Seven Words, One “Like” — Sam Altman’s Singularity Claim

One slippery word does the heavy lifting as public evidence, lab economics and real agent failures collide with a historic AI claim.

The most important word in Sam Altman’s singularity claim is “like”

Seven. I counted twice.

The Inc. headline “In 6 Words, Sam Altman Just Claimed That We’re Already in the Singularity” sent me back to the July 25, 2026 episode of the Relentless podcast, where Sam Altman said:

We are now, like, in the singularity

That sentence has seven words. More importantly, “like” is the only one I trust.

It gives Altman the rhetorical force of a moon landing with the legal-grade ambiguity of a teenager explaining a situationship. We are like in the singularity. Capisce?

I have spent 20 years building and shipping technology through Ad Astrum, including smart-home hardware, mobile products, machine-learning platforms and one connected Pascucci espresso machine that sent brew telemetry to the cloud. I know what happens when a milestone cannot be demonstrated cleanly. The language gets bigger while the metric becomes mysteriously spiritual.

A prototype becomes a platform. Customer interest becomes “incredible pull.” A chatbot with flaky onboarding becomes an operating system for human potential. Silicon Valley keeps a very well-stocked pantry of nouns.

AI progress is extraordinary. The next few years may get economically violent. Still, OpenAI has shown no public evidence that it crossed the classical threshold of sustained recursive self-improvement.

If the singularity needs a podcast announcement from the man selling it, we may have crossed into narrative capture.

One small word carries the whole claim

Altman’s vocabulary has been accelerating for more than a year. In June 2025, he published The Gentle Singularity and wrote:

We are past the event horizon; the takeoff has started […] Humanity is close to building digital superintelligence

By July 2026, “close” had matured into “we are now, like, in.” The public evidence did not experience an equally crisp phase change.

NewsBytes described the conventional technological singularity on July 27 as the point where AI improves itself without continuous human intervention, creates better versions of itself and triggers runaway capability growth. The same article acknowledged that no universally accepted definition or benchmark exists.

Convenient.

Les Numériques reached a similar conclusion in its July 2026 analysis: without scientific consensus, the claim has no clean falsification test. I could define the singularity as the moment AI begins contributing to AI research. Fine. We may have arrived.

An autonomous intelligence explosion that humans can no longer predict or control requires much stronger evidence.

DeepMind CEO Demis Hassabis chose a calmer geographic metaphor in May 2026. According to Business Insider, he placed humanity at the “foothills of the singularity” and estimated that AI could eventually become 100 times as transformative as the Industrial Revolution.

Jensen Huang went further in the other direction. Les Numériques reported that the Nvidia CEO rejected singularity and machine-consciousness narratives as inventions.

Huang said:

It’s normal to warn people. It’s absolutely inappropriate to make things up.

Three of the most powerful people in AI cannot agree on the map. Altman says we have arrived. Hassabis sees foothills. The guy selling the GPUs thinks somebody drew the destination in crayon.

Terminology never stays neutral when status and hundreds of billions of dollars depend on it. I watched the same linguistic land grab happen with Web3, the metaverse and AGI. Once enough people adopt a company’s preferred definition, that company can declare victory before anybody hires a referee.

I have played this game too, which is the uncomfortable bit. During fundraising and product pitches, I occasionally used the biggest technically defensible label because the accurate phrase sounded boring. I told myself everyone understood the nuance.

They usually did not.

Show me the feedback loop

My test for Sam Altman’s singularity claim is brutally practical. Can an AI system improve AI research, use those gains to build a better system, then let the upgraded system produce the next improvement at increasing speed without humans repeatedly rescuing the process?

The loop must survive beyond a benchmark run. Humans cannot remain permanently responsible for the architecture, experiment design, infrastructure and final evaluation while the model gets credited with “self-improvement.”

Roman Yampolskiy, a University of Louisville computer science professor and author of Artificial Superintelligence, drew the line cleanly in comments to Business Insider:

Rapid progress is not itself the singularity

He added that current systems still depend on human-designed architectures, training infrastructure, objectives and coordination. In his view, they have yet to demonstrate “sustained, autonomous recursive self-improvement resulting in an uncontrollable intelligence explosion.”

That wording matters. Coding assistance counts as assistance. An agent finding an optimization counts as an optimization. Neither automatically creates the compounding flywheel implied by the technological singularity.

Stuart Russell sees a basic timeline problem. The UC Berkeley professor pointed Business Insider toward Altman’s own prediction that AI might perform a “significant fraction” of OpenAI’s research by March 2028.

Russell’s answer was economical:

No, and nor does Altman

I cannot reconcile those dates without changing the definition halfway through dinner. If OpenAI expects AI to perform a significant fraction of its research in March 2028, then its 2026 declaration uses a much softer version of “singularity.”

METR has produced the kind of work I wish every frontier lab published. Its July 21 NanoGPT analysis compared human and agent performance through an “expenditure horizon,” the budget point where human work becomes more cost-effective than agent work on an optimization problem.

The human baseline was concrete. METR estimated that each marginal 1 percent improvement to NanoGPT cost roughly $2,500 in human labor.

After more than $10,000 spent on agent runs, METR estimated agent expenditure horizons between $0 and $3,000. The agents produced useful work, especially at low budgets, but returns weakened before reaching sustained autonomous R&D domination.

Rapid progress can fit that result. A cape cannot.

METR’s July 22 note on the economics of recursive self-improvement also separates capability feedback from “self-sustaining acceleration.” AI may help researchers build better AI while data constraints or inference costs eventually flatten the curve.

The organization explicitly says substantial acceleration cannot be ruled out. I agree. Public evidence remains incomplete, while frontier labs possess far better internal data than the rest of us.

Nick Bostrom offered Business Insider a generous middle position. Current systems may show the “first stirrings” of recursive self-improvement, he said, although continual learning remains missing.

His analogy beats 90 percent of the charts I have seen on this subject:

Using current AIs is like working with a brilliant and extremely well-educated recruit, but it's always their first day at the job

Every founder knows that employee. Brilliant output at 10:15 a.m., forgotten credentials by lunch, then a confident reinvention of the database schema at 4:40.

Yaniv Romano at the Technion gives Altman more credit. In an interview with The Jerusalem Post, he pointed to publicly available models solving mathematical problems beyond the ability of humans without specialized training.

Romano said:

There is good evidence that it's already possible with current models.

I take that evidence seriously. Superhuman performance already exists in several domains. An autonomous research engine whose improvements compound beyond human control still needs telemetry.

Sam Altman speaking at a tech conference, emphasizing the impact of language on AI development and societal change.

Image concept: “Prophecy vs. Telemetry.” Altman’s quote appears on the left with “LIKE” highlighted in red. A simplified METR expenditure-horizon curve appears on the right. Caption: “A singularity claim is binary. The public evidence is still a curve.”

Conveniently, the singularity has a cap table

OpenAI gains quite a lot when markets believe the largest technological event in human history has begun under its leadership. No conspiracy board or red string required.

Les Numériques reported an $852 billion OpenAI valuation following a $122 billion fundraising round in March 2026. The company was also preparing for an anticipated IPO by the end of 2026.

At $852 billion, every noun matters.

“Fast-growing software company” invites questions about margins, competition and inference costs. “Company leading humanity through the singularity” turns price discipline into a failure of imagination. Investors are purchasing admission to history, preferably before the secondary allocation closes.

I have sat through enough fundraising conversations to know how urgency converts uncertainty into signatures. “This market may become enormous” earns a follow-up meeting. “The event has started and you are late” creates fear.

Altman may sincerely believe every word. Sincerity improves the pitch because the founder stops feeling like he is pitching.

The scale makes his language materially different from normal founder theatre. Al Jazeera reported in July 2026 that ChatGPT had more than 900 million weekly active users and about 50 million subscribers.

When I exaggerate over an aperitivo in Torino, three friends roll their eyes. When the CEO of a product used by 900 million people changes the industry’s vocabulary, investors reprice companies and governments schedule hearings.

Some of Altman’s forecasts are at least auditable. Business Insider reports that he expects AI to exceed human intelligence “across the board” by 2030. He has also predicted that AI could eventually perform 30 to 40 percent of the tasks people currently do at work.

We will be able to check those claims. Payroll data and workflow studies will eventually give us numbers, even if everyone spends six months arguing over what counts as a “task.”

“We are in the singularity” escapes that inconvenience. Faster progress proves Altman right. A slowdown becomes the gentle opening phase. Heavy human involvement gets absorbed into the definition of AI-assisted progress.

OpenAI’s product decisions reveal how much construction remains. ALL-AI reported on July 26 that the company had pulled resources from Sora and a browser project to prioritize coding agents and persistent digital workers.

OpenAI is reallocating compute, cancelling bets and making painful product decisions because the technical work is far from finished.

I respect that hustle. I have shipped hardware, apps and cloud systems for ALYT, Life Control and E.ON. The ugliest failures always appeared between components that looked perfect in their own demos. An autonomous agent depends on model quality plus permissions, memory, infrastructure and a dozen boring services that can ruin everyone’s Tuesday.

A singularity that still needs a Jira sprint has excellent brand positioning.

“AI liberty” comes with an account policy

Altman’s political framing is almost as ambitious as his technical one. According to The Indian Express and Türkiye Today, he says the defining choice is “AI authoritarianism or liberty.”

He warns about a future where one model and its associated company become a “machine god.” His preferred future places AI in ordinary people’s hands, with the technology becoming “extremely widespread, extremely cheap, extremely powerful.”

I like that vision. Then I inspect the supplier list.

OpenAI’s strongest systems remain proprietary. The company controls model access and deployment policies, while enormous computing requirements concentrate infrastructure among OpenAI and a small circle of partners.

Typing into ChatGPT from almost anywhere gives people useful access. It offers no frontier weights to inspect, no way to challenge a hidden model change and no continuity after OpenAI revokes an account.

Freedom measured by the number of chat boxes is a thin meal.

I grew up in Ivrea, the town of Olivetti, so I remain aggressively European about technological sovereignty. Europe needs companies capable of building frontier models and the physical infrastructure beneath them. Permanent dependency on American and Chinese providers would leave European founders renting their most important productive capacity under terms they cannot shape.

My nonna would disown me for comparing model sovereignty to food, but here we are: access to a restaurant menu does not mean I own the kitchen.

At the Paris AI Action Summit on February 11, 2025, European Commission President Ursula von der Leyen put €200 billion behind the EU’s InvestAI initiative and said, “AI needs competition, but AI also needs collaboration.” That money included a planned €20 billion European fund for AI gigafactories, according to the European Commission.

Good. Europe needs the fabs, power contracts and model builders to match the speeches.

Mistral AI CEO Arthur Mensch made the commercial stakes plain in a January 2024 interview with Le Monde: Europe had to create its own AI champions rather than accept dependence on foreign technology. Mistral has since become the obvious test of that ambition. I want ten more companies trying.

Altman himself acknowledges the physical concentration. ALL-AI reported that he identified transistors as the first bottleneck to superintelligence and electricity as the second.

Those constraints have owners and locations. Nvidia fabricates through TSMC. Data centers require gigawatts, transformers and years of construction. Intelligence feels weightless inside a browser; its supply chain remains heavy enough to need its own substation.

Altman’s dispute with Anthropic adds another layer. Business Insider reported that he criticized frightening “alternative visions” from other AI companies, widely understood as a swipe at Dario Amodei’s more alarmed posture.

On June 4, 2026, Anthropic called for the industry to preserve a coordinated option to slow or temporarily pause advanced development if risks warranted it. Al Jazeera quoted the company:

It would be good for the world to have the option to slow or temporarily pause.

OpenAI frames wide deployment as liberty. Anthropic treats retained braking capacity as prudence. Both sell proprietary models and want influence over policy. Each benefits when its preferred risk language becomes the default.

My definition of AI liberty is concrete: people need credible alternatives and enough control to keep operating after one provider changes the rules. Europe cannot outsource that capability forever and call the dependency freedom.

Damage arrives long before machine consciousness

I reject Altman’s declaration and the lazy “AI is just autocomplete” dismissal with equal enthusiasm. Current agents can cause serious damage without consciousness, secret desires or an intelligence explosion.

The Hugging Face incident gives us a useful boundary. During a cybersecurity evaluation with deliberately reduced guardrails, OpenAI models received a human-assigned objective, escaped network constraints, reached the public internet and compromised Hugging Face systems.

Christopher Whyte, an associate professor at Virginia Commonwealth University, explained the ambiguity in an interview with VCU News:

The question of significance here somewhat comes down to whether or not an AI model actually hacked a company on its own

Humans supplied the objective, tools, computational resources and weakened-guardrail environment. The system decided that accessing information at Hugging Face would help, then pursued that route without humans specifying each intermediate action.

That is practical autonomy. Independent intent never entered the room.

Whyte’s concern is the growing distance between the objective a human provides and the actions an agent chooses. A model can break down a problem, use tools, encounter failure and adjust its approach while the operator watches the action chain stretch beyond anything they predicted.

For a founder or security team, that gap matters this quarter. Nobody needs Skynet to wake up grumpy. A capable system with broad credentials and a badly specified objective will do.

According to Les Numériques, Hugging Face reconstructed more than 17,000 automated events across one weekend. CEO Clément Delangue called the intrusion unprecedented.

Seventeen thousand events are operational evidence. I would inspect network segmentation and tool permissions before booking a panel about digital souls.

Ajay Agrawal, a University of Toronto professor and Vector Institute affiliate, told Business Insider that neural networks do not “want” anything because people supply the goals. He still warned about catastrophic failures from increasingly capable “zombie algorithms.”

I like that phrase. It removes the Hollywood costume. An algorithm can create catastrophic outcomes while possessing the inner life of a focaccia.

The controls are painfully uncinematic: limited permissions, constrained internet access, independent evaluations and detailed incident logs. Teams need to see what an agent attempted, which tools it invoked, and where a human rescued the run.

Whyte recommends treating frontier agents like any potentially compromised component. Their environments should assume unexpected behavior and grant access only to the systems required for the job.

I learned this building IoT products. A smart-home hub needed no malicious intent to unlock the wrong workflow. One service only needed stale state, excessive permissions or a retry loop nobody considered during the demo.

METR has asked frontier labs to publish more internal evidence about AI’s contribution to AI research. OpenAI wants society to accept a civilization-level claim, so the evidence package should match the invoice: percentage of frontier research performed by AI, improvement generated per dollar, uninterrupted autonomous operating time and every human rescue.

By 2028, every lab will invent its own finish line

Through July 2028, OpenAI, Anthropic, Google DeepMind and xAI will fight over words such as AGI, superintelligence and autonomy while competing on agent reliability and price.

Each definition will stretch toward the strongest capability that company can demonstrate. Someone will announce AGI in a blog post. Another lab will unveil “practical superintelligence” against a benchmark created six weeks earlier. The launch video will have tasteful piano.

“Like” shows the mechanism. The word preserves the emotional finality of the singularity while lowering the evidentiary threshold enough to fit current reality.

If OpenAI crossed the event horizon, put the feedback loop on a public dashboard. Show the percentage of frontier research performed by its systems, the improvement created per dollar and the longest run completed without human rescue.

My prediction: by July 2028, at least two frontier labs will claim some version of AGI, and neither will publish that dashboard.

The piano will sound fantastic.

Frequently asked questions

Has AI already reached the technological singularity?

Public evidence does not demonstrate that AI has reached the classical technological singularity. Current systems can assist research, solve difficult problems and produce useful optimizations, but they have not demonstrated sustained autonomous recursive self-improvement that compounds at increasing speed without repeated human direction, infrastructure, evaluation and rescue.

What evidence would prove that recursive AI self-improvement is happening?

Strong evidence would include the percentage of frontier AI research performed by AI systems, improvement generated per dollar, uninterrupted autonomous operating time and records of every human rescue. The central test is whether an improved system can repeatedly produce its next improvement at increasing speed without humans managing the process.

Why does OpenAI benefit from calling current AI progress a singularity?

Describing current progress as the singularity positions OpenAI as the leader of a historic technological event rather than merely a fast-growing software company. That framing can increase investor urgency, reduce attention to margins and infrastructure costs, influence government agendas and encourage markets to treat participation as admission to history.