Rafael Enrique | LightRocket | Getty Images
OpenAI will publicly launch its ChatGPT chatbot in late 2022, giving the public a taste of how new generative AI chatbots can provide far more thorough, creative and conversational answers to web queries compared to the traditional searches that consumers have used to look for online information for the past 25 years. Google, Microsoft and others have followed suit, announcing competing products.
The industry quickly moved from text responses to AI-generated photos and videos, and now we're seeing the rise of AI agents.
Agents are built to increase productivity and complete tasks, not just provide answers — the domain of chatbots and image generators. For better or worse, agents are AI tools that can make decisions “without a human,” Kvamme said.
Grace Isford, a partner at venture firm Lux Capital, said tech investors have “dramatically increased” interest in startups focused on building AI agents. Collectively, the companies have raised hundreds of millions of dollars and valuations have risen along with the overall generative AI market.
According to PitchBook, generative AI is set to explode in 2023, with $29.1 billion invested across nearly 700 deals, up more than 260% in deal value from 2018. Meanwhile, the non-AI investment environment has been in a protracted downturn for more than two years, following record fundraising during the COVID-19 pandemic.
If 2023 is the year that AI hype peaked, 2024 will be the year of early adoption.
“Since the introduction of ChatGPT, we've seen a veritable whirlwind of innovation in the market,” Jared Spataro, corporate vice president of AI at Work at Microsoft, told CNBC. Microsoft is OpenAI's largest backer, investing billions in its own generative AI models and products in addition to the billions it has poured into ChatGPT developers.
The term AI agents isn't clearly defined across the tech industry. Industry experts who spoke to CNBC about this emerging trend generally see agents as a step beyond chatbots in that they are designed for specific business functions and can be customized with large-scale AI models. Think JARVIS, Tony Stark's multifaceted AI assistant in the Marvel Universe.
AI agents are often described as advanced generative AI tools that can carry out complex, multi-step tasks on behalf of users and generate their own to-do lists, eliminating the need for users to walk them through a process step by step.
“The Assistant doesn't just give you an answer, it automates a series of steps,” said François Ajzenstat, chief product officer at digital analytics company Amplitude.
Microsoft CEO Satya Nadella said on an earnings call earlier this year that he wants to offer AI agents that can complete more tasks on users' behalf, but that “there is still a lot of work to be done.” Executives at Meta and Google have also touted efforts to make their AI assistants more productive.
At Google I/O in May, Google announced Project Astra, the company's latest advancement toward an AI assistant being developed by Google's AI division DeepMind.
Google's demo video shows the Assistant using video and voice to help users remember where they put their glasses, look up codes, and answer questions about displayed objects. It's still just a prototype for now, but Alphabet CEO Sundar Pichai said the company hopes to roll it out to users later this year.
The demo came a day after OpenAI showed off a similar voice conversation with ChatGPT, which it positions as an AI assistant that can act as a conversation partner, language translator, math tutor and code collaborator.
At its Build developer conference, Microsoft announced a partnership with Cognition AI to bring to customers Cognition's own AI agent, Devin, which Cognition describes as its “first AI software engineer.”
Devin quickly became a social media sensation for his ability to handle multi-step processes. Rather than just generating simple lines of code, Devin creates a process for solving a problem, writes the code, tests it, and then ships it.
Martin Kong, chief operating officer of enterprise AI startup Cohere, said AI agents could take on tasks like booking flights, processing expenses, recommending interest rates for loans, and emailing customers their arrival time and updating Salesforce accordingly.
So far, these tools have been largely limited to tasks like helping people write code: At Microsoft's GitHub, for example, roughly 46% of all code “across all programming languages” is generated by AI, CEO Thomas Dohmke said in a blog post earlier this year.
While the line between AI coding tools and true AI agents is blurry, most experts who spoke to CNBC said the defining feature of agents is that they are moving far beyond a single use case and are beginning to approach full-fledged personal assistants.
Anthropic and other startups are already working toward that goal: The first step is giving chatbots the ability to interact with external tools and services on customers' behalf.
Microsoft's Spataro said the development process for the company's coding agent, Copilot, was “like being strapped to a rocket.” He said a big part of what Microsoft is doing is moving from one- or two-step tasks to multi-step tasks, which could include looking at a user's calendar and providing a 30-second summary of their priorities for the day.
Fred Havemeyer, head of U.S. AI and software research at Macquarie, said in a recent note to investors that the firm expects to see more AI agents on the rise.
“We believe that agent-like AI that can self-direct itself to accomplish tasks will be the tool that unlocks the value of GenAI for everyday users,” Havemeyer wrote.
Roman Huet, head of developer experience at OpenAI, told CNBC that the concept of AI agents gained traction last year, but the company quickly realized there was work to be done to make the tools more autonomous.
“Our models are getting more and more powerful and are able to capture user intent in a much better way than before, but we're still in the early stages in terms of building agents,” Huett said.
The big advance, he said, will be when AI agents learn user preferences and are able to “act on your behalf” without you having to ask them to.
AI agent startups are raising a lot of money from investors, and while it's not the same as the billion-plus dollars being pumped into AI model companies, their valuations are still far above the fundamentals of their businesses.
Adept, led by OpenAI and Google alumni, was valued at more than $1 billion last year and says on its website that its technology “silences users to the complexities of software tools.”
French AI agent startup H raised a $220 million seed round in May from investors including Amazon, Samsung, UiPath, and former Google CEO Eric Schmidt. Artisan AI, a Y Combinator-backed startup developing AI agents it describes as “AI employees for enterprises,” recently closed a $7.3 million seed round, saying more than 100 companies have participated so far.
Jaspar Carmichael-Jack, founder and CEO of Artisan AI, said that true AI agents won't be able to begin developing until 2022, as chatbots such as ChatGPT will be the first to allow regular consumers to interact with such tools.
“People are talking about the VC market being generally down,” Carmichael Jack said, “but for us, it's like 2021 for AI startups.”
Braden Hancock worked at Facebook Research and Stanford University's Artificial Intelligence Lab before co-founding Snorkel AI in 2019. He said the market is in a “similar hype cycle” to self-driving cars, and it will take a similarly long time for broader AI agents to go mainstream.
Hancock said agents have to be “many times more” competent for people to be “acceptable to putting something on autopilot.” He added that there is a “very high bar” to clear when it comes to using technology to sign your name and send money on your behalf.
Imbue, a three-year-old startup founded by Kangjun Qiu, is backed by Amazon's Alexa Fund and Eric Schmidt and is valued at over $1 billion. Based on the company's own user research, Qiu said that the current characterization of AI agents (generally intelligent personal assistants that handle delegated tasks) is not what users actually want because by design they are “not completely trustworthy.”
“Even as a CEO, it's hard to delegate work to your executive assistant,” Qiu says. “I've known her for two years, and she's great.” And when it comes to new jobs, “it's still hard to fully gauge: 'Okay, is this going to live up to my expectations?'” Qiu says.
Imbue is developing a no-coding way to create your own AI software agents that can run in the background to cater to your personal needs, like creating a way to track the news, building a bot to book travel, etc. These types of AI models do not need to be trained with user data because each use case is personalized.
Rather than delegating tasks to agents built by companies like OpenAI or Google and being centrally managed and controlled by those companies, Imbue envisions the agents delegating control into the hands of the user.
“The idea with agents is that they allow anyone to write software,” Qiu says. Users “ask the agents to write code on the computer and make the computer do what they want it to do.”
clock: “It will take 10 years to solve AI”
