Also this week, Goldman Sachs Chief Information Officer Marco Argenti told me his team is also interested in using DeepSeek, but the model needs to go through a security screening. That’s presumably to guard against hidden data known as prompt injections, which is a security concern for any company that uses an AI model to power its applications, whether that model is Chinese or not. (If you need a helpful glossary to understand other AI jargon that’s showing up everywhere, this one is hot off the presses!)

If DeepSeek is good enough for Goldman, that should be the latest indication open source models have truly arrived after a long struggle to catch up to closed-source rivals. 

Goldman’s AI practices are also a sign of what’s to come more broadly in the business world, given the scale of its software purchasing. 

For starters, the bank, like many other businesses, has gotten productivity gains from coding software. After giving its more than 10,000 software developers access to Microsoft’s coding assistant, GitHub Copilot, in June, Argenti said they are now 10% more efficient at coding on average, which he likened to having 1,000 additional software engineers at his disposal.

Of course, AI developers including OpenAI are developing ambitious products that would effectively do the work of a senior software engineer, not just auto-complete code as human engineers type, which is Copilot’s specialty. Goldman is currently testing more complex AI coding tools, but Argenti didn’t elaborate. 

Even the basic productivity gains give the investment bank the option to hire fewer engineers if it wants to and it’s “literally freeing up capacity” among human engineers, he said.

Goldman wants its bankers, traders and other staff to use AI too. The bank this month launched an AI chatbot it developed for employees to generate slides for client presentations or extract insights from long documents. Employees using the chatbot can pick AI models from Meta Platforms (Llama), Google and OpenAI, as each of the models may be better than the others at handling different tasks.
High-Flyer’s DeepSeek release was a big deal in part because it showed that financial firms could influence the AI sector just as much as tech companies. Argenti, however, isn’t interested in investing in training models from scratch, and Goldman doesn’t plan to release any.