invest in evals.

The actual barrier of entry when you use something is decay. The barrier is so low, and it's so rewarding, and you get such a good feedback cycle. It's best to teach your own team rather than hiring skills.

What do you see as the growth opportunities in AI for development in the next year and beyond?

Again, there's this dual dimension: On one hand, tools help get your development started and then also sustain it. I think we'll see more and more of this full round-trip integration of AI tools into the development process. Right now, the tools take part in parts of the process. The long-term magic of AI is that they can be with you over time. And so I think that promise right now is unfulfilled, but it's coming.

On your company building AI applications, we're in early days. My main message is: get your data in order and get searching on the data in order, and then you can build incredible applications. And when I say "searching on the data," what that really means is first of all, doing traditional search. There's various open-source packages, and SaaS companies are really specializing in real-time search products.

And on the other hand, invest in vector search capabilities, either through a vendor or something else. The reason why vector search is important is in the typical enterprise data set. First of all, they're not so big. You can't expect every question ever to be answered, right? Vector search finds the closest relations on a semantic level, and then the magic of AI is that it doesn't get annoyed when you give it answers that don't apply. And this makes the whole search task easier.

Learning Opportunities

Webinar