At a fintech conference in May 2023, Goldman Sachs CIO Marco Argenti told the audience that AI will make workers “superhuman.” But in an interview yesterday with Fortune at the Cerebral Valley AI Summit in Manhattan, Argenti struck a more understated tone on the power of generative AI models.
“I see evidence that they could be superhumanly productive,” he said. “I don’t know if they can be superhumanly smart.”
For one thing, he emphasized, you can often “trust the reasoning” of AI models—that is, how the model analyzes information, makes inferences, and draws conclusions to make decisions—more than the model’s actual output, which can sometimes be inaccurate. However, he said, that is hardly unique to AI.
“Every human develops logic and experience, but we don’t even do a great job of retaining facts,” he said. Still, the firm is moving slowly on green-lighting AI tools because “there is so much that we don’t know yet about these models.”
That measured pace can be difficult in an era of generative AI innovation, and Argenti said companies “need to have a sense of urgency, for sure.” But as a CIO at a highly regulated bank like Goldman Sachs, he added, “my job is not to be too impatient.”
The potential for increased productivity, rather than intelligence, is what was front and center in Goldman Sachs’ announcement yesterday that it will roll out the first generative AI tool across the firm by the end of the month. The new tool, an integration of Microsoft’s GitHub Copilot for generating software code, will be available to over 9,000 Goldman Sachs developers, or about 80% of the firm’s engineers, on the company’s centralized internal AI platform, called GS AI.
“Copilot is integrated in our software development life-cycle processing,” Argenti explained, adding that the code that is generated goes through the same series of quality checks as all of the firm’s software code.
Argenti told Fortune that the deployment of GitHub Copilot would offer developers significant efficiency gains, but it would not lead to job cuts. Instead, he described it as an investment in increased productivity.
“We are a growing company; we are in a competitive market,” he said. “So it’s more for us [that] we’re gaining additional bandwidth.”
Argenti had previously made clear that Goldman Sachs has no plans to build its own large language models from scratch and would, instead, integrate existing AI models, such as OpenAI’s GPT-3.5 and GPT-4, Google’s Gemini, and Meta’s Llama into its own platform.
“I might be completely wrong,” he said in October 2023, but “I don’t believe at this point … that it is necessary to start from scratch.”
Now that automating software coding has been tackled, Argenti said the firm’s next AI implementation will help knowledge workers with what’s known as document management, or repetitive tasks like summarizing text, producing excerpts, or augmenting documents with other data. However, the need to glean highly accurate information makes this a tougher challenge than code generation, he cautioned.
“It is harder because of the amount of information that is at play and the complexity of data sources,” he said. While he declined to say when the document management tool would be released across the firm, he noted that Goldman Sachs’ centralized GS AI platform would help with safety because it is designed with guardrails to allow the firm “to experiment safely at higher speed.”