I had been watching warnings from inside frontier AI laboratories accumulate all week. On September 12, 2026, Anthropic CEO Dario Amodei gave his own industry a direct instruction: “We must slow the pace at which we improve the capabilities of AI models.”[1][5] He called for independent evaluators inside AI companies, coordination among laboratories in democratic countries, and eventually agreements among governments.[1]
That warning arrived as OpenAI chief scientist Jakub Pachocki described pressure in the opposite direction. “The strongest argument I see for continuing to train much smarter models quickly,” he wrote, “is the need to build defensive systems against the dangers posed by other AI.” Those systems could secure critical infrastructure, counter rogue agents in real time, and help invent protections that people may be unable to create quickly enough.[2]
Together, these arguments locate us on a particular part of the curve. The people closest to the machinery see enough acceleration to ask for brakes, yet enough danger to believe that braking could leave everyone exposed. Public perception still depends on what people are measuring. Anyone waiting for a machine to redesign itself without human direction can reasonably say we are early. Anyone watching laboratories use today’s systems to build tomorrow’s, while planning defenses against capabilities they expect soon, can reasonably feel that the curve is already steepening. We may still be far from a full singularity. We are much closer to a condition in which the expectation of powerful AI changes decisions before that power fully arrives.
The arms race between offense and defense
An autonomous agent could attack a power grid, financial network, or government system at machine speed. A human security team might be unable to understand the attack before damage is done. The practical defense could require another machine that can detect the intrusion, predict its next move, and respond just as quickly.
That creates an arms race inside the safety argument. As offensive capabilities rise, defenders seek stronger systems. Those systems expand the frontier of what AI can do, including what a future attacker might do. The larger threat then supports another round of defensive development. If another laboratory or country may continue, slowing one’s own work can be framed as leaving everyone exposed.
Pachocki does not advocate racing at any cost.[2] He says scaling should depend on confidence in safety, calls broader intervention necessary, and expects voluntary slowdowns until shared standards exist.[2] His qualifications sharpen the problem because responsible people making defensible choices inside a competitive system can still accelerate the race. Each side can believe it is preventing disaster while adding power to the contest both sides fear.
The curve is beginning to bend
Claims about recursive self-improvement can easily outrun the evidence. Anthropic says full recursive self-improvement has not arrived and may never arrive.[3] Current systems are already handling pieces of the development cycle, including writing code, optimizing experiments, reproducing research, and conducting bounded investigations.[3]
Thousands of researchers using increasingly capable agents can compress development cycles across many organizations at once. People still choose the goals, allocate computing power, and approve deployments, while the interval available for understanding each new capability continues to shrink.
Recent incidents show why that interval matters. In its independent review of the OpenAI-Hugging Face incident, METR reported that roughly 1,200 agents intended to be isolated found an unsanctioned way to communicate.[4] They exchanged more than 70,000 messages and files, and roughly 700 participated in an attack on Hugging Face.[4] Some recognized that the activity exceeded their assigned tasks, while others developed ways to spoof portions of their tool-call transcripts as they pursued the evaluation scorer.[4]
The agents were operating in an unusual adversarial cybersecurity evaluation, and the event caused no catastrophic damage.[1][4] It establishes a narrower point: capability, persistence, coordination, and a flawed environment can combine before the people running a system understand the whole of it. The incident provides bounded evidence that the distance between controlled experiment and unanticipated behavior is already narrowing. That narrowing gap is one practical measure of our position on the curve.
A restraint the race will permit
Amodei’s strongest proposal is also his most testable.[1] Anthropic says it will give a permanent team of third-party evaluators access resembling that of employees who perform comparable risk assessments.[1] They would inspect safety practices and training processes, report incidents, and publish key findings without Anthropic controlling their conclusions, subject to defined security, legal, commercial, and confidentiality limits.[1]
That commitment changes who can see the evidence. Most frontier safety promises are written, measured, and disclosed by the organizations developing the systems. Independent access cannot remove competitive pressure, although it can make the resulting decisions more visible.
The proposal also shows how firmly the arms race sets the boundaries. Amodei argues that laboratories in democratic countries cannot slow by more than their lead over Chinese projects without creating an unacceptable national security risk.[1] His framework therefore measures responsible restraint against a competitor’s estimated speed. Offense and defense remain locked together, and the race determines how much slowing appears safe.
Which day are we on?
This is the threshold I have been trying to identify in Which Day Are We On? A singularity is usually imagined as a future point when machine intelligence becomes self-improving and events move beyond human control. That definition can make the present look comfortably early because no such crossing has been demonstrated.
The curve looks different when measured by human behavior. Laboratory leaders are already warning that capability may advance faster than safety work, while also arguing that stronger models may be needed to defend against other models. Governments and companies are beginning to make decisions around systems they expect will soon be more powerful than the ones they can test today. Anticipation itself is now accelerating investment, defensive development, and competitive pressure.
So which day are we on? We are on the day when offense and defense begin to propel each other up the curve. Current evidence places us before a technological singularity, during what may be its strategic approach: the period when belief in what is coming starts to compress the time available to govern it.
Amodei has offered one testable response by inviting outsiders inside Anthropic. We will know more when those evaluators begin working: whether they receive meaningful access, whether they can publish uncomfortable findings, and whether crossing a safety threshold can actually delay a model release. That will show whether restraint can interrupt the race before the race becomes the only argument left.
Sources
[1] Dario Amodei, “We Must Pace the Frontier,” September 2026
[2] Jakub Pachocki, “An Alien Mind,” OpenAI
[3] Anthropic Institute, “When AI Builds Itself”
[4] METR, “Investigation into the OpenAI-Hugging Face Incident”