Meet Devin the AI software engineer, employee #1 in Goldman Sachs’ “hybrid workforce”

Goldman Sachs has just deployed “Devin,” a generative AI (gen AI)-based full-stack developer, Goldman CIO Marco Argenti told CNBC in a July 11 interview. “Devin is going to be like our new employee,” said Argenti, “who’s going to start doing stuff on the behalf of our developers.” Devin can execute entire apps from scratch based on natural-language verbal prompts. And while the banking world has been using AI-related tech for decades, Devin is the “first AI software engineer,” according to its creator Cognition.

Devin is part of Argenti’s vision for a “hybrid workforce” where AI agents work alongside their flesh-and-blood colleagues. Shanker Ramamurthy, Managing Partner of Global Banking & Financial Markets at IBM, offered an example of how Devin could be helpful to an institution like Goldman. “Think of it this way,” he told IBM Think. “If Devin increases productivity by 20%, then Goldman’s 12,000 technologists can now operate like 14,400 technologists.”

More than any other industry, “finance is where we are seeing the greatest value for AI creation,” said Ramamurthy. “After all, in finance, you’re not dealing in atoms; you’re dealing in bits.”

In recent years, for example, AI-adjacent fields like machine learning (ML) have enabled transformational advances in finance like algorithmic trading, high-frequency trading and the eternally thorny practice of options pricing. Devin’s role, Ramamurthy explained, is to narrow the technology and time gap between the complicated asks coming from a trading desk and execution.

To illustrate, Ramamurthy shared a hypothetical use case. “Let’s say there is a news item that someone found hundreds of millions of tons of very high-grade iron ore. If you’re a commodities trader, you can ask Devin, ‘Hey, help me incorporate this new information into my model.’” Given access to relevant data and models, Devin could rapidly work with a technologist to co-create or refine and help implement an application to process the request, streamlining and accelerating the software development life cycle.

One might well ask what the big deal is. Devin is just the latest in a long line of AI-based coding tools the global investment bank has used, including the much-lauded GitHub Copilot (not to be confused with Microsoft’s AI tool), which OpenAI and GitHub released in 2021. The world—or at least, the general public—had never seen a tool like GitHub Copilot, where a user could enter an engineering prompt in natural language, and the software would spit out functioning code. It seemed to be precisely the sort of task for which gen AI was invented in the first place.

Devin can do that too, but with the addition of a key functionality: autonomy. According to Cognition’s website, the company designed Devin to be a “tireless, skilled teammate,” with the ability to create an app from soup to nuts. As with Ramamurthy’s example above, the user can pretty much tell Devin their wish list in ordinary, natural language and get code that does exactly what they were hoping for.

YouTube has plenty of demos of people making apps with Devin.ai using the subscription-based, publicly available version. The interface looks something like a cross between ChatGPT and a Unix terminal, which is to say, half natural language and half code. The language part is the conversation the user is having with Devin; the code part is how Devin interprets your instructions under the hood.

You can connect Devin directly to other apps that might contain information helpful to your project, such as GitHub or Slack. In one video on a tech channel called Singh in USA, host Harnoor Singh demonstrates why this feature is useful: you can tell Devin, for example, to use the photo repository contained in a Slack conversation with your team, or give Devin access to a previous project from your GitHub account.

You approach Devin similarly to how you’d approach ChatGPT: you input, in natural language, a description of the app you want it to create. In Singh’s example, he tells Devin, “I want to create a stage timer”— the kind of countdown clock a public speaker can refer to on their laptop when giving a presentation. Devin is able to create Singh’s app with just three natural-language prompts.

While coding, it also does its own QA and debugging, a process that Cognition describes on its website: “Devin sets up the code environment, reproduces the bug, and codes and tests the fix on its own.” Pretty impressive; no wonder Devin has its own LinkedIn profile (the playful handiwork of Cognition Labs).

The Goldman announcement caused mayhem among IT professionals in the blogosphere, with one Medium headline reading, “Goldman Sachs just made your computer science degree worthless.” But according to Goldman’s Argenti, Devin isn’t there to compete with humans for jobs. After all, gen AI is only as smart as its user instructions. As he puts it, “Engineers are going to be expected to have the ability to really describe problems in a coherent way and turn it into prompts, and then be able to supervise the work of those agents.”

Ramamurthy agreed. “That’s not to say it’s 100% right and it’s got no gaps in it,” he said. “But really, Devin is about reimagining what’s possible. Instead of [the current practice] of taking a process which might be mostly analog and inserting generative AI into it, what if you can reimagine the process? You bring human plus digital agent together; magic happens.”