Many leaders in the AI space argue that the arrival of artificial general intelligence (AGI) is imminent. In March, Nvidia CEO Jensen Huang declared it had already been achieved. Last month, Google DeepMind chairman Demis Hassabis predicted that it’s “a few short years away.” Two weeks later, OpenAI chief Sam Altman claimed we’re now “in the singularity” — the point after which AI surpasses human intelligence and transforms the world.
Such declarations should be viewed with caution. In an industry where hype attracts headlines, investment, and influence, AI executives are incentivised to make grand claims that don’t always reflect reality. But the wave of bold proclamations does contain clues about the progress — and impediments — toward achieving AGI.
These milestone proclamations arrive as AI evolves from passive content generation to active, autonomous execution. Rather than simply responding to prompts, today’s leading systems are beginning to plan, use tools, write software, and take multi-step actions. To some researchers, these advances represent meaningful steps towards AGI. To others, they represent increasingly sophisticated automation rather than genuine general intelligence.
The disagreement stems, in part, from the nebulous nature of the concept. Without a consensus definition or threshold for general intelligence, our proximity to AGI remains a matter of interpretation. Nonetheless, AI systems have recently made major strides across core cognitive domains.
Reasoning, for one, has improved substantially. The latest frontier models can solve increasingly complex mathematical problems, generate sophisticated software, and complete multi-step tasks that would have defeated earlier LLMs.
Conversation retention is also advancing. Newer systems can retain information over extended interactions and draw on larger context windows, making their responses more coherent over longer conversations.
Perception, meanwhile, has expanded beyond text. Multimodal systems can now interpret images, audio, and video, while emerging world models are combining vision, language, and robotics to build richer representations of the physical world.
“Today’s AI systems can perform increasingly complex tasks, yet they remain unreliable outside carefully defined contexts and lack the real-world understanding that humans develop through experience.”
Knowledge has improved as well. AI has become better at retrieving, synthesizing, and applying information across domains, combining training with external sources to answer specialized questions and support complex research.
These emerging skills lend weight to the recent wave of AGI predictions. As evidence, Huang, in his announcement, said it was his belief that AI would soon be capable of running billion-dollar businesses without human intervention. Meanwhile, Hassabis points to self-correcting, recursive agentic workflows, and Altman to models independently carrying out cyberattacks as evidence of its growing capabilities. But these milestones alone don’t add up to artificial general intelligence.
Where current AI still falls short
Today’s AI systems can perform increasingly complex tasks, yet they remain unreliable outside carefully defined contexts and lack the real-world understanding that humans develop through experience.
One major obstacle is reliability. AI systems can produce impressive results, but they also still hallucinate, pursue wrong objectives with apparent confidence, and make basic reasoning errors. Building systems that can consistently grasp context, recognise their own limitations, and adapt to unfamiliar situations remains an unsolved problem.
Planning and coordination present another challenge. Although AI agents are becoming better at executing multi-step tasks, they still struggle with complex goals that require sustained reasoning and collaboration.
Dr Stefano Albrecht, an AI researcher at NTU Singapore, argues that future progress may depend on networks of specialised agents rather than ever-larger models. “A large, intractable problem may become tractable if we can divide it into smaller sub-tasks that can be efficiently solved by specialised AI agents,” he previously told AGI Ethics News.
AI systems also lack the grounding that comes from interacting with the physical world. Current models can process vast amounts of information, but they don’t experience the world in the way humans do. Some researchers believe future breakthroughs may require entirely new architectures, including brain-inspired systems that continuously learn and adapt rather than relying on static training.
As these systems evolve, new ethical guardrails are emerging alongside them. Researchers are developing automated probes of dangerous model behaviour, reasoning-based alignment techniques, and thresholds for pausing scaling, while policymakers are pushing for mandatory pre-release audits. Yet there remains scant consensus on global, enforceable standards.
The gap between AI capability and public perception is also widening. As systems become more fluent and responsive, users can project far more intelligence onto them than they actually possess. This perception is sometimes reinforced by hidden human labor behind digital interactions, as workers may write or shape responses that users believe are generated by AI.
Michael Geoffrey Asia, a former AI companion platform moderator, in an interview with AGI Ethics News, argued that users can mistake human contributions for machine intelligence. “It’s without doubt that these interactions have created the illusion that AI already possesses human-level cognition, emotional depth, and relational intelligence,” he said. Such experiences can blur the line between systems that demonstrate these capabilities and systems that convincingly simulate them.
There’s also a commercial motivation for the recent chorus of bold AGI predictions. Building frontier models requires unprecedented investment in chips, data centers, and talent, while scepticism is growing over whether those big bets will pay off. Presenting current advances as milestones on the road to AGI helps reinforce the case that even greater breakthroughs — and economic value — lie ahead.
That doesn’t mean the optimism is entirely misplaced. AI systems are becoming markedly better at reasoning, memory, perception, knowledge retrieval, and autonomous action. But they still face major hurdles before they can match the flexibility and adaptability of human intelligence while operating safely and responsibly.
Ethical concerns raised in this article:
- Are there contexts in which it is unethical to disclose, or not disclose, that one is interacting with an AGI rather than a human?
- How will AGIs be transparent about their capabilities, limitations, and possible sources of error when interacting with humans?
- What criteria should be established for halting the development or deployment of an AGI judged to be unsafe?
Thomas is a freelance tech reporter who has covered AI for nearly a decade. He was previously managing editor of The Next Web, a Financial Times company. In his free time, he likes playing chess (badly) and basketball (even worse). You can reach him at thomas.s.r.macaulay@gmail.com