On the heels of ChatGPT’s second anniversary, it seems that generative AI still has hype left in the tank going into 2025.
Despite some signs of faltering business potential this year and some worries about slowing progress in the next, AI bellwether Nvidia and other players in the AI race have continued to see massive growth—though Microsoft’s stock has underperformed the S&P 500 this year—and AI startups like OpenAI, Elon Musk’s xAI, and Anthropic have pulled in monster investments this quarter.
With plenty of predictions flying around about what the next year in the technology’s evolution holds, we talked to some experts and looked for themes. What’s clear from these conversations is that many in the space expect generative AI to look a bit different next year—more dedicated agents that can perform tasks beyond the scope of chatbots, as well as more specialized LLMs. Experts expect private data caches to become more valuable, and overall costs to come down. Meanwhile, human oversight and safeguards could become more important.
Agents, agents everywhere
Talk of agents as the next big era of GenAI advancement has been percolating all year, but this trend reached new heights in recent months as Salesforce rebranded its Copilot product as Agentforce, Microsoft rolled out its agent studio, and other enterprise companies offered up agent builders.
“You have all manner of agents. SAP has theirs, and I think ServiceNow has theirs, and people are just announcing agents left, right, and center,” Babak Hodjat, CTO of AI at Cognizant, told Tech Brew.
Experts expect these entities to coalesce into coordinated multi-agent networks in the next year and beyond. Hodjat, who helped build the technology that would become Apple’s Siri as a multi-agent architecture, said the idea isn’t new, but it’s gained more purchase in the business world lately.
“Multi-agent systems have been around since the ’90s, and they’re resurgent, but they were mainly in research,” Hodjat said. “Now, industry is seeing agents and agentification as the next big step from generative AI…And what is a very natural next step from that is connecting these agents. And that’s what we [at Cognizant] foresaw, and weren’t sure when it’s going to happen. But it’s unfolding faster than anyone expected.”
More human supervision
But if companies are already squeamish about risks with customer-facing chatbots, delegating tasks to autonomous systems will require even more trust in AI. Ece Kamar, director of Microsoft’s AI frontier labs, told us it will be a while before AI agent systems are able to perform without robust human oversight.
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
Scott Beechuk, a partner at Norwest Venture Partners, said he expects the rise of agentic systems could spawn a new role within companies.
“How are quality assurance engineers going to monitor these systems? Observability between agents as they communicate, so that if you have a system of, say, 20 agents all working together to complete a complex task, how are we going to trace that?” Beechuk said. “I think it’s going to quickly become a huge, huge issue that’s going to need human beings who are dedicated to it.”
Meet the specialists
Companies like OpenAI have more or less trained foundation models on the entire free and available internet at this point, and that, along with some other issues, has caused some consternation about how to scale these massive models in the next year.
PwC’s US Chief AI Officer Dan Priest said that industry and domain-specific models will become more important in 2025 as industries leverage specialized technical knowledge and data.
“Every single industry has their domain, their domain-specific data that’ll be part of the tuning that happens,” Priest told Tech Brew. “Every single function has their domain-specific data, and so those specialized data pools will drive the innovation that we expect to see in 2025.”
The end of “play money”
Of course, how businesses actually make revenue off AI has been a dominant question since the start of the boom, and Alois Reitbauer, chief technology strategist at Dynatrace, said this demand is only getting more urgent.
“That’s starting now—they got their play money for one and a half years,” Reitbauer said. “Now, the boards are coming back [to say], ‘How are we making money out of this? If this is the biggest thing since the internet, why aren’t we making money out of this?’”
Reitbauer said original, differentiated, and domain-specific innovations will be key to answering that question.
Natalie Glance, chief engineering officer at Duolingo, expects AI’s cost efficiency to improve in the next year, but the main focus remains on creating the best possible product.
“[We think] about product first and speed of execution second, and cost efficiency comes third, so that we’re kicking the ball to where the player is going to be instead of where the player is right now,” Glance said.