Argenti also likened the advent of powerful generative artificial intelligence systems such as ChatGPT to the invention of the printing press, and predicted the technology will transform how businesses store and organize institutional knowledge, according to the email. He also raised the question of whether A.I. could make rising inequality worse.

Goldman Sachs declined to comment on Argenti’s message.

Printing press

In the email, Argenti said that while others have said generative A.I. will be more impactful than the discovery of fire, the debut of the internet, or the move to cloud computing, he believed that a better analogy is the invention of the printing press, which had the effect of both democratizing access to knowledge as well as massively accelerating the codification of knowledge.

Argenti said that while “efficiency gains are capturing a lot of the mindshare” he believed “LLMs are a breakthrough in knowledge more than they are in productivity.” He said this was similar to what happened when the printing press made books must cheaper and faster to produce. LLM is an acronym for large language models—the A.I. systems that encode the statistical connections between words and phrases learned from vast amounts of human text—that  underpin chatbots such as OpenAI’s ChatGPT, Microsoft’s new Bing chat, Google’s Bard, and a half dozen others.

In the letter, Argenti argues that LLMs represent a breakthrough in knowledge dissemination by eliminating barriers to understanding complex subjects. With LLMs, users can adjust the level of detail and complexity that the LLM produces, making knowledge more accessible and comprehensible, he said.

‘Tribal’ knowledge

Argenti broached the idea that Goldman might train its own “ChatGS” LLM on all of the knowledge the company possess and which could then help users find and surface that knowledge in the format that would be most useful for them.

The Goldman CIO said that the problem with any large corporation today is that a lot of critical knowledge resides only in the brains of its employees. “Within a corporation, most knowledge is not codified. It’s tribal,” Argenti wrote. “It resides in the minds of ‘experts,’ connected by an internal social network that takes years to master.” That knowledge can also be lost if one of those key experts leaves the company. Argenti postulates that a “ChatGS” system could help Goldman breakdown these knowledge silos and ensure that vital information was recorded, stored, and accessible.

In the message, Argenti also touched on concerns about the downsides of generative A.I., including the idea that A.I. could pose an existential threat to humanity and whether research into more powerful A.I. should be paused to give society and government regulators a chance to grapple with the implications of the new technology. He also broached the idea that the efficiency gains from powerful A.I. could make inequality worse than it already is. But Argenti did not offer his personal take on these pressing issues. Instead, he asksed a series of rhetorical questions.

“In this particular moment, one of the greatest unknowns, and risks, for humanity lies in the answer to the following questions: would AI increase or close the societal gap?” he wrote. “Would it pose a threat to the very existence of humankind as we know it? Should we give ourselves time to reflect before we push ahead at full speed?”

Argenti then concludes the letter with the idea that the democratization of knowledge from LLMs “would sway the pendulum in a benign and desirable direction of a better, fairer world.”

Corporate knowledge bases

The idea of corporate knowledge bases, knowledge graphs, and other software-enabled “knowledge management systems,” not dissimilar to what Argenti is suggesting LLMs will enable, enjoyed a brief heyday in the late 1990s, shortly after graphical web browsers spurred the internet boom. It was during this period that many companies created corporate “intranets” to supposedly codify and store knowledge. But many of these corporate knowledge management systems never lived up to their promise, in part because they were often difficult to search—it was often easier to just walk down the hall and ask the internal expert or send them an email—and even more difficult to keep up-to-date.

The same statistical encoding process that underpins LLMs can be used to create “vector databases” that allow people to find information without having to search for exact keywords. That could solve the search issue that sank many corporate knowledge graphs. But it does not solve the problem of how to get the knowledge out of people’s heads and written down in the first place so that an LLM can be used to encode it. Most experts are too busy doing their day jobs to stop and write down step-by-step instructions or explanations for much of what they do. And some expertise is based on intuition, built over years of experience, that can be difficult to articulate.