Slow Singularity
The story we tell ourselves about Artificial General Intelligence (AGI) is primarily one of sheer Hollywood drama: a sudden, world-flipping event. A "hard takeoff." One day, we are puttering along, and then suddenly, a digital god emerges from the server racks, rendering humankind either irrelevant or extinct.
It makes for a great story, but it is wrong.
What I see happening is not a sudden shockwave. Rather a gradual, disruptive rise like the Internet's invention, which reshaped our world so subtly at first that we only realized its magnitude in hindsight. A slow singularity. By a slow singularity, I am referring to a slow build up of special (super-human) capabilities that through integration will create a forms of generality across systems but not necessarily consciousness.
And the danger is that we are treating this slow shift as if it were not urgent. As if we are not already living inside the event.
AGI Isn't a Person, It's a Process
The first mistake we make is picturing AGI as a singular conscious entity, a ghost in the machine waiting to wake up. That's not what the evidence suggests.
Instead, what we're witnessing is the emergence of a mosaic of superhuman skills. Everything including “agency” are stitched together through tools, APIs, and orchestration layers. As Matthew G. Johnson’s recent AGI Report Card points out, models today routinely perform at PhD levels in specific tasks while still stumbling on common-sense reasoning.
Because capability is modular, liability is too. Each new skill, code synthesis, market trading, scientific discovery, creates its own regulatory blind spot. Bind enough of those blind spots together and instead of a supervillain, you have an unaccountable system that no single agency or nation can limit or control.
That system spans clouds, APIs, and human operators, so responsibility scatters just as widely.
Intelligence does not live in the model weights alone; it lives in the supply chain of models, tools, and humans. We are not building a sentient mind so much as an increasingly capable engine for action. And as our understanding of how to build these building these workflows improves and unlock more real-world usefulness, AGI seems to be arriving in pieces.
And because AGI is arriving in chunks, we have been adapting to it in a reactive and somewhat chaotic way.
National strategies for the "AGI transition" keep stalling as corporate deployment outruns them. The bottom-up adaptation continue to be messy as an emergent product of millions of individual and institutional responses.
Our Reactive Dance with Technology
We are course-correcting in real time, and not exactly designing a new society.
When AI-generated misinformation threatens an election, we scramble to build new content verification tools and public awareness campaigns.
When artists see their styles replicated, legal and ethical battles erupt over copyright and data provenance.
When companies discover AI can dramatically boost productivity, they invest, and the workforce has to adapt.
This is not a planned transition as we are simply implementing solutions, redesigning infrastructure, and adapting systems, all while facing the continued challenges brought by these changes. It is an adaptation, yes, but not the kind we need. These responses are fragmented, uneven, and often too late. And it is worrying that we may normalize the chaos, instead of preparing for what’s next.
Today’s patchwork fixes target symptoms rather than addressing the underlying system. Each isolated reaction fuels market pressure for faster implementation, creating a positive feedback loop: more disruptions lead to quicker deployments, which in turn result in even less oversight.
This cycle highlights why fragmented approaches to governance are failing. A coordinated new approach to policy effort must begin immediately, not when an imagined "true AGI" finally emerges.
The Governance Problem is Already Here
This brings me to the most critical point: the need for governance is not a future problem to be solved before we can "unleash" AGI.
For example, the EU AI Act’s 2024 rules focus on transparency and documentation for general-purpose AI (GPAI), but they don’t require full security checks before deployment. It gives flexibility for lower-risk uses but could miss risks with powerful systems, like those with 100-billion parameters. While rules require clear documentation and public awareness, they don’t mandate universal security audits. This shows how hard it is for regulations to keep up with fast-changing AI, stressing the need for regular updates to stay safe and responsible.
However, open-weights AI labs and closed-weights giants alike share one incentive: capture first-mover advantage before systemic safeguards mature. That race-to-scale dynamic, more than any single breakthrough, threatens orderly adaptation.
The conversation has to shift. It’s not about predicting the arrival of AGI or building some mythical "control panel" for it. The ethical, economic, and legal dilemmas posed by these systems are already on our doorstep. Agentic AI systems are reshaping all aspects of work and society in general. Agentic systems already adjudicate credit, triage patients, and even police content, decisions with immediate human consequences.
Yet those decisions are often shrouded in opacity, devoid of recourse, and flimsy oversight. All these surface problems that are becoming crisis points, and overlooking them will be at our peril.
I am all for acceleration, but a lack of coherence from the major labs as regards what the future may look like and imagining AGI governance as something only relevant "after the revolution" is of major concern.
The future isn’t some far-off destination we’re racing toward. It’s already here, arriving in a steady stream of daily updates. OpenAI rolled out multiple model iterations in just 18 months, setting the tempo for this reality. Each new API, model, or application is an unfolding of the Singularity right now, piece by piece, and we’re all living through it, adapting as we go.
The pressing question is: Do we act with the urgency this slow-motion singularity demands, or do we drift further, until steering becomes impossible?