Join top executives in San Francisco on July 11-12 to hear how leaders are integrating and optimizing AI investments for success. Learn more
If you thought the pace of AI development had accelerated since the release of ChatGPT last November, well, brace yourself. Thanks to the emergence of autonomous AI agents such as Automatic GPT, babyAGI And AgentGPT in recent weeks, the race to get ahead in AI has only gotten faster. And, many experts say, more concerning.
>>Don’t miss our latest special issue: Data centers in 2023: doing more with less.<
It all started in late March, when developer Toran Bruce Richards, under the name @significantgravitas, launched Auto-GPT, an “experimental open-source application” tied to OpenAIs GPT-4 by API. Auto-GPT ran on Python and had Internet access, long/short-term memory, and, by stringing GPT calls together in loops, could act autonomously without requiring a human agent to perform each action.
Then, on March 29, @yoheinakajima launched BabyAGI, a “task-driven autonomous agent” using GPT-4, Pine cone‘s vector search, and LangChainAI‘s framework to “autonomously create and execute tasks based on an objective.”
Fast forward a few weeks, and now Auto-GPT has more GitHub stars than PyTorch (currently 82K) and is “the fastest growing GitHub repo in history, eclipsing ten year old open source projects in 2 weeks. ” Fortune says BabyAGI is “taking Silicon Valley by storm” and OpenAI’s Andrej Karpathy, who was previously AI director at Tesla, called instruments the “next frontier of rapid engineering.”
Are AI agents a game-changer?
Jay Scambler, an Oklahoma City-based consultant and strategist who builds AI tools for small businesses and creatives, told me via a Twitter post last week that the tools feel like a game changer. “I don’t want to sound dramatic, but we now have the power and responsibility to have a coordinated team of AI agents at our fingertips without much effort,” he said. “This team is never tired, executes code *almost* flawlessly (depending on who you ask) and can find answers to almost any problem using tools like LangChain.”
Others are not so optimistic. Nvidia AI scientist Jim Fan tweeted: “I see AutoGPT as a fun experiment, as the authors also indicate. But nothing more. Prototypes are not intended to be production ready. Don’t let the media fool you – most of the ‘cool demos’ are heavily singled out.”
Anyway (and of course there is a but), at the moment both Auto-GPT and BabyAGI require developer skills and are not accessible to the average ChatGPT user. And even Nicola Bianzino, chief technology officer at EY, told me in an interview that Auto-GPT is “fascinating,” but he admits he doesn’t yet understand the details of how it actually works. This is happening so fast, he explained, that there are already a large number of versions on top of the original. “I personally don’t know the different variations that exist in the wild,” he said.
Serious concerns about AI agents in the wild
Although the AI agents are “profound”, there are also serious concerns. Daniel Jeffries, former chief information officer at Stability AI and general manager of the Alliance for AI Infrastructure, told me last week that “the challenge is we don’t really know what an error looks like. Currently Auto-GPT fails reasoning 15-30% of the time, I think we are becoming less tolerant of errors as they become more autonomous.
And while the current use cases are limited, such as Fortune article pointed out that there are other risks lurking – including the continuous chain of prompts from the AI agent quickly racking up significant bills with OpenAI; the possibility of malicious use cases such as cyber-attacks and fraud; and the danger of autonomous bots taking action in a way the user didn’t intend, such as buying items, making appointments, or even selling stock.
That doesn’t seem to slow down the race to develop tools for AI agents, though. Last week eg. HyperWritea startup known for its generative AI writing extension unveiled an experimental AI agent that can surf the web and interact with websites just like a human user.
HyperWrite CEO Matt Shumer said his team is very focused on security issues. “We want to find the right way to do it, and that’s kind of a thread through all of this, we’re taking our time doing this the right way,” he said.
I also got a chance to talk to the developers behind AgentGPT, a browser-based AI agent launched on April 8 that offers easier access to the non-technical user.
Three developers with a day job worked on autonomous agents in their spare time, with a view to use cases for internal tooling. Seeing the explosive popularity of Auto-GPT and BabyAGI, they decided to implement their project and get feedback. In just nine days, AgentGPT has over 14,000 stars on GitHub and over 280,000 users, and developers sleep a few hours a night to keep things running.
“It’s been pretty crazy,” said Srijan Subedi, one of the founders of AgentGPT. “We did about 2x every day.”
While AgentGPT doesn’t have web browser capabilities yet, they say they’ll implement it in a week or two. But the bigger vision behind AgentGPT, the developers say, is to move beyond web browsing and integrate the AI agent with other tools, such as Slack, email, and even Facebook.
Are AI Agents Just Hype and Bustle?
Some say the new focus on AI agents is just another example of “busy brothers”, with hyperbolic claims from “get-rich schemers” who want to play off the excitement surrounding the potential of these tools.
That may be true, but to me it seems like the pace of AI development in this space is real, meaning it’s worth keeping a close eye on, especially as the risks and dangers become crystal clear. It may be impossible for me to keep up, but with developers starting to use Auto-GPT on their phones, I think we should all strap in for a quick ride.
VentureBeat’s mission is to become a digital city plaza where tech decision makers can learn about transformative business technology and execute transactions. Discover our Briefings.