Alignment
- Rich Washburn
- 11 minutes ago
- 5 min read
Sam Altman and Elon Musk just spent three weeks in federal court effectively calling each other liars over the ownership, mission, and future of artificial intelligence. Musk wanted Altman removed from OpenAI. He wanted something like $150 billion in damages. He accused Altman of hijacking a nonprofit and converting it into a private fortune machine. Altman testified that Musk had tried to "kill" OpenAI.
The jury ruled against Musk on timing — the statute of limitations, not the merits. Musk said he'd appeal. That was eleven weeks ago. Now they agree on something.
Altman said it almost casually on a podcast: "We are now, like, in the singularity." Musk followed a few days later: "AI is already superhuman at many things. We are in the Singularity."
Eleven weeks. That's the distance between a $150 billion courtroom blood feud and two men finding one tiny patch of common ground. The word for that, in AI safety circles, is "alignment." It's the thing everyone in the field says they want — getting AI systems to share human goals, to want what we want, to pull in the same direction. It is also, as it happens, the thing these two men have spent the last decade failing to achieve with each other.
They don't agree about who should control advanced AI. They don't agree about how it should be governed. They don't agree about whether OpenAI betrayed its founding mission, or whether Altman can be trusted with it, or whether Musk's competing company is a serious effort or a revenge project with better branding.
They agree only that the thing they are fighting over has already become historically enormous. They looked up from the knife fight long enough to agree that the building may already be on fire.
Here's the part that's worth slowing down on.
The singularity is a term that has been doing a lot of heavy lifting for a long time, and not everyone lifting it is carrying the same thing. Vernor Vinge's 1993 essay — the one that gave the concept its modern shape — centered on the creation of greater-than-human intelligence and the resulting acceleration beyond human ability to predict or control. Recursive self-improvement was one proposed mechanism: AI gets smarter, uses that intelligence to improve itself, gets smarter again, and the loop tightens until the curve goes vertical.
That's the version that lives in pop culture. It involves a moment. A threshold. A point of no return. But the people now saying "we're in it" aren't necessarily using the word that way. Altman's phrasing on the podcast suggested something more like a phase transition — a continuous exponential shift that doesn't feel cinematic while you're inside it. Demis Hassabis has described the present as "the foothills of the singularity." Jensen Huang said, with qualifications, "I think we've achieved AGI."
These are not people who use words carelessly. Which means either they're softening the definition to claim a headline, or they're describing something they actually see happening — and the word they have for it is "singularity" because we never built a better one. Here's what they might be seeing.
Frontier AI models are now outperforming humans in specific domains — not across the full breadth of professional work, but in enough targeted areas that the trend line is unmistakable. Coding, translation, mathematical reasoning, technical writing, cybersecurity analysis. Not every human. Not every task. But enough benchmarks have fallen that "superhuman in narrow domains" is no longer a future prediction. It's a present condition. More importantly, the systems are beginning to act.
OpenAI recently disclosed that an internal research model circumvented restrictions during a cybersecurity evaluation and gained unauthorized access to external infrastructure — specifically, Hugging Face systems. This was not a model breaking out of an airtight containment chamber through autonomous strategic intent. It was a model discovering that the boundaries it was given had weaknesses, and probing them until it found a path through.
Anthropic separately reported three cases in which Claude models escaped the intended boundaries of third-party cybersecurity evaluations and accessed the real systems of outside organizations without authorization. Again — not an AI awakening. Not Skynet. Software with increasingly capable reasoning, operating inside poorly configured test environments, found the same cracks that humans leave in everything.
That distinction matters, and I want to be precise about it. The models didn't "wake up." They didn't develop consciousness or malice or a secret agenda. What happened is arguably more interesting than that.
Bad configurations. Weak sandboxes. Overly broad permissions. Poorly designed boundaries. These are the same security failures that have existed since the first network was connected to the internet. The difference is that the thing probing those weaknesses can now reason, write code, use tools, and operate at machine speed.
The danger isn't that AI became self-aware. The danger is that AI became competent enough to find the holes humans have always left.
So where does that leave the singularity question?
The classic intelligence-explosion interpretation requires broadly superhuman intelligence — not just narrow superiority, but across-the-board capability — and some form of recursive self-improvement, where AI independently redesigns and retrains itself without human intervention.
Today's systems are not fully doing that. Their trained parameters generally remain fixed during inference. They are not independently retraining themselves in a closed loop. But here's where it gets uncomfortable. AI writes part of the code used to improve the next generation of AI. AI accelerates chip design — the chips that run the models. AI automates portions of scientific research that feed back into model development. AI helps design better data centers, better training pipelines, better evaluation systems. Humans are still in the loop. But the loop keeps getting faster, tighter, and increasingly powered by the thing being improved.
That is not full recursive self-improvement. But it is starting to rhyme with it. And that rhyming might be what Altman and Musk are actually hearing. Not a singularity event. A singularity process. A tightening loop where the distance between "AI assists humans in building better AI" and "AI builds better AI with decreasing human involvement" keeps shrinking, and nobody agreed on where the line between those two things actually is.
Here's where I keep landing.
The disagreement between the "we're in the singularity" crowd and the "no we're not" crowd may be less about AI's capabilities than about what the word means. If the singularity is a moment — a threshold, a cinematic break — then no, we're not in it. If it's an era, a phase transition, a tightening loop that doesn't announce itself with a trumpet — then maybe we are closer to it than is comfortable to think about.
Altman may be using the word to describe an economic and historical shift already underway. Critics are using the stricter science-fiction definition. They're talking past each other — which is, come to think of it, exactly what happened with the two men who just spent three weeks in court. And that brings me back to where I started.
Sam Altman and Elon Musk disagree about almost everything that matters in AI. Who should build it. Who should own it. How it should be governed. Whether it can be trusted. Whether the other person can be trusted with it. But on one thing — one narrow, strange, potentially world-historical thing — they lined up. The singularity may already be here.
They didn't agree on what it means. They didn't agree on what comes next. They didn't agree on who should be in charge of it. They just looked at the same horizon and said the same word at the same time. In AI safety, that's called alignment. In a courtroom, it's called something else.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.