The Japanese startup Integral AI has announced that it has created the “world’s first AGI-capable model.” According to a company representative, the team has achieved one of the ultimate goals of artificial intelligence development: enabling their AI model to learn new tasks “without pre-existing datasets or human intervention.”
“This is more than just a technical achievement; it marks a new chapter in the story of human civilization,” said Jad Tarifi, PhD, CEO and co-founder of Integral AI, in a statement published via Businesswire. “Our task now is to scale this model, whose AI-powered capabilities are still in their infancy, into an embodied superintelligence that expands the freedom and possibilities of collective governance.”
Integral AI proposes its own three-part definition of AGI.
The first is Autonomous Skill Learning: the model must independently acquire new skills in new domains without pre-prepared datasets and without human involvement.
The second is Safe and Reliable Mastery: learning must take place without “catastrophic” side effects; as the company explains, a hypothetical “kitchen robot” learning to cook should not start a fire.
The third is Energy Efficiency: the total energy cost of training should be comparable to that of a human learning the same skill.
Integral AI claims that its multi-layer architecture is inspired by the structure of the neocortex and allows the system “to grow, abstract, plan, and act” as a unified whole. In the reported early robotics trials, robots allegedly acquired new skills in real-world conditions without direct supervision.
The claims are extremely bold. It is important to maintain a measure of healthy skepticism here. The weakest points of the announcement are obvious.
First, there is no public technical report or peer-reviewed paper describing the architecture, experiments, and metrics.
Second, there are no independent benchmarks or reproducible demonstrations that would allow external researchers to test the claim of “self-learning without datasets,” and in particular to evaluate the energy-efficiency criterion in comparison with humans.
Third, the phrasing “world’s first AGI-capable” rests on the company’s own definition of AGI, which is not an industry-wide standard.