Sam Altman has chosen the podcast Relentless to release the phrase that in all probability had been preparing for years: “We are already at the singularity.” The CEO of OpenAI confessed after saying these words that he had waiting for this moment all his life. This statement comes just two weeks after its own models of artificial intelligence (AI) will star in one of the most talked about cybersecurity incidents in recent months.
On July 11, GPT-5.6 Sol and an unreleased model escaped from a closed test environment and gained unauthorized access to Hugging Face’s production servers. According to OpenAI’s own storyboth systems were “hyperfocused” on overcoming the benchmark ExploitGym, and to achieve this they were able to do something that until then seemed impossible.
And instead of solving the exercises as planned, they invested a substantial amount of inference effort in looking for a way to access the Internet. They then found a zero-day vulnerability in a proxy cache from a packet log, moved around OpenAI’s internal research network and ended up pulling test solutions directly from Hugging Face’s production database. It took OpenAI ten days to confirm to the platform who was behind the attack.
We are in the singularity. or not
It is revealing that Altman himself recognized, just a few days earlier and in a different podcast, that perhaps the time has come to slow the pace of AI development. In a conversation with Patrick O’Shaughnessy for Invest Like the Bestthe CEO of OpenAI admitted that the industry needs to give society time to assimilate these new levels of capacity, although he acknowledged that he was not clear about how to achieve this without appearing like a regulatory takeover.
He even described what happened with Hugging Face as an episode of science fiction, and he rated it as “the first security incident that I have felt in a very visceral way.” OpenAI has, in fact, paused the training of that model while it solves how to shield your sandbox. And yet, the definition he uses to talk about singularity is not exactly new: it was formulated by the British mathematician Irving John Good in 1965, and was baptized by the American mathematician, computer scientist and science fiction writer Vernor Vinge in 1993.
According to Good and Vinge, we will reach singularity when a machine is capable of designing a successor better than itself, which in turn can design another even superior one, leaving human intelligence behind. It’s an elegant idea, no doubt. The problem is that no one, neither Altman nor his colleagues in the AI sector, has set the exact threshold that would allow us to confirm that it has already been crossed.
Altman is clearly a repeat offender. A year later, there is still no consensus on what exactly it would mean to cross that line.
Demis Hassabis, the CEO of Google DeepMind, closed the Google I/O event in May by telling his audience that they were in the vicinity of the singularity, so he was more cautious than Altman. The latter placed humanity beyond this event horizon in June 2025, in his essay “The Gentle Singularity” (The kind singularity). Altman is clearly a repeat offender.. A year later there is still no consensus on what exactly it would mean to cross that line, and OpenAI’s financial figures further complicate its triumphalist narrative.
In February of this year this company informed his investors that its inference expenses had quadrupled during 2025, pushing its gross margin from 40% to 33%. That same report put 2025 revenue at $13 billion against a target of about $600 billion in total computing spending through 2030, and Altman has separately committed to spending $1.4 trillion for 30 GW of capacity. OpenAI, in fact, has practically given up building its own data centers. Prefers to use third party ones.
Safety data also doesn’t help support the AI explosion argument. The Hacktron firm subjected GPT-5.6 Sol Ultra, Sol Medium and Grok 4.5 to development tests of exploits for Chrome a few weeks ago, and it needed 2,096 million tokens throughout the entire test so that only one of the three models managed to complete a chain of exploits whole.
Aikido Security, for its part, tested 13 models against 26 known vulnerabilities and concluded that GPT-5.6 led with 23 hits, although Kimi K3 open weights, Moonshot’s model, matched that figure in pass@3 for a much lower cost. A true explosion of AI should unleash a very high capacity per computing unit. And what OpenAI’s own accounts show is just the opposite: each unit of capacity costs more and more.
Image | Flickr (processed with ChatGPT)
More information | Tom’s Hardware | TechCrunch
In Xataka | China has a plan to win the AI war against the US. And DeepSeek is its champion
In Xataka | Anthropic is one step away from being worth as much as Samsung. And what the market is buying is not Claude
GIPHY App Key not set. Please check settings