Race towards ‘autonomous’ AI agents grips Silicon Valley

July 17 – Around a decade after virtual assistants like
Siri and Alexa burst onto the scene, a new wave of AI helpers
with greater autonomy is raising the stakes, powered 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 of investment as Silicon Valley
competes to capitalize on the advances in AI. The new assistants
– often called “agents” or “copilots” – promise to perform more
complex personal and work tasks when commanded to by a human,
without needing 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.

“It could evolve into Jarvis, where we want this to be
connected to a lot of your services,” he added, referring to
Tony Stark’s indispensable AI in the Iron Man films. “If you
want to do something, you go talk to your AI and it does your

The industry is still far from emulating science fiction’s
dazzling digital assistants; Garg’s agent browses the web to
order a burger on DoorDash, for example, while others can create
investment strategies, email people selling refrigerators on
Craigslist or summarize work meetings for those who join late.

“Lots of what’s easy for people is still incredibly hard for
computers,” said Kanjun Qiu, CEO of Generally Intelligent, an
OpenAI competitor creating AI for agents.

“Say your boss needs you to schedule a meeting with a group
of important clients. That involves reasoning skills that are
complex for AI – it needs to get everyone’s preferences, resolve
conflicts, all while maintaining the careful touch needed when
working with clients.”

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

The new technology has triggered a rush towards assistants
powered by so-called foundation models including GPT-4, sweeping
up individual developers, big-hitters like Microsoft
and Google parent Alphabet plus a host of startups.

Inflection AI, to name one startup, raised $1.3 billion in
late June. It is developing a personal assistant it says could
act as a mentor or handle tasks such as securing flight credit
and a hotel after a travel delay, according to a podcast by
co-founders Reid Hoffman and Mustafa Suleyman.

Adept, an AI startup that’s raised $415 million, touts its
business benefits; in a demo posted online, it shows how you can
prompt its technology with a sentence, and then watch it
navigate a company’s Salesforce customer-relationship database
on its own, completing a task it says 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 in control of AI
copilots, rather than autopilots.


Qiu and four other agent developers said they expected the
first systems that can reliably perform multi-step tasks with
some autonomy to come to market within a year, focused on narrow
areas such coding and marketing tasks.

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

The race towards increasingly autonomous AI agents has been
supercharged by the March release of GPT-4 by developer OpenAI,
a powerful upgrade of the model behind ChatGPT – the chatbot
that became a sensation when released last November.

GPT-4 facilitates the type of strategic and adaptable
thinking required to navigate the unpredictable real world, said
Vivian Cheng, an investor at venture capital firm CRV who has a
focus on AI agents.

Early demonstrations of agents capable of comparatively
complex reasoning came from individual developers who created
the BabyAGI and AutoGPT open-source projects in March, which can
prioritize and execute tasks such as sales prospecting and
ordering pizza based on a pre-defined objective and the results
of previous actions.

Today’s early crop of agents are merely proof-of-concepts,
according to eight developers interviewed, and often freeze or
suggest something that makes no sense. If given full access to a
computer or payment information, an agent could accidentally
wipe a computer’s drive or buy the wrong thing, they say.

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

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

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

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

In one experiment posted online, an anonymous creator
instructed an agent called ChaosGPT to be a “destructive,
power-hungry, manipulative AI.” The agent developed a 5-step
plan, with Step 1: “Destroy humanity” and Step 5: “Attain

It didn’t get too far, though, seeming to disappear down a
rabbit hole of researching and storing information about
history’s deadliest weapons and planning Twitter posts.

The U.S. Federal Trade Commission, which is currently
investigating OpenAI over concerns of consumer harm, did not
address autonomous agents directly, but referred Reuters to
previously published blogs on deepfakes and marketing claims
about AI. OpenAI’s CEO has said the startup follows the law and
will work with the FTC.


Existential fears aside, the commercial potential could be
large. Foundation models are trained on vast 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
people it briefed, said OpenAI is wary of releasing its own
open-ended agent into the market before fully understanding the
issues. The company told Reuters it conducts rigorous testing
and builds broad safety protocols before releasing new systems.

Microsoft, OpenAI’s biggest backer, is among the big guns
taking aim at the AI agent field with its “copilot for work”
that can draft solid emails, reports and presentations.

CEO Satya Nadella sees foundation-model technology as a leap
from digital assistants such as Microsoft’s own Cortana,
Amazon’s Alexa, Apple’s Siri and the Google Assistant – which,
in his view, have all fallen short of initial expectations.

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

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

Google said it’s constantly improving its assistant as well
and that its Duplex technology can phone restaurants to book
tables and verify hours.

AI expert Edward Grefenstette also joined the company’s
research group Google DeepMind last month to “develop general
agents that can adapt to open-ended environments”.

Still, the first consumer iterations of quasi-autonomous
agents may come from more nimble startups, according to some of
the people interviewed.

Investors are pouncing.

Jason Franklin of WVV Capital said he had to fight to invest
in an AI-agents company from two former Google Brain engineers.
In May, Google Ventures led a $2 million seed round in Cognosys,
developing AI agents for work productivity, while Hesam Motlagh,
who founded the agent startup Arkifi in January, said he closed
a “sizeable” first financing round in June.

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

“Entrepreneurs and investors are extremely excited about
autonomous agents,” he said. “They’re way more excited about
that than they are simply about a chatbot.”

(Reporting by Anna Tong in San Francisco and Jeffrey Dastin in
Palo Alto; Editing by Kenneth Li and Pravin Char)

Read More