Here Come the AI Assistants

The race for generative AI (Gen AI) leadership hit an inflection point when a recent Chinese development of a Gen AI model known as DeepSeek, surprised industry watchers with its low-cost AI tool, raising questions about the massive spending by US technology companies on Gen AI.

The development may lead to steep declines in per-token costs that goes into the pre-training infrastructure, but it won’t necessarily mean the spending on Gen AI was not useful, according to George Lee, co-head of Goldman Sachs Global Institute.

“First of all, this [development] doesn't mark the end and perhaps it marks the beginning of even more pre-training activity by more people who can afford to embark on this at lower capital costs,” says Lee on Goldman Sachs Exchanges podcast.

Lee is alluding to the Jevons paradox where more efficiency in a resource would lead to higher adoption and consumption rather than lower it. This paradox originated from William Stanley Jevons in 1865.  He argued how more efficient steam engines would not decrease use of coal in British factories, but rather increase it. In his view, cheaper fossil fuels would lead to more demand for the resource (rather than less), as more steam engines would be built.

For Gen AI, this could mean expanding on use cases across the business landscape beyond just automating simple tasks – complex tasks such as scientific research and legal assistance could also be automated, says Kim Posnett, global co-head of Goldman Sachs Investment Banking.

However, there’s one parallel phenomenon on the rise that is changing the way we use this technology – AI assistants.

What are AI agents or assistants?

Conversational AI is already pretty ubiquitous, that is “sort of like personal assistants for everyone in every context – personally, professionally,” says Posnett. But Gen AI can do more than chat, it can potentially execute complex multistep tasks autonomously.

When you hear about the development of Gen “AI agents” or “AI assistants,” they are essentially a system of “models, computation, and resources that complete linked tasks and allow you to complete more multistep complex tasks in business or personal life,” explains Lee.

For example, a user can ask the AI agent to book a flight, hotel, and car rental, and it would complete all of these tasks autonomously. It gives instructions to various applications that may involve opening up a webpage, taking a picture of the webpage, discerning pixels, identifying text entry boxes and buttons. It could even take hold of your computer cursor and begin to execute tasks on your behalf.

That said, these types of consumer AI agents currently in development are still in the early stages. Lee notes that developers want to enroll people in refining it. In its current form, it’s reportedly slow and deliberately herky-jerky while it executes the tasks.

“Nonetheless, [it’s] an inspiring direction of travel for the technology,” says Lee.

Consumer AI agents are also only one stream of work in the agent space – enterprises and Gen AI research shops are also promoting and developing their own models of an AI agent to assist with their work.

Posnett agrees that it’s still early days for AI agents: “I believe they will also be ubiquitous over time – who knows what timeframe will be.”