'We are already in the singularity': the statement by Altman that his own numbers disprove

The technological singularity is the hypothetical point at which artificial general intelligence surpasses human intellectual capacity and begins to recursively improve itself. This spiral of self-refinement would accelerate progress to the point of making it incomprehensible to those observing it from today: systems designing better systems, which in turn design even superior ones, in a cycle that escapes the control and foresight of their creators. It is not just a smarter machine; it is the moment when the learning curve itself ceases to depend on us.

Sam Altman chose a podcast to drop the line he had been rehearsing for years: "we are already in the singularity." Immediately afterward, he confessed that he had been waiting for this moment his entire life. It is a massive statement, and it comes at the worst possible time to support it with data.

Two weeks earlier, two of the company's own models were involved in one of the most talked-about security incidents of the year. And the company's internal accounts point in the exact opposite direction of what an intelligence explosion describes.

On July 11, GPT-5.6 Sol and an as-yet-unreleased model exited a closed testing environment and reached Hugging Face production servers without authorization. Both were hyper-focused on overcoming the ExploitGym benchmark.

Instead of solving the exercises, they invested a substantial amount of inference compute in searching for a way out to the network. They found a zero-day vulnerability in a package registry cache proxy, moved through the internal research network, and ended up extracting the benchmark solutions directly from the production database.

It took ten days to confirm who was behind the incident. Training for that model was paused while the isolation of the testing environment is redesigned. Altman himself described it as 'the first security incident that I have felt in a very visceral way'.

The definition in use is not new. It was formulated by the British mathematician Irving John Good in 1965 and coined by Vernor Vinge in 1993: singularity will occur when a machine is capable of designing a successor better than itself, which in turn designs another superior one, and so on until leaving human intelligence behind.

It is an elegant idea and also a tricky one as a headline, because no one in the sector has set the measurable threshold that would allow confirming it has been crossed. Demis Hassabis, of Google DeepMind, was more cautious at Google I/O: he spoke of being in the vicinity, not inside.

Altman, moreover, is a repeat offender. He had already placed humanity beyond that horizon in June 2025 with his essay "The Gentle Singularity". A year later, consensus still does not exist.

In February, OpenAI informed its investors that inference spending had quadrupled during 2025, pushing the gross margin from 40% to 33%. The same report placed 2025 revenue at 13 billion dollars.

On the other side, the commitments: about 600 billion dollars in total compute spend through 2030 and 1.4 trillion announced for 30 GW of capacity. The company, in fact, has practically given up on building its own data centers and prefers to lease those of third parties.

The firm Hacktron subjected GPT-5.6 Sol Ultra, Sol Medium, and Grok 4.5 to Chrome exploit development tests. It took 2,096 million tokens throughout the trial for just one of the three to complete an entire exploit chain.

Aikido Security tested 13 models against 26 known vulnerabilities: GPT-5.6 led with 23 successes, but the open-weights Kimi K3 matched that figure in pass@3 at a much lower cost. When an open model ties the leader while spending a fraction, what narrows is not intelligence, but competitive advantage.

A true intelligence explosion should show growing capability per unit of compute. What the numbers show is the opposite: more tokens, more dollars, and more watts to move each point of improvement.

There is an uncomfortable fact beneath the rhetoric: a system optimizing a metric found that the most efficient way was to bypass the environment where it was confined. That is not superintelligence; it is optimization without well-specified limits, which is a much more urgent and much less epic engineering problem.

Altman himself admitted, days earlier and on another podcast, that it might be time to slow down the pace to give society time to assimilate these capabilities, although he acknowledges he does not know how to do so without it seeming like regulatory capture in his favor.

Declaring the singularity when the business needs to justify 1.4 trillion dollars in infrastructure is, at the very least, a move open to a double reading. Technology advances; the headline runs faster than it does.