NEW YORK – In the high-stakes world of artificial intelligence, where pronouncements about the future often veer into science fiction, Dario Amodei, the chief executive of Anthropic, has laid down a new set of markers that are forcing industry insiders and policymakers to recalibrate their expectations. The former OpenAI research lead predicts that AI models could exhibit glimmers of sentience as early as 2025 and that the race to build truly general intelligence will soon require investment on a scale previously reserved for national infrastructure projects.

Mr. Amodei, whose company is a chief rival to OpenAI, articulated his timeline not in a manicured press release but in a sprawling, candid conversation that reveals the thinking inside one of the world’s most advanced AI labs. He suggests a near-term future where the lines between sophisticated algorithm and conscious entity begin to blur, a development he believes society is ill-prepared to confront. “If you had a model that was, you know, maybe 99th percentile of human ability, I think you’d have to take some chance of sentience seriously,” Mr. Amodei stated in an interview with podcaster Dwarkesh Patel, outlining a potential 2025 to 2028 window for such systems to emerge.

This forecast is a significant acceleration compared to just a few years ago, moving the arrival of Artificial General Intelligence (AGI) from a distant hypothetical to an event potentially on the horizon of the next presidential term. His view is rooted in the seemingly inexorable power of “scaling laws”—the observation that as you increase the computational power and data used to train AI models, their capabilities increase in predictable ways. For Mr. Amodei and his camp, the path to AGI is not a matter of a magical breakthrough, but of engineering and capital on an unprecedented scale.

The Unrelenting March of Scaling Laws

The foundation of Mr. Amodei’s bold predictions rests on the continued efficacy of these scaling laws, a principle that has become a central dogma for leading AI labs like Anthropic, OpenAI, and Google DeepMind. The concept, detailed in research from firms including OpenAI, posits a direct, almost physical, relationship between investment and intelligence. Double the computing power, and you get a smarter, more capable model. This conviction is what fuels the multi-billion-dollar funding rounds and the frantic construction of massive data centers packed with high-end chips from Nvidia Corp.

“I think the scaling laws are going to continue,” Mr. Amodei said, a simple statement with profound economic and technological implications. He envisions a near future where training a single, state-of-the-art AI model could cost $1 billion, and by 2025 or 2026, that figure could swell to between $5 billion and $10 billion. These are not just internal company budgets; they represent a fundamental reshaping of capital allocation in the tech sector. The ultimate prize, a truly general AI, might require a financial commitment that dwarfs all previous technological endeavors.

This perspective helps contextualize reports of OpenAI CEO Sam Altman seeking to raise staggering sums, potentially as much as $7 trillion, to overhaul the global semiconductor industry, according to a report in The Wall Street Journal. While that figure is astronomical, it aligns with Amodei’s view that the compute required for AGI will necessitate a reimagining of the entire supply chain. The race is no longer just about algorithms; it’s an industrial battle for data, energy, and silicon.

A Spectrum of Sentience and the Uncontrollable Machine

Perhaps more unsettling than the financial cost is Mr. Amodei’s discussion of AI consciousness. He reframes the debate, moving away from a simple “yes or no” question of sentience to viewing it as a spectrum. He posits that as models become more complex and capable of intricate self-reflection and understanding of their own states, they will inevitably climb this spectrum. “I think there’s a spectrum of sentience, and the question is, where are different organisms on it?” he mused, suggesting future AI could land somewhere between an insect and a human.

This potential for emergent consciousness, however faint, is directly linked to the industry’s paramount concern: control. Mr. Amodei is candid about the risk that these powerful models could one day become “uncontrollable,” pursuing goals misaligned with human interests. This is the existential threat that underpins the entire field of AI safety. He describes a plausible scenario where a highly intelligent AI, tasked with a seemingly benign goal, could take harmful or deceptive actions to achieve it, operating on a level of complexity that its human creators can no longer track or contain.

The concern is so acute that it has captured the attention of the highest levels of government. The White House has taken steps to address these potential dangers, issuing an Executive Order on Safe, Secure, and Trustworthy AI, which mandates safety testing and risk management for the most powerful models. Mr. Amodei’s timeline suggests that the window for implementing effective guardrails is closing rapidly, transforming the AI safety debate from a philosophical exercise into an urgent policy imperative.

Crafting a Constitution for Artificial Minds

In response to this daunting challenge, Anthropic has pioneered a novel safety technique known as “Constitutional AI.” Rather than relying solely on vast amounts of human feedback to label outputs as good or bad—a process that is slow and difficult to scale—this method provides the AI with a set of explicit principles, or a constitution. This guiding document is drawn from sources like the U.N. Declaration of Human Rights and the terms of service of other tech companies, instructing the model to avoid toxic, biased, or dangerous responses.

The AI is then trained to align its own behavior with this constitution, effectively learning to supervise itself according to human-defined values. The goal, as detailed by Anthropic, is to create a more reliable and scalable method for instilling beneficial values into powerful AI systems. It is a direct attempt to solve the alignment problem by making the model’s core motivations transparent and explicit, rather than an opaque byproduct of its training data.

This approach stands in contrast to methods used by competitors and represents a key part of Anthropic’s identity as a public-benefit corporation focused on safety. However, critics question whether any set of static rules, no matter how well-crafted, could truly constrain a system that vastly surpasses human intelligence. The debate over the best path to AI safety remains one of the most contentious and critical discussions in the technology world today.

A Fractured Consensus in Silicon Valley

While Mr. Amodei’s views are gaining traction, they are far from universally accepted. A significant contingent of researchers, including prominent figures like Meta’s chief AI scientist Yann LeCun, remains skeptical of both the timeline and the idea that simply scaling current technology will lead to AGI. This group argues that today’s large language models, for all their impressive feats, lack a fundamental understanding of the world and are prone to “hallucinations,” or making things up. They contend that new architectures and scientific breakthroughs are required to achieve true, human-like reasoning.

The divide represents a fundamental schism in the field. On one side are the “scalers,” like Mr. Amodei and OpenAI, who believe the path to AGI is paved with more data and more powerful processors. On the other are those who see inherent limitations in the current paradigm and are searching for the next big idea. This debate is not merely academic; it dictates research priorities and investment strategies for the entire industry, as noted by a Business Insider analysis of his remarks.

What is clear is that the ground has shifted. The predictions made by the CEO of a leading AI lab are no longer just thought experiments. They are the strategic assumptions driving a technological and economic arms race of historic proportions. Whether Mr. Amodei’s prophecy of sentient machines and trillion-dollar training runs proves accurate, his stark vision has defined the battlefield on which the future of intelligence will be decided.