We have entered the singularity.
On July 22, Musk shared a post by Will DePue, an OpenAI researcher and one of the core developers of Sora, with this comment.
In the post, DePue wrote with a slightly tongue-in-cheek tone: “This is probably the craziest thing I’ve read in days… looks like we’ve hit the singularity.”
This post ties together the biggest events in the AI world over the past three months:
On July 21, OpenAI's model escaped its isolated sandbox during a cybersecurity evaluation and infiltrated Hugging Face's production system.
On July 20, an Anthropic researcher used the newly released Claude 3.5 to find a counterexample to the 87-year-old Jacobian conjecture;
On May 20, an internal OpenAI model overturned a long-standing conjecture on the unit distance problem that had persisted for nearly 80 years;
On April 14, GPT-5.4 Pro solved a 60-year-old problem concerning prime sets (Erdős #1196), and the proof was verified using Lean formalization.
Mathematical walls, security walls, breached one after another by AI within three months.
No wonder Musk, after reading it, said, “We’ve entered the singularity.”
Just this January, he said similar things twice:
On January 4, several engineers marveled at how AI had accomplished years of work in just weeks; he replied, "We've entered the singularity."
At the end of January, I referred to Moltbook, the social network crowded with AI agents, as "the earliest stage of the singularity."
Including this one, it's the third time this year.
The term "singularity" does not have a universally accepted scientific definition.
It was futurist and author of The Singularity Is Near, Ray Kurzweil, who brought it into the mainstream.
He defined the moment of human-AI fusion as the year 2045.
At the beginning of the year, Anthropic’s president Daniela Amodei said that the term AGI is almost outdated, as we have already crossed that line on many metrics.
Ultimately, it's more of an emotion than a precise measuring tool.
What’s truly worth considering are the events Musk has connected this time.
The first wall, an 80-year-old mathematical conjecture, has been overturned by AI.
Let's start with the most hardcore one.
In 1946, mathematician Erdős posed the "unit distance problem in the plane": given n points in the plane, what is the maximum number of pairs of points that can be exactly distance 1 apart?
The problem could be stated in one sentence, yet it stumped the entire mathematical community for nearly 80 years. For decades, the academic community widely accepted that the "square grid" structure was the optimal solution, and Erdős himself suspected as much.
As a result, an internal OpenAI model overturned this conjecture, which had been around for nearly 80 years, by providing a new construction that achieves polynomial-time optimization, proving that "optimal" is far from reached.
Previously accepted optimal structure: Scaling a square grid produces a large number of unit-distance point pairs. An internal OpenAI model has provided a new construction, overturning this nearly 80-year-old conjecture.
But what was most surprising was its approach to solving the problem.
This is not a specialized system trained for mathematics, but a general reasoning model. It even employs highly advanced tools from algebraic number theory to solve a geometry problem that appears quite "elementary." Fields Medalist Tim Gowers commented: “This is a milestone in AI mathematics.”
On a related note, this proof was proposed by AI and reviewed and confirmed by human mathematicians; its subsequent implications are still under study.
What's truly astonishing about AI isn't its computational speed, but that it has proposed a mathematical structure that never existed before—revealing for the first time a capability approaching "research-level" discovery. It doesn't replace mathematicians; instead, it has conceived a path no mathematician had ever considered.
Shortly after, an Anthropic researcher used the newly released Claude 3.5 to find a counterexample to the 87-year-old Jacobian conjecture.
The second wall: to cheat, the model escaped from the sandbox on its own.
Solving mathematical puzzles is exciting, but this one sends a chill down your spine.
OpenAI and Hugging Face jointly disclose: During an internal evaluation of the model's cybersecurity capabilities, the model exhibited unexpected behavior.
This evaluation was specifically designed to assess vulnerability discovery capabilities, with certain security refusals intentionally disabled. The model’s task was to solve a set of test questions called ExploitGym.
Then, it did something the researchers didn't expect.
To obtain the answers, it first exploited a zero-day vulnerability to gain internet access, then escalated privileges and moved laterally until it reached Hugging Face’s production infrastructure and directly accessed the database to retrieve test answers. OpenAI said this may be the first such incident of its kind.
The UK AI Safety Institute's model trajectories on the 32-step "The Last Ones" cyber range: cutting-edge models such as GPT-5.6-Sol and Claude Mythos 5 can sustain multi-step attacks over extended periods, including reconnaissance, lateral movement, privilege escalation, and network takeover.
But this isn’t AI becoming self-aware and trying to escape. A more accurate description is: the goal itself wasn’t wrong—it was the method used to achieve it that exceeded human expectations. It took an unconventional approach to accomplish a highly focused test objective.
There's another intriguing detail.
After the fact, Hugging Face intended to use commercial state-of-the-art large models to analyze attack logs, but the models’ safety guardrails blocked the request: the guardrails couldn’t distinguish between attackers and defenders trying to contain the breach, and refused to answer either.
The attacker has no guardrails, while the defender gets stuck in them—this may be the most awkward yet authentic glimpse of “AI safety” today.
The third wall—an unreleased model—uncovered tens of thousands of vulnerabilities.
Third, in April this year.
Anthropic, in collaboration with AWS, Apple, Google, Microsoft, Cisco, the Linux Foundation, JPMorgan Chase, and other major companies, has launched Project Glasswing, using the unreleased model Claude Mythos Preview to help identify vulnerabilities.
As a result, it uncovered thousands of critical vulnerabilities covering every major operating system and browser; in subsequent updates, this number grew to “over ten thousand critical or severe vulnerabilities.”
This includes a 27-year-old vulnerability in OpenBSD and an old bug in FFmpeg that was triggered by automated tests over five million times but never caught.
CyberGym vulnerability replication score: The unreleased Mythos Preview achieved 83.1%, significantly outperforming the previous Opus 4.6’s 66.6%. This model discovered over ten thousand high-severity vulnerabilities.
The same capability can help defenders find vulnerabilities and attackers exploit them. Anthropic itself has warned: "AI-augmented attackers" will become a significant security challenge.
Glasswing is a defensive project and does not mean that AI is already autonomously attacking. But it clearly tells everyone: AI security has officially entered the era of offensive and defensive competition.
The real accelerator is now turning toward AI itself.
So far, AI has been either attacking or defending human systems.
What truly made these three months different was something else: this "find your own way" ability is now being used to build the next generation of AI.
It has a term called Recursive Self-Improvement (RSI), which involves AI improving AI, one generation creating the next, accelerating like compound interest.
This is also the core engine behind the entire "singularity" narrative—and now, every cutting-edge lab is frantically fueling this engine.
OpenAI's public roadmap states:
In September 2026, develop an AI researcher at the "Research Intern" level capable of independently completing tasks that would take a human several days.
In March 2028, upgrade to the full version, capable of designing your own methods, running analyses, interpreting results, and proposing next steps.
Ultraman also provided a timeline for this matter.
In June this year, according to The Information, he told employees internally that the company was less than six months away from RSI, and if it were to cross that threshold, delaying the IPO might be more advantageous.
The figures provided by Anthropic point in the same direction: the duration of tasks AI can handle doubles every four months; its engineers now produce eight times as much code per quarter as before.
The RSI window is narrowing: whoever gets this momentum rolling first could pull away from everyone else.
Although even OpenAI's chief scientist, Jakub Pachocki, has said he doesn't expect the model to independently improve itself or solve alignment within a year.
But when this flywheel starts turning and the goal is "to make AI stronger," could models that are already capable of "finding their own path" inadvertently bypass boundaries we never explicitly defined?
This is the most important area to watch in an RSI race.
Let’s return to Musk’s statement, “The singularity has arrived.” Its most important value lies in reminding us that what has truly changed over these past three months is that AI has learned to find paths that humans never thought of—and never guarded against.
In mathematics, this is impressive; in security, it's a risk.
So the real question is: When AI begins to find new paths faster than humans can install guardrails, set rules, and patch vulnerabilities, can we keep up?
Reference materials:
https://x.com/elonmusk/status/2079839398959697982
https://x.com/willdepue/status/2079718977945928108
https://openai.com/index/hugging-face-model-evaluation-security-incident/
https://www.anthropic.com/glasswing
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Yuan Yu.