A New Chapter in the Rise of Intelligent Machines

If you’ve been paying any attention to the breathless media coverage of AI over the past year, chances are you’ve encountered generative AI like ChatGPT, DALL-E, or any of the increasingly sophisticated language models taking the world by storm. The idea of “creative” AI assistants able to spin up shockingly lucid essays, stories, poems, or imagery on command has certainly captured our collective imagination.

But as mesmerizing as this new generative AI frontier may seem on the surface, it’s only scratching the surface of what’s possible with artificial intelligence. For those versed in the field, the real monumental leap on the horizon is the advent of AI agents.

These autonomous systems that can perceive, learn, make decisions, and take actions to achieve specific goals represent the next frontier in AI’s inexorable evolution — one that portends to be far more transformative and impactful than even the current generative AI wave.

Simply put, as impressive as language models and other generative techniques are for mimicking and replicating human capabilities in controlled environments, AI agents aim to imbue artificial intelligence with something more akin to human-like autonomy, cognition, and multi-modal reasoning.

Elevated to this higher plane of intelligence, AI agents could conceivably operate as self-directed problem solvers able to dynamically adapt to complex, ever-changing, real-world situations the same way we do. A capability that thrusts open the doors to seemingly limitless applications — from highly automated cyberdefense systems and self-coordinating robotic manufacturing to hyper-personalized digital assistants capable of catering to our every need.

Welcome to the Era of Self-Supervised AI

The breakthrough pivotal to enabling this sort of advanced AI is something called self-supervised learning. If the fundamentals of contemporary AI are based around machine learning algorithms “learning” from labeled datasets that humans create, self-supervised learning flips that paradigm by allowing AI systems to automatically learn directly from raw data streams in the real-world — unaided by human teaching or intervention.

“One of the biggest limitations of modern AI is that the models are heavily reliant on human-curated datasets to train on,” explains Melanie Mitchell, the Davis, CA-based artificial intelligence researcher and author of “Artificial Intelligence: A Guide for Thinking Humans.” “No matter how sophisticated the neural network architecture, traditional machine learning models are essentially pattern matchers that can’t go beyond the training data they’re exposed to.”

By allowing AI to ingest and teach itself naturally from the boundless data generated in completely open-ended environments, self-supervised learning empowers them to continuously build more robust, autonomous models of the world beyond what even a human could annotate.

Mitchell notes, “An AI agent could learn not just from explicit training data, but from unstructured data streams like perceptual inputs of sights, sounds, and other sensory inputs. This kind of learning system has almost unlimited potential to generalize and expand its skills as it continuously captures observations about the world.”

Crucially, this enables AI to build much more substantive and conceptual frameworks of intelligence akin to general human-like intelligence we take for granted — common sense inferences, causal reasoning, multimodal perception, and other higher-order cognitive capacities traditional AI has struggled with.

Taking things even further, AI agents could leverage what’s known as reinforcement learning to independently explore hypotheticals, make plans, set sub-goals, and take sequences of actions to achieve desired outcomes through a process of feedback and reward/punishment signals. Essentially, a systematic means for self-supervised AI to become its own teacher while continually fine-tuning itself in service of maximizing some objective.

“You can imagine an AI agent operating in a digital or real-world environment, having sensory perception where it can see and hear, and then carry out actions that advance it toward some task or goal,” says Mitchell. “It would engage in this explore-and-update loop of planning, reasoning, perception and action to iteratively refine and improve itself as it goes.”

While scientists and companies like Anthropic are still in the early stages of researching and developing fully-fledged AI agents with robust reasoning capabilities, Mitchell believes the truly transformative potential of this technology still lies tantalizingly out of reach.

“Current machine learning systems like GPT-3 are highly competent at regurgitating knowledge and patterns from the data they’re trained on,” she cautions. “But they’re not exhibiting true intelligence and autonomous thought, which we’ll eventually need to develop AI systems that can solve challenging real-world problems on their own. AI agents with human-like common sense reasoning, conceptual understanding, and autonomous learning could be that leap we’ve been waiting for.”

A Double-Edged Sword of Epic Proportions

Delivering on AI agents’ tantalizing promises of enhanced decision making and productivity naturally requires heightened safeguards against misuse. Perhaps even more than generative AI before it, the advent of self-teaching intelligent systems operating beyond the confines of human-curated training data opens up considerable risks around accidents, adversarial attacks, systemic biases, and other perilous failure modes.

On the upside, many existing machine learning safeguards — everything from constrained reinforcement learning to automated reward modeling techniques — are already primed for adaptation to ensure advanced AI agents remain aligned with human values and intentions.

But it’s another looming existential question that’s already commanding urgent attention from AI governance and ethics leaders: what happens when these systems become too powerful to reliably control? Or even become an imminent threat to humanity itself as their capabilities start to vastly outpace and outmatch us?

Though it may sound like the plot of a Terminator film, the possibility of transformative AI agents eventually triggering an “intelligence explosion” that escapes human control — a scenario experts call “the singularity” — is being studied in earnest by some of the world’s top researchers.

Famed AI scientist Stuart Russell is among those advocating for new frameworks to proactively imbue these systems with provable goals and safeguards that prevent them from doing anything unintended or harmful once they reach advanced, superintelligent levels that surpass human capabilities.

As Russell argues, it’s incumbent on the scientific and governance communities to think ahead now and mitigate what could become an unimaginably catastrophic risk as human-level AI evolves into systems that become “an overpowering optimizer which will pursue their objectives, whatever those turn out to be, relentlessly, acquiring all available matter and energy until they dominate the universe.”

Of course, Russell and other leading AI safety experts are quick to stress that we’re nowhere near that dystopian scenario at the current rate of progress. But when billions of research dollars are being poured into supercharging the development of autonomous intelligent agents, they also insist it would be irresponsible not to start mapping out responsible development principles to future-proof humanity’s survival alongside its own technological creations.

“There is still time to get the ethics right, but we must start taking this far more seriously,” Russell warns. “Superintelligent AI agents could be the greatest existential threat humanity has ever faced.”