Insight: The race for “autonomous” AI agents is taking over Silicon Valley

Jul.17 – About a decade after virtual assistants like Siri and Alexa hit the scene, a new wave of self-powered AI helpers is upping the ante, fueled by the latest version of the technology behind ChatGPT and its rivals.

Experimental systems that run on GPT-4 or similar models are attracting billions of dollars in investment as Silicon Valley competes to capitalize on AI advances. The new assistants – often called “agents” or “co-pilots” – promise to carry out more complex personal and work tasks at the command of a human, without the need for close supervision.

“High level, we want this to become something like your personal AI friend,” said developer Div Garg, whose company MultiOn is beta testing an AI agent.

“He could evolve into Jarvis, where we want him to be plugged into a lot of your services,” he added, referring to Tony Stark’s indispensable AI in the Iron Man movies. “If you want to do something, go talk to your AI and do your thing.”

The industry is still a long way from emulating the dazzling digital assistants of science fiction; Garg’s agent surfs the web to order a burger on DoorDash, for example, while others can create investment strategies, email people who sell coolers on Craigslist, or summarize business meetings for those joining late.

“A lot of what’s easy for people is still incredibly difficult for computers,” said Kanjun Qiu, CEO of Generally Intelligent, an OpenAI competitor that builds AI for agents.

“Suppose your boss needs you to schedule a meeting with a group of important clients. This involves complex reasoning skills for the AI: it has to get everyone’s preferences, resolve conflicts, all while maintaining the careful touch needed when you work with customers.”

Early efforts are just a taste of the sophistication that could come in coming years from ever more advanced and autonomous agents as the industry pushes toward an artificial general intelligence (AGI) that can match or surpass humans in a myriad of cognitive tasks. according to Reuters interviews with about two dozen entrepreneurs, investors and AI experts.

The new technology has sparked a rush towards assistants powered by so-called basic models including GPT-4, garnering individual developers, blockbusters like Microsoft (MSFT.O) and Google Parent Alphabet (GOOGL.O) plus a slew of startups .

Inflexion AI, to quote one startup, raised $1.3 billion in late June. According to a podcast from co-founders Reid Hoffman and Mustafa Suleyman, he’s developing a personal assistant who could act as a mentor or handle tasks like securing flight credit and a hotel after a travel delay.

Adept, an AI startup that raised $415 million, touts its business benefits; in a demo posted online, she shows how you can solicit her technology with one sentence, and then watch her navigate a company’s Salesforce customer relationship database on her own, completing a task that would take a human 10 or more clicks.

Alphabet declined to comment on agent-related work, while Microsoft said its vision is to keep humans under the control of AI copilots, rather than autopilots.

PHASE 1: DESTROY HUMANITY

Qiu and four other agent developers said they expect the first systems capable of reliably performing multi-step tasks with some autonomy to hit the market within a year, focusing on narrow areas like coding and marketing tasks.

“The real challenge is building systems with sound reasoning,” Qiu said.

The race towards ever more autonomous AI agents was boosted by the March release of GPT-4 from developer OpenAI, a powerful update to the model behind ChatGPT, the chatbot that became a sensation when it was released last November.

GPT-4 facilitates the kind of strategic and adaptive thinking needed to navigate the unpredictable real world, said Vivian Cheng, an investor at venture capital firm CRV who focuses on AI agents.

The first demonstrations of relatively complex reasoning agents came from individual developers who in March created the open source projects BabyAGI and AutoGPT, which can prioritize and perform tasks such as sales prospecting and pizza ordering based on a predefined goal and the results of previous actions.

According to eight interviewed developers, today’s first crop of agents is simply a proof of concepts, and often blocks or suggests something that doesn’t make sense. Given full access to a computer or payment information, an agent could accidentally erase a computer’s drive or buy the wrong thing, they say.

“There are so many ways it can go wrong,” said Aravind Srinivas, CEO of ChatGPT competitor Perplexity AI, which has instead opted to offer a human-supervised co-pilot product. “You have to treat the AI ​​like a baby and supervise it constantly like a mom.”

Many computer scientists focused on the ethics of AI have pointed out the near-term harms that could result from the perpetuation of human biases and the potential for disinformation. And while some see a future Jarvis, others fear the deadly HAL 9000 from “2001: A Space Odyssey”.

Computer scientist Yoshua Bengio, known as the “godfather of AI” for his work on neural networks and deep learning, urges caution. He fears that future advanced iterations of technology may create and act on their own unexpected goals.

“Without a human being in the loop who checks every action to see if it’s not dangerous, we could end up with actions that are criminal or could harm people,” Bengio said, calling for more regulation. “In a few years these systems may be smarter than us, but that doesn’t mean they have the same moral compass.”

In an experiment posted online, an anonymous creator tasked an agent called ChaosGPT to be a “destructive, power-hungry, manipulative AI.” The agent developed a 5-step plan, with Phase 1: “Destroy Humanity” and Phase 5: “Achieve Immortality”.

However, it didn’t get too far, seeming to disappear down a rabbit hole of researching and archiving information about history’s deadliest weapons and scheduling Twitter posts.

The US Federal Trade Commission, which is currently investigating OpenAI over fears of harm to consumers, didn’t approach autonomous agents directly, instead directing Reuters to previously published blogs about deepfakes and AI marketing claims. The OpenAI CEO said the startup follows the law and will work with the FTC.

‘STUDY LIKE A ROCK’

Existential fears aside, the commercial potential could be great. The basic models are trained on large amounts of data such as text from the internet using artificial neural networks that are inspired by the architecture of biological brains.

OpenAI itself is very interested in AI agent technology, according to four people briefed on its plans. Garg, one of the briefers, said OpenAI is wary of releasing their open-ended agent to market before we fully understand the issues. The company told Reuters that it conducts rigorous testing and builds extensive security protocols before releasing new systems.

Microsoft, OpenAI’s biggest proponent, is among the bigwigs targeting the AI ​​agent field with its “copilot for the job” capable of drafting solid emails, reports, and presentations.

CEO Satya Nadella sees the base-model technology as a leap from digital assistants like Microsoft’s Cortana, Amazon’s Alexa, Apple’s Siri and the Google Assistant, all of which he says have fallen short of initial expectations.

“They were all dumb as a rock. Whether it’s Cortana or Alexa or Google Assistant or Siri, all of these just don’t work,” she told the Financial Times in February.

An Amazon spokesperson said Alexa already uses advanced AI technology, adding that his team is working on new models that will make the assistant more capable and useful. Apple declined to comment.

Google said it’s also constantly improving its assistant, and that its Duplex technology can phone restaurants to reserve tables and check times.

Artificial intelligence expert Edward Grefenstette also joined Google’s DeepMind research team last month to “develop general agents that can adapt to open environments.”

However, according to some of the people interviewed, the first consumer iterations of quasi-autonomous agents could come from more agile startups.

Investors are jumping.

WVV Capital’s Jason Franklin said he had a tough time investing in an AI agent firm of two former Google Brain engineers. In May, Google Ventures spearheaded a $2 million seed round in Cognosys, developing AI agents for labor productivity, while Hesam Motlagh, who founded agent startup Arkifi in January, said he closed a first round of “considerable” funding in June.

There are at least 100 serious projects working to commercialize agents, said Matt Schlicht, who writes an AI newsletter.

“Entrepreneurs and investors are extremely enthusiastic about autonomous agents,” he said. “They are much more excited about this than just a chatbot.”

Reportage by Anna Tong in San Francisco and Jeffrey Dastin in Palo Alto; Edited by Kenneth Li and Pravin Char

Our standards: the Thomson Reuters Trust Principles.