Content
summary Summary

Google Deepmind researchers say the ability to act in an open-ended way is crucial for developing superintelligent AI.

Using ever-larger datasets is not enough to achieve ASI, the researchers say. They are referring to current scaling strategies, which tend to focus on using more computing power and more data. Instead, AI systems must be able to generate new knowledge and improve their learning capabilities on their own in an open-ended way.

The researchers define an AI system as "open-ended" if it constantly produces new and learnable artifacts that are difficult to predict, even if an observer has become better at predicting them by studying past artifacts.

Whether AI-generated content is open-ended depends on if an observer sees the generated things as new and if they can learn something from them. For example, a mouse would consider airplanes new but can't understand the principles behind them. For aerospace engineering students, the designs are new and learnable - which makes the system open in this scenario. An alien with extensive aviation knowledge may not find the designs new, but it can analyze and understand them. | Image: Google Deepmind

The researchers give some examples of open systems and their limitations. AlphaGo can constantly develop new strategies to beat learning humans - but only within the rules set by humans. Despite its openness, such a system is too narrow to discover new science or technology, the researchers say.

Ad
Ad

Another example is the "AdA" agent, which solves tasks in a 3D world with billions of possible variations. Through training, AdA gradually gains skills for increasingly complex and unfamiliar environments.

To human observers, this seems open-ended, as AdA keeps showing new abilities. However, the results suggest that the novelty reaches a limit over time and can only be reactivated by adding more content. So openness is limited in time.

The POET system trains agents that are each connected to an evolving environment. By switching between environments, the agents eventually solve very complex tasks that could not be solved by direct optimization.

However, this openness is limited by the settings of the environment itself: At some point POET, like AdA, reaches a limit.

Foundation models based on passive data will soon hit a wall

In contrast, current foundation models, such as large language models, are not "open-ended" because they are trained on static datasets; they would eventually stop producing new artifacts. They will "soon plateau," according to the paper.

Recommendation

However, they could lead the way to open systems if they were continually fed new data and combined with open-ended methods such as reinforcement learning, self-improvement, automatic task generation, and evolutionary algorithms.

Foundation models would bring together human knowledge and could guide the search for new relevant artifacts, the researchers say. This makes them a possible building block for AI's ability to generalize.

"Our position is that open-endedness is a property of any ASI, and that foundation models provide the missing ingredient required for domain-general open-endedness. Further, we believe that there may be only a few remaining steps required to achieve open-endedness with foundation models."

From the paper

The scientists also warn of major safety risks. The artifacts generated must remain understandable and controllable by humans. Safety must be built into the design of open-ended systems from the outset to prevent a superintelligent AI from getting out of control.

But if the challenges can be overcome, the researchers believe that open-ended foundation models could enable major breakthroughs in science and technology.

Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.

Deepmind co-founder Demis Hassabis hinted a year ago that Google Deepmind was testing combining the strengths of systems like AlphaGo with the knowledge of large models.

However, the Gemini model that Google has released so far does not yet offer such breakthrough innovations, and it can barely keep up with OpenAI's GPT-4.

OpenAI's Q* also likely aims to combine ideas like those known from Google Deepmind's AlphaZero with language models - similar to what Microsoft researchers recently showed with "Everything of Thought."

Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Google Deepmind researchers argue that the ability to constantly improve itself and create new adaptive artifacts is a key characteristic of artificial superintelligence (ASI). They call this characteristic "open-endedness."
  • A system is considered "open-ended" if it constantly generates new artifacts that are difficult to predict and learn from the perspective of an observer. Examples such as AlphaGo, AdA, and POET demonstrate varying degrees and limits of openness.
  • According to the researchers, current scaling strategies that rely solely on more computing power and data are not sufficient to achieve ASI. Instead, AI systems must be able to generate new knowledge and improve their learning capabilities on their own. The combination of foundational models and open methods is only a few steps away, they say.
Sources
Online journalist Matthias is the co-founder and publisher of THE DECODER. He believes that artificial intelligence will fundamentally change the relationship between humans and computers.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.