Life,  Tech

The Rise of AI Agents and Their Limitations

Driven by the large language model (LLM), the AI ​​agent (AI agent) exploded out of the circle.

AI agent, a kind of AI agent, can perceive the environment and respond, complete certain tasks through independent decision-making and action, and iterate through learning and reflection.

Due to the super autonomy, the whole process hardly depends on human instructions and supervision.

This new technology has swept through Silicon Valley, and individual developers, technology giants such as Microsoft and Google’s parent company Alphabet, and start-up companies have almost all entered the game.

Take one startup, Inflection AI, which raised $1.3 billion in funding at the end of June.

According to co-founders Reid Hoffman and Mustafa Suleyman in a podcast, the company is developing a personal assistant that it says can take on an advisory role, handling tasks like getting flight credits and hotels after travel delays.

Another startup, Adept, raised $415 million.

The model Action Transformer (ACT-1) developed by Adept can enter the company’s Salesforce customer relationship database by itself with only one prompt, and complete a task that is said to require 10 or more human clicks.

Meanwhile, investors have piled in.

Jason Franklin of WVV Capital (Vivo Capital) said he had to work hard to get two former Google Brain engineers to invest in an AI agency.

In May, Google Ventures led a $2 million seed round for startup Cognosys to develop AI agents.

Hesam Motlagh, founder of agency startup Arkifi, said the company closed a “sizeable” first round of funding in June.

In addition, according to Matt Schlicht, a well-known journalist in the field of AI, there are at least 100 serious projects dedicated to commercializing AI agents. He said:

Entrepreneurs and investors are very excited about autonomous (AI) agents . They are more interested in AI agents than chatbots.

AI company MultiOn is beta testing an AI agent app it developed, and developer DivGarg said:

For the most part now, we want it to be your personal AI friend.

He added:

It could evolve into (Tony Stark’s AI assistant in Iron Man) Jarvis, which we hope will be able to connect to many of your services. If you want to do something, you can talk to your AI assistant and it will do it for you.

Large language model, with flying AI agent
In March, OpenAI released the large language model GPT-4, further promoting the race in the field of AI agents.

In just a few weeks, autonomous AI agents have sprung up like mushrooms.

The most typical examples are Baby AGI and Auto-GPT, these two agents can determine the priority and execution of tasks according to preset goals and results.

Auto-GPT takes output from GPT-4, feeds it back to itself via an impromptu external memory, iterates further on a task, and can correct errors or suggest improvements.

Baby AGI operates similarly to Auto-GPT, but with a different task preference orientation.

Vivian Cheng, an investor at CRV, a venture capital firm focused on AI agents, said GPT-4 elevates the strategic and adaptive thinking needed to navigate the unpredictable real world.

Five AI agent developers, including Kanjun Qiu, CEO of OpenAI competitor General Intelligence, which creates artificial intelligence for agents, said they expect the first systems capable of reliably performing multi-step tasks with some autonomy will be within a year. Listed, with a focus on niche areas such as coding and marketing tasks.

“Stupid as a rock”
At present, AI agents are still in their infancy and need human “nanny-style” guidance and supervision, otherwise mistakes will occur.

“There are so many ways it can go wrong,” said Aravind Srinivas, CEO of ChatGPT competitor Perplexity AI. As a result, he favors proxies offering human-supervised products.

You have to treat AI like a baby and constantly supervise it like a mom.

“The real challenge is to build systems with strong reasoning capabilities,” Qiu said.

Microsoft CEO Satya Nadella sees the underlying model technology as a leap forward for its own digital assistants like Cortana, Amazon’s Alexa, Apple’s Siri and Google Assistant, all of which, in his view, have fallen short of original expectations.

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

Alexa already uses advanced AI technology, an Amazon spokesperson said, adding that the team is developing new models to make the assistant even more powerful and useful.

Google claims it’s also constantly improving its Assistant, with its Duplex technology making calls to restaurants to reserve tables and check hours.

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

While far from perfect, the potential for AI agents could be staggering, as OpenAI was the first to notice.

Recently, OpenA co-founder Andrej Karpathy said in an offline event for developers that if a paper proposes a different model training method, OpenAI will scoff at it, but when a new AI agent paper comes out, they will Seriously excited discussion.

Karpathy said excitedly at the time:

AI agents represent a future of AI!

error: Content is protected !!