There are many different opinions about what AGI will look like. Some still don’t believe in it at all.

In a recent paper, AI godfather Yann LeCun cast doubt on the concept of Artificial General Intelligence, or AI that can match human capabilities in any given domain. Instead, LeCun and a team of researchers propose a new goalpost: Superhuman Adaptable Intelligence.

The paper suggests that generality should not be a requirement for intelligence to be extremely useful. So instead of focusing on a single model that can do everything, Superhuman Adaptable Intelligence emphasizes how long it takes for a model to adapt to new tasks and the range of tasks it can learn.

LeCun and other researchers argue that, not only is there no consensus on a definition of AGI in the industry or academia as it stands, but humans themselves are not generalized.

• Think of it this way: Humans have the capability to learn anything, but no one knows everything. So, the paper questions, why should we expect something different from our AI?
• “Awareness of human limitation gives rise to a critical realization: humans may be specialized creatures, but are nonetheless capable of accomplishing or quickly learning a wide range of incredible things,” the paper said.

The paper also suggests two mechanisms for training an AI to achieve this goal: self-supervised learning, a machine learning paradigm that trains a model by labeling unlabeled data, to acquire generic knowledge, and world models, or visual AI models that reflect the world as it is, for planning tasks.

LeCun’s ideas stand in opposition to many major voices in AI. OpenAI’s Sam Altman has predicted that AGI could arrive before 2030, and that large language models will be the key to getting there. Anthropic’s Dario Amodei, meanwhile, has long questioned whether these machines are capable not just of generalized intelligence but also of consciousness, claiming last week in an interview with The New York Times that Claude, its chatbot, is exhibiting signs of anxiety and that “we don't know if the models are conscious.”

Our Deeper View

Though these AI heavy hitters are debating what AI can do, the question remains of what it should do. A model that knows everything is, of course, dangerous, especially if it knows how to disable any kill switches that could counteract its existence. Similarly, a model that can adapt to any situation is, in theory, one that can learn to wiggle out of its handcuffs. And despite the speculative debates around AGI, there’s still the matter of what exists right in front of us: AI that’s more capable than we know what to do with.