Is Artificial General Intelligence Closer Than Tech CEOs Admit?

I was thinking about a talk that hung between certainty and conjecture during a quiet time during the AI summit in Paris. A timeframe of five to 10 years for Artificial General Intelligence was provided by DeepMind’s smart and thoughtful CEO, Demis Hassabis. As usual, things were “complicated.”

The message was yet sufficiently obvious to inspire action in spite of the hedging. AGI is becoming a planning assumption at high-level discussions between government agencies and IT businesses, rather than a hypothetical concept. Some even see it as a strategic necessity.

In the last three years, we have seen a surge in claims from well-known AI figures. AGI will be available in the “reasonably close-ish future,” according to Sam Altman, while Dario Amodei of Anthropic said it will be two or three years away, if advancements proceed as planned. More provocatively, Ali Ghodsi, CEO of Databricks, asserted that AGI may already exist in some form, if not all.

The catch? On the definition of “AGI,” nobody is entirely in agreement. Definitions are still pliable and flexible, sometimes based on measurable standards and other times existing in abstract futures. AGI is either the ultimate goal of machine cognition or a catch-all word for investor deck marketing, depending on the story.

This is not a coincidental ambiguity. Instead, it’s a feature.

Companies maintain plausible deniability despite utilizing significant investment by maintaining a blurred boundary. By the end of 2025, over $130 billion had been invested in AGI-related projects, motivated by both worries of falling behind and aspirations for change. This cash flow changes economic priorities in addition to advancing research.

Language models are being strategically deployed by businesses to streamline processes and free up human talent. Despite not being fully “general,” these models are remarkably adaptable and far quicker at accomplishing a wide range of jobs that were previously only performed by experts. But the transition from tool to thinker is a big one, and public opinion frequently minimizes it.

Particularly enhanced are the simulations. Neural nets called “world models” from DeepMind, which are intended to predict and interact with virtual environments, have shown very comparable patterns of reasoning to those of human game strategies. These models, however, are still limited to customized grid simulations or training environments like StarCraft.” Good generalization is still a challenge.

During a panel, an AI researcher compared these systems to “chess masters who have never left a tournament hall.” I remember that. What a fitting analogy. They struggle with turmoil but thrive on structure.

Through the utilization of extensive datasets, businesses are improving their models to replicate human speech and reasoning. But comprehending and mimicking are two different things. When faced with ambiguity, these systems continue to be vulnerable to hallucinations, frail reasoning, and brittleness. One of the biggest obstacles is still the ability to make decisions based on context, which humans do automatically.

However, optimism is still high. Moreover, with good cause.

Through strategic partnerships, labs have been testing multi-agent systems, which are essentially digital swarms of artificial intelligence agents that cooperate, plan, and negotiate in limited settings. Imagine them as models of AGI societies of the future. Particularly inventive are the outcomes. However, they’re also hard to understand. Artificial systems don’t always behave in ways that make sense to humans.

This uncertainty creates opportunities for advancement as well as overreach.

Geopolitically, artificial intelligence is quickly turning into a battlefield. Countries that had argued over semiconductor policy are now joining forces to secure AI. Western leaders were spooked by China’s DeepSeek, which claims near-AGI performance without using American chips. The European Union is striving to maintain free research while increasing its laws. In the meantime, U.S. regulators alternate between supporting initiatives and examining them reactively.

Making sure that, when it is developed, AGI represents human goals rather than optimizing for unforeseen ones is at the heart of the controversy. The fundamental idea behind the establishment of businesses like Safe Superintelligence is to address alignment before capabilities surpass our capacity.

While confidence in technology increased during the pandemic, it decreased in institutions. Its impetus hasn’t diminished. Furthermore, the concern is not just whether artificial intelligence (AGI) will exist, but also whether we will be able to identify it when it does, as we trust opaque models to make increasingly complicated decisions.

Last autumn, I was sitting at a roundtable when I couldn’t help but feel uneasy. With confidence, a policy expert said that AGI will “solve governance.” In my opinion, it was both dangerously reductive and utopian.

There is more than one type of artificial general intelligence. Unlike a new iPhone, it is not a single system that has yet to be revealed. This constellation of developing technology reflects our hopes, fears, and sometimes even our illusions.

In the years to come, we’ll probably witness systems that feel universal and function remarkably fluidly across disciplines. But we must not mistake superficial proficiency for profound comprehension. Emotion, memory, circumstance, and culture all influence human intelligence, which is complex and multifaceted.

This is being addressed by researchers by incorporating more comprehensive ethical frameworks into development pipelines. There has been progress, and the governance tools have significantly improved. However, this process needs to quicken.

It’s possible that AGI is coming. Or it might always elude us, changing shape every time we approach. What counts most, in either case, is the story we tell about it. We are creating machines that surpass our capacity to control them, or instruments that magnify our values?

Like AGI, the solution will probably be a mess.