Updated Mar 13, 2024 - Technology

The road map to AI's next level could be nature

Illustration of a collection of zeroes and ones in the style of an antique scientific illustration.

Illustration: Annelise Capossela/Axios

The path to artificial general intelligence (AGI) requires a different approach than today's generative AI models, one inspired by natural ecosystems, argues a newly out-of-stealth-mode firm that's hatched such a project.

Why it matters: Digital intelligence based on a web of intelligent agents is potentially cheaper, more environmentally sustainable, and more geopolitically defensible than one vast system trained on billions of data points.

Verses AI is entering the debate with the bold claim that the machine-learning methods behind ChatGPT and the rest of AI's advances for the past 20 years will never get the industry to AGI.

AGI is the industry's grail — a human level of artificial intelligence that can reason and learn in new ways — and Verses' founders say today's most advanced large language models, like OpenAI's GPT-4, can't deliver it.

  • "There's no evidence of an ability to act outside its training data," CEO Gabriel René told Axios.

The big picture: Verses is working instead on what it calls distributed intelligence, using biology as its starting point, in the belief that AGI is only possible with a system that can self-organize and retrain in real time — as biological organisms do.

  • Verses chief scientist Karl Friston is betting this requires higher degrees of autonomy and computing efficiency than the current school of large model development allows.
  • Building on 30 scientific papers from its researchers, the company has developed Genius, an operating system for "continually learning autonomous agents" operating at the edge of our connected devices.
  • NASA's Jet Propulsion Laboratory and Volvo are among the beta users of Genius.

"Minimizing complexity" is his "fundamental drive," Friston says. Instead of building ever-bigger AI models, Verses aims to deliver "99% smaller models" without sacrificing quality.

  • "90% of the neural net is not useful. There has to be a move away from big data to sparse data that is very well selected," Friston said, including the ability to forget data that is not relevant.

Flashback: AI's roughly 70-year scientific journey has taken many twists and turns, and just as the conventional wisdom of 25 years ago was shattered by the rise of machine learning, today's dominant paradigm of massive scale could be displaced by a different approach.

  • Computing's history is full of pendulum swings between monolithic and distributed systems, the center and the edge.
  • Jeff Hawkins, the inventor of the PalmPilot, was one pioneer of distributed intelligence theory — including a "thousand brains theory."
  • The roboticist Rodney Brooks pioneered a distributed approach to autonomous machines that was featured in Errol Morris' documentary "Fast, Cheap and Out of Control."

René offered two analogies for the current moment in AI.

  • The contest between advocates of big AGI versus a web of smaller intelligent agents is "AOL versus the web" all over again, he says.
  • The AI world, he predicts, is also headed for a "smartphone moment" in which the AI equivalents of Nokia, Blackberry and Motorola are rapidly superseded by a better architecture.

The problem today is that "most AI out there is just good input and output mapping, without any notion of agency or context sensitivity" that true agents would display, Friston said.

  • Today's AI models are pre-trained and memorize content, spitting it back out in new formations on demand — but since the systems do not have critical thinking skills, they lack autonomy.
  • Only a model that can identify its mistakes and correct them by re-training in real time would qualify as "superintelligence," René said.

How it works: Friston proposes an approach called "active inference" in which intelligent agents with predictive and adaptive abilities autonomously share knowledge with other agents and generate new agents, creating a self-sustaining web of intelligence.

  • Friston argues that AGI can't be achieved without belief structures. "I have to model my beliefs, not just the data," including quantifying uncertainty, just as humans attempt to assess what they don't know when they reach judgments, he said.

"Instead of just scaling a machine, you're growing an ecosystem," said René, in which agents are gathering evidence for their own use and exchange with other agents — just as organisms in a human body do.

  • Ninety percent of your body is autonomous agents working without your input, and your body is constantly seeking balance, René said.
  • Verses' approach imagines a constant process of rebalancing among agents, similar to the equilibrium achieved within Wikipedia by the interactions of the people who build it.

Friction point: Verses has suggested that generative AI market leader OpenAI is breaching its charter by not engaging with Verses.

  • Verses took out a full-page ad in the New York Times in December that announced its progress and methods. Verses challenged OpenAI to cooperate with it — asking the nonprofit firm to fulfill its charter promise to "stop competing" with any "value-aligned, safety-conscious project that comes close to building AGI before we do."

Reality check: Verses, based in Culver City, California, has around 100 employees across 60 remote locations, and has raised $65 million.

  • It's a small player entering a battlefield of giants, and even if its founders are right in theory, there's no guarantee that their approach will take off in the market.

What's next: Verses says it will launch a public beta version of Genius in the summer.

Go deeper