Altman Reviews PHP Crisis Week: AI Scans Out Massive Vulnerabilities, Advocates Rebuilding the Underlying Architecture
In an interview with Stripe founder John Collison, Altman reviewed "PHP Crisis Week," saying that large models' ability to uncover vulnerabilities continues to grow exponentially. Previously no one dared use it to scan the entire internet's public code; after actually doing so, the number of critical vulnerabilities exposed exceeded everyone's expectations. He believes fixing software alone is no longer enough, and that the lasting solution is to rebuild the entire stack from router firmware and operating system kernels to upper-layer applications, and that AI tools capable of writing code in parallel have already led a group of startups to begin. Ten days after the crisis, he still dares not declare victory, and says humanity's existing threat-model assumptions are failing.
Why read this
AI score 56
In the interview, Altman reviewed the scale of vulnerabilities exposed by AI's full-code scans and offered the judgment that the underlying architecture should be rebuilt rather than continuously patched.
Topics
View original postStripe founder John Collison's just-completed 52-minute exclusive interview with OpenAI CEO Sam Altman—holy crap, the aftereffects hit hard after watching it, the information density is way too high, it's absolutely explosive, meow, Compared with those previous PR interviews full of empty platitudes, it feels like this is the first time Altman has completely torn off his guard and laid out, without reservation, OpenAI's most terrifying internal technical progress, the most real mathematical battle results, and his immense melancholy over the end of the old world, The most explosive part of the whole session's information density was that math-crunching weekend known as "Hold My Beer." Altman admitted that when the internal team took the new model to tough out all of humanity's hardest frontier scientific problems, the model conquered them one after another, For a single proof related to the Millennium Prize Problem, the Navier-Stokes equations, they poured in $10 million-level compute, equivalent to instantly running through hundreds of millions of dollars' worth of human scientific research. But his reaction right after was not revelry, but rather, after two consecutive nights of excitement-induced insomnia, a huge sadness welling up in his heart: He watched with his own eyes as the temple of humanity's pure intellectual labor was being permanently sealed away by compute in an irreversible way. Just like seeing the last batch of farmers plowing with horses on Irish soil back then, an old chapter of carbon-based brains taking on truth single-handedly has completely ended. Many people still treat Agents as a few chatbots tossing messages at each other, but Altman is redefining the bottom line of organization and safety: He even dislikes the term "Agent Swarms" and insists on calling them "Agent Teams." Because frontier users can already conduct "ten thousand parallel assaults" through chains of thought, Facing this kind of capability expanding at the speed of light, he admits that inside OpenAI there is widespread extreme pessimism about "sandboxes being able to contain superintelligence," and the only thing they can rely on is internal monitoring and alignment. He even uttered a chillingly cold safety prophecy: "I would rather have a trillion AIs of slightly superhuman intelligence checking and balancing each other as they enter society than one ultimate behemoth with an IQ of a trillion ruling alone." What stings all entrepreneurs and ordinary workers even more is his public slap in the face to the whole industry's narrative: Altman said bluntly that the AI blueprint Silicon Valley paints for the public today is a "terrible sales pitch"— The giants boast about curing cancer and making prices cheaper on one hand, while on the other depicting a utopia where ordinary people lose their jobs, hand over control, and have to scrape by on welfare and live in a surveillance society. His core reason for insisting on iterative deployment is that a tiny group of elites within a few square miles of San Francisco must never decide the future for billions of people around the world. From smashing PHP's underlying vulnerabilities and triggering a full-stack rebuild, to $10 million in compute overturning frontier mathematics; From cloud memory's ruthless mockery of local compute, to Agents completely erasing "fake work" like web forms. Listening to the whole conversation, there is no arrogant preaching, only an extremely real sense of weightlessness: The door has already been thrown open with a roar, and the physical and logical laws of reality are being rewritten line by line of code, Those human experiences and self-esteem we once thought indestructible are, before this era's most violent surge of compute, as fragile as a fleeting sunset on the horizon.
View original
Stripe 创始人 John Collison 对 OpenAI CEO Sam Altman 刚做完的这期 52 分钟专访，卧槽看完后劲好大，信息密度台大了，简直炸裂喵个咪， 相比之前那种套话连篇的公关访谈，感觉这是奥特曼第一次彻底撕掉防备，把 OpenAI 内部最恐怖的技术进展、最真实的数学战果，以及他对旧世界落幕的巨大惆怅，毫无保留地全部和盘托出了， 全场信息密度最炸裂的，是那个被称为“Hold My Beer（看我的）”的数学攻坚周末。 奥特曼坦言，当内部团队拿到新模型去硬啃全人类最难的前沿科学难题时，模型一个接一个地将其攻克， 单单一个千禧年大奖难题 Navier-Stokes 方程相关的证明，他们就砸进去了 1000 万美元级别的算力，相当于瞬间跑完了人类几亿美元的科研量。 但紧接着他的反应不是狂欢，而是连着两晚兴奋到失眠后，涌上心头的那份巨大悲伤： 他亲眼看着人类纯智力劳动的殿堂，正在被算力以一种不可逆的方式永久封存。 就像当年在爱尔兰土地上看到最后一批用马耕作的农夫一样，一个属于碳基大脑单挑真理的旧篇章，彻底终结了。 很多人还在把 Agent 当成几只互相丢消息的聊天机器人，奥特曼却在重新定义组织与安全的底牌： 他甚至讨厌“Agent Swarms（蜂群）”这种词，坚持称之为“Agent Teams（团队）”。 因为前沿用户已经可以通过思维链进行“一万次并行攻坚”， 面对这种以光速膨胀的能力，他坦白 OpenAI 内部对“沙盒能关住超级智能”普遍极其悲观，唯一能依赖的只有内部监控与对齐。 他甚至说出了一句冷峻至极的安全预言： “我宁愿有一万亿个略微超人智力的 AI 彼此制衡进入社会，也绝不要一个智商一万亿的终极巨兽独裁。” 更刺痛所有创业者和普通打工人的，是他对全行业叙事的公开打脸： 奥特曼直言，硅谷今天向公众描绘的 AI 蓝图是“糟糕透顶的推销（Terrible sales pitch）”—— 巨头们一边吹嘘能攻克癌症、物价更便宜，一边却描绘一个普通人失去工作、交出控制权、不得不靠低保和监控社会苟活的乌托邦。 他坚持迭代部署的核心理由，就是绝不能让旧金山几平方英里内的一小撮精英，替全世界几十亿人决定未来。 从砸碎 PHP 底层漏洞导致全栈重构，到 1000 万美元算力推翻前沿数学； 从云端内存对本地算力的无情嘲讽，到 Agent 彻底抹杀网页表单这种“假工作”。 整场对话听下来，没有狂妄的说教，只有一股极度真实的失重感： 大门已经轰然推开，现实的物理与逻辑法则正在被一行行代码重写， 那些我们曾经以为坚不可摧的人类经验与自尊，在这个时代最猛烈的算力狂飙面前，脆弱得就像地平线上一抹转瞬即逝的落日。
Body · AI translation
Honestly, no one ever dared to take AI and scan the entire internet's code before, but after hearing Altman recap the "PHP crisis week" at the start of the interview, I think every programmer would feel suffocated:
The critical vulnerabilities uncovered by large models far exceed human imagination. The security myth that the modern internet prides itself on has, it turns out, been running naked on its foundation all along.
Stripe founder John Collison's interview with Altman was real, simple and blunt, with sharp questions and no small talk or pleasantries. It opened by dragging the whole world back to that still-harrowing "PHP crisis week."
In an extremely calm tone, Altman said something that sends a chill down the spine of every security expert: The ability of large models to find vulnerabilities has been growing exponentially, but before this, no one had ever truly dared to take it and scan all the public code running across the entire internet.
When this actually happened, the number of fatal vulnerabilities that surfaced stunned everyone.
The vast majority of the infrastructure of modern digital civilization is, in essence, a series of dangerous buildings erected on slapdash craftsmanship.
Many low-level scripts written decades ago, and ancient protocols running on hundreds of millions of servers worldwide, originally survived on an unspoken reliance on obscurity—"the code is too complex, hackers can't find it with the naked eye";
but when large models with advanced reasoning capabilities became fully automated radar, all that self-deceiving cover was torn to shreds in an instant.
Faced with this torrent of vulnerabilities, John Collison asked the core question: Are we doomed to forever patch passively?
Altman's answer thoroughly rejected the tinkerer mindset of traditional cybersecurity, with its endless patching.
He said: fixing software alone has long been useless; the only lasting solution is to completely rebuild the entire underlying architecture.
Two years ago, if anyone had dared to say they would rewrite router firmware, operating system kernels, and the full-stack applications above them from scratch, every investor would have treated them as insane; but today, precisely because there are AI tools that can write code in parallel thousands of times over, a batch of brand-new startups is already tearing down the entire stack built years ago.
Recapping the situation ten days after this crisis, Altman said he still dares not declare victory.
Although the speed at which companies used AI to quickly write patches made the first day much better than the worst-case disaster expected;
the deepest fear in security circles has never been software-level overflows and injections.
It is that all the "threat model assumptions" humanity has established in the past are now comprehensively failing in a physical sense: when AI's comprehension penetrates down to the underlying transistors, it can even ferret out from microscopic hardware those ghost vulnerabilities that require entirely new laws of physics to explain. This is the cruelest reshuffle in the history of software engineering: the balance of attack and defense has been completely overturned.
In the past, hackers found vulnerabilities through years of persistent flashes of insight, while defenders ran themselves ragged; in the future, if AI cannot be used to thoroughly overturn and rewrite the dirty code humanity has left behind over decades, any device connected to a network cable will become a target running naked on the wasteland.
Perhaps the era of patching up the old world is truly over, and those traditional architectures once revered as sacred cornerstones are being torn down into ruins by the very creations they themselves nurtured.