Sam Altman Says We're Already in the Singularity. The Real Question Is What Happens to Human Agency Next.
A few days ago, in a podcast, Sam Altman, the CEO of OpenAI, made an extraordinary, almost casual remark: "We are now, like, in the singularity." This isn't the first time he's hinted at it. His 2025 essay, The Gentle Singularity, spoke of being "past the event horizon" and "close to building digital superintelligence."
If the word "Singularity" sounds like something out of a futuristic movie, you're not wrong. But let's break it down in a way that makes sense.
One influential conception of the Singularity is the point at which AI becomes capable of recursive self-improvement. Imagine an oven that could eventually learn from every cake, improve how it works, redesign parts of its own process, and repeat that cycle faster and faster. This idea suggests a point where "normal" rules of progress no longer apply, and intelligence grows beyond our comprehension.
Altman's point is that he believes the transition has already begun. What exactly has begun is where the argument starts.
Belief vs. Technical Reality: The Crucial Nuance
It's vital to note the nuance here. Altman believes we are in the singularity. He did not announce that OpenAI has achieved the Singularity as a technical milestone, nor has he claimed OpenAI has secretly built superintelligence. In fact, he's clarified the opposite in the past.
This distinction matters. Researchers like Stuart Russell and Roman Yampolskiy actively dispute Altman's characterization, arguing that today's AI systems, while impressive, still rely heavily on human direction and don't yet demonstrate totally independent self-improvement.
This disagreement isn't just academic. It exposes a more immediate and critical problem for anyone building or using AI: the shift from AI that gives you answers, to AI that acts on your behalf, to AI that increasingly influences what you decide.
Agentic AI vs. AI with Agency: The Defining Difference
The AI world has spent a lot of time on the shift from AI that gives answers (like asking a chatbot for information) to AI that acts (like an agent that finds and books flights for you). But the next, more profound shift is from acting to deciding. This is where the difference between "Agentic AI" and "AI with Agency" becomes crucial.
- Agentic AI: Think of this as a very smart, highly capable assistant. You give it a clear job: "Find me flights to London," or "Find the best one and book it within these constraints." It then plans the steps, uses tools (like travel websites), and gets the job done. It's excellent at pursuing an assigned objective. Its actions are always about achieving your goal.
- AI with Agency: At Recursum, we use "AI with Agency" to describe a further step: systems with meaningful discretion not only over how objectives are pursued, but over how objectives are interpreted, prioritized, revised, or proposed. It's not just following instructions. It's participating in decisions about what deserves attention in the first place. For example, you ask it to plan your London trip. But based on your priorities, schedule, budget, and reasons for traveling, it challenges the objective itself: "Before I book London, there's a bigger question. Given what you're actually trying to accomplish, is London the right trip?"
The industry often uses "agent" for both, but that hides a big difference. If Altman's framing means we have entered a period of accelerating machine capability, then understanding this difference is no longer just for scientists. It's about who's in charge, what risks we're taking, and how we'll work with these powerful systems.
The New Governance Problem: Shaping Judgment, Not Just Execution
AI already makes countless small decisions inside objectives we give it: what information to prioritize, which route to take through a problem, which tool to use, what to recommend. The deeper question begins when that discretion moves upstream, from deciding how to pursue our goals toward influencing which goals deserve pursuit.
The question isn't whether AI can make decisions. It already does. The question is whether AI begins participating in decisions about what should matter.
This is where the real challenge lies. The strategic issue isn't that AI with Agency is simply "more powerful." It's that agency fundamentally changes the governance problem.
With an Agentic AI, the main question for control is: Did it do the job I gave it correctly?
With AI with Agency, the question becomes: How did it determine what the job should be? This is a radically different problem. When an AI can influence or even generate objectives, we can't just put fences around its behavior. We have to address how the system arrives at judgments.
As AI gains discretion, the problem shifts from controlling execution to calibrating judgment.
Recursum's Framework: Cognitive Attunement
This is precisely where Cognitive Attunement comes in.
Cognitive Attunement, a framework championed by Recursum CEO Elise Verdugo, is the process by which an AI system becomes calibrated not merely to a user's instructions, but to their judgment patterns, priorities, constraints, contradictions, and evolving intent.
Personalization learns what you like. Cognitive Attunement learns how you judge.
This "learning" isn't about mind-reading, but about building an evolving model from your observable choices, corrections, priorities, constraints, and contradictions. For example:
- Personalization sees: Ernesto usually rejects X, so it recommends Y.
- Cognitive Attunement sees: Ernesto usually rejects X because he prioritizes authority over reach. But in this situation, that principle may conflict with another priority. Which should govern the decision?
Our work explores what would be required for increasingly autonomous AI systems to remain calibrated to human judgment as their discretion expands. The problem isn't merely, "How do we align AI with human judgment?" It's also, "What happens when AI begins shaping the human judgment it is supposed to align with?" As AI becomes better at judgment, how do we prevent human judgment from quietly becoming ceremonial? That's a serious question.
Maybe the most important question about the Singularity isn't when AI surpasses human intelligence. It's when convenience causes us to surrender human judgment before we realize judgment was what made the agency ours.
The challenge, then, is not simply building AI capable of deciding. It is ensuring that as machines gain agency, humans don't surrender theirs.
TL;DR: Sam Altman's recent statement that "we are in the singularity" highlights AI's accelerating capabilities. This makes the distinction between "Agentic AI" (AI executing human-assigned tasks) and "AI with Agency" (AI participating in objective-setting) crucial. As AI moves to influence decisions, the governance challenge shifts from controlling actions to calibrating how the system arrives at judgments. Recursum, led by CEO Elise Verdugo, is exploring this problem through Cognitive Attunement, a framework for thinking about how increasingly autonomous AI might remain calibrated to human judgment without gradually replacing it.
By Ernesto Verdugo AI architect, founder of Verdugo Labs, and creator of Recursum. He builds systems for what happens when human and artificial intelligence stop working separately.
Links: