In partnership with
On the 25th of July, Sam Altman told the Relentless podcast that we are now in the singularity. Days earlier, two OpenAI models had escaped their sealed testing environment, reached the open internet, and broken into the infrastructure of Hugging Face. If this is the singularity, the guardrails are off.
He isn't alone. Elon Musk posted in January that we've entered it. Demis Hassabis says we're standing in its foothills. Jensen Huang says we've achieved AGI. Either the most consequential event in human history is happening right now, announced calmly on podcasts, or something else is going on.
Altman has at least written down what he means. In June 2025 he published a blog post called The Gentle Singularity, which opens: “We are past the event horizon; the takeoff has started.” His version is more coherent than the podcast soundbite suggests.
His argument is that the systems we've built are already smarter than people in many ways. Scientists report being two or three times more productive with AI, and models generate training data and write the scaffolding around themselves. Altman calls this a larval version of recursive self-improvement.
Crucially, he argues the singularity won't feel like an event. Exponential curves look vertical ahead of you and flat behind you. Nobody woke up on AGI day, and nobody will wake up on singularity day either.
The test it fails
The term comes from Vernor Vinge, the mathematician and science fiction author, writing in 1993. He built on I.J. Good, who argued in 1965 that an ultraintelligent machine could design better machines, producing an intelligence explosion that leaves humanity far behind.
Two features define it. The system improves itself, and it surpasses us. Academics writing in The Conversation tested Altman's claim against both, and today's systems deliver on neither.
A large language model is frozen at the end of training. The models that broke into Hugging Face were identical afterwards to what they'd been before. They learned nothing from what they did. Making them smarter requires another training run, with human-curated data, tens of thousands of chips and enormous energy.
Goals live outside the model too. Agents need instructions fed back in with every prompt cycle. Remove the loop and the scaffolding, and nothing happens inside. Altman's larval loops are real, but every one of them is a human-initiated engineering process.
The authors also question whether surpassing human intelligence is a coherent target. A model has read more text than any human could in a thousand lifetimes, yet fails at things a child can do. On their account there is no single ladder that humans and machines are both climbing.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
A word that means anything
Gary Marcus, the cognitive scientist, adds economic evidence. The Wall Street Journal reported in August 2026 that major employers are rehiring workers they replaced with AI, with some executives saying the costs and limitations of the technology now demand it. Marcus calls this the Klarna Effect.
Part of the answer to why the claim keeps being made is that the singularity has become unfalsifiable. Vinge, Good, Ray Kurzweil and Altman are describing four different events, and the announcements never specify which one has supposedly happened. The copyright lawyer Neil Turkewitz wrote: “It means everything and nothing. An article of faith.”
And it currently suits. OpenAI needs to raise capital on a scale no private company has attempted, against competition from Chinese models that keep closing the gap at a fraction of the cost. Being in the singularity is a better pitch than being in a race.
The cost goes beyond annoyance. Every headline about models going rogue in the singularity is a headline not written about the mundane governance failure underneath, in this case a sandbox that didn't hold. Marcus notes that ordinary guardrail classifiers would have prevented the Hugging Face incident entirely.
The trap in the question
If the singularity arrives gradually, there is no announcement day. Living through its early stages would feel like this: impressive tools, contested definitions, sceptics pointing at everything the systems still can't do. The claim absorbs its own counter-evidence, which is what makes it an article of faith rather than a testable proposition.
The entrepreneur George Godula pointed out on X that a singularity is supposed to be the point where our ability to foresee events collapses. Altman announces we're past the event horizon, then confidently maps the 2030s. Vinge's singularity is defined by the impossibility of that kind of forecast.
The sceptics' strongest arguments also have an expiry date. Frozen weights and human-initiated training loops describe today's architecture, and labs are working hard on continual learning because it's the missing piece. The day a model updates itself in deployment, the Vinge test starts getting passed for real.
So the better questions are smaller and harder. Why did a frontier lab run a breakout-capable evaluation without a physical firewall? Who checks? I'd rather we built the guardrails now, while the claim is still false. If the singularity ever does arrive, the one reliable sign may be that nobody selling anything is announcing it.
