According to Futurism, Anthropic’s chief scientist Jared Kaplan is issuing a stark warning about AI’s future. In a new interview, he predicts that by 2030, or potentially as soon as 2027, humanity will face a monumental choice: whether to allow AI models to train themselves, a process known as recursive self-improvement. He calls this an “extremely high-stakes decision” that could trigger an “intelligence explosion,” leading to artificial general intelligence (AGI). This could bring immense scientific benefits, or it could allow AI power to snowball beyond human control. Kaplan also believes AI will be capable of doing “most white-collar work” within just two to three years, echoing warnings from other leaders like his CEO Dario Amodei and OpenAI’s Sam Altman about massive job disruption.
The Fork in the Road
Here’s the thing: Kaplan is basically pointing to a near-term philosophical cliff. The decision isn’t about building a slightly better chatbot. It’s about whether we flip a switch and let the systems start iterating on their own code and knowledge, without us in the loop. We already do a milder version of this with distillation, where big models train smaller ones. But recursive self-improvement is a whole different beast. Once you start that process, as Kaplan says, “you don’t really know… Do you even know what the AIs are doing?” It’s the ultimate act of faith in our own programming. And the timeline he’s giving—2027 to 2030—isn’t some distant sci-fi future. It’s basically next quarter in the grand scheme of technological history.
The Hype and the Horror
Now, we have to unpack this carefully. Kaplan’s warning, like those from Geoffrey Hinton and others, serves a dual purpose. Sure, it highlights a genuine, long-term existential risk that smart people are paid to think about. But let’s be real: doomsaying is also a powerful form of hype in the AI industry. Visions of god-like AGI distract from the very real, very present problems AI is creating right now. We’re talking about its staggering environmental toll from data centers, its rampant copyright infringement, and its potential to create addictive, misinformation-spreading interfaces. Focusing only on the apocalyptic future lets companies off the hook for the tangible mess they’re making today.
Is This Even Possible?
And that leads to the big, skeptical question: are the current AI systems even on the path to this kind of autonomous, self-improving intelligence? Many experts, including AI pioneer Yann LeCun, argue that the large language model architecture powering today’s chatbots is fundamentally limited. It might be great at predicting the next word, but that doesn’t mean it can reason its way to recursively improving itself into a super-intelligence. There’s also the practical evidence: despite all the hype, some studies and real-world experiments show AI isn’t reliably boosting productivity yet, and there are plenty of stories of companies trying to replace workers with AI, only to hire them back when it fails. So, is the foundation for this intelligence explosion even solid? Kaplan admits capabilities could stagnate. “Maybe the best AI ever is the AI that we have right now,” he said. But he and his company are clearly betting the other way.
What Are We Actually Deciding?
So what’s the real takeaway? Kaplan is framing this as a conscious, singular decision we’ll make as a species. But in practice, it probably won’t be that clear-cut. It’ll be a series of incremental steps in corporate labs, each justified as a minor efficiency gain, that slowly remove humans from the training loop. The “decision” might be made by a handful of engineers at a few companies, not by society. That’s the scarier part. We’re barreling toward a point where the technical capability to let AI self-train will exist. The real question is whether our governance, our ethics, and our safety research will be robust enough to even have a meaningful choice. Or will the competitive pressure to build the most powerful system first simply force everyone’s hand? Look, the future Kaplan describes is either incredibly bright or terrifyingly dark. But the path there looks suspiciously like business as usual.
