Altman says hoarding memory in personal computers is wasteful and it should be moved to the cloud
OpenAI CEO Sam Altman said in a public interview that users around the world hoarding large amounts of memory in personal computers is extremely wasteful, with most machines running at only 5% utilization, and that this high-bandwidth memory should be consolidated into cloud data centers rather than kept locally to run models. He believes that global advanced memory production capacity is a rigid physical bottleneck, and that idle personal devices are diverting strategic materials that should be used for data centers. The article also points out that this statement aligns with OpenAI's business stance of promoting cloud subscriptions, and that cloud efficiency does not equal users' privacy, security, and independent sovereignty.
View original postAfter downing a few glasses of dark beer, OpenAI CEO Altman hesitated in front of the camera, and finally could not hold back. Somewhat losing his composure but speaking very truthfully, he revealed one of OpenAI's deepest secrets. Altman's exact words were, "I probably shouldn't say this, but since we're drinking, I'll say it." A few days ago, after the new ChatGPT model sliced through mathematicians' lifelong research problems like cutting vegetables, he was so excited that he stayed up two nights in a row, yet at dawn he fell into deep sorrow: because he had watched with his own eyes the era in which humanity relied on flesh-and-blood brains to explore truth die completely before him 😭😭 Immediately afterward, he told the story of the crazy weekend inside OpenAI known as "Model Name-Hold My Beer." As soon as the new model ran successfully, a group of the world's top scientists went completely wild in the internal channel. As if posing exam questions to a deity, they threw in all at once the hardest research problems accumulated over decades of their academic careers, unsolved mysteries spanning every discipline, and dead knots that everyone was certain "a machine could absolutely never solve." Then, before everyone's eyes, that model, like autumn wind sweeping away fallen leaves, knocked down these ultimate fortresses of human intelligence one after another. For a single proof related to the Millennium Prize Problem, the Navier-Stokes equations, they poured in $10 million worth of compute. Altman admitted that the actual scientific discoveries produced far exceeded the several hundred results they had publicly released. He was so excited that he stayed awake all night two nights in a row, but the most spine-chilling part of the whole scene was the psychological cliff after his excitement: "I also had more sadness and melancholy than expected, and a feeling that... something is ending." This was absolutely not put-on PR rhetoric, but the instinctive lament of a person who had witnessed the Creator descend, uttered over the ruins. For thousands of years, the reason humanity could hold its head high and call itself the lord of all creation was that we firmly believed the flesh-and-blood brain contained divinity and insight that could not be reduced to formulas. No matter how advanced the tools, those deepest mathematical truths, physical laws, and scientific discoveries would always belong to the supreme crown of a few chosen ones. But on that weekend, that crown was violently and elegantly removed by row after row of cold GPU racks at a compute cost of tens of millions of dollars. The hardest propositions became a question of compute; a lifetime of ascetic pursuit became code run overnight. It proved that the upper limit of humanity's scientific exploration had been infinitely expanded, but it also sounded the death knell for the era in which humanity proudly looked down on the universe relying purely on flesh-and-blood intelligence. The person closest to superintelligence in the world held far more truths than were shown to the outside world, but after drinking, he sounded sadder than anyone 😭
View original
几杯黑啤下肚之后,OpenAI CEO 奥特曼在镜头前犹豫了一下最后没忍住,有点失态但又非常真实的吐露了OpenAI 最深的一个秘密, 奥特曼原话是“我其实可能不应该说这个,但既然我们在喝酒,我就说了。” 前几天在ChatGPT新模型把数学家们一辈子的研究难题像切菜一样打穿之后,他兴奋得连熬两夜,却在天亮时陷入了深深的悲伤: 因为他亲眼看着人类依靠肉身大脑探索真理的时代,在眼前彻底死去了😭😭 紧接着,他讲了 OpenAI 内部那个被称为“Model Name-Hold My Beer(看我的)”的疯狂周末。 新模型刚跑通, 一群全世界最顶尖的科学家在内部频道里彻底玩疯了, 大家像是在给神明出考题一样, 把各自学术生涯里积攒了几十年最难啃的研究、横跨各个学科的未解之谜、以及所有人都笃定“机器绝对不可能做出来”的死结,一股脑全扔了进去。 然后,那个模型在所有人面前,像秋风扫落叶一样,接二连三把这些人类智力的终极堡垒全部击倒。 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 万美元的算力。 奥特曼坦言,实际跑出来的科研发现, 远比他们对外公开发表的那几百个成果多得多。 他连着两个晚上兴奋得彻夜无眠, 但全场最让人脊背发凉的,是他兴奋之后的心理断崖: “我也有比预期更多的悲伤和忧郁,以及一种……某种东西正在结束的感觉。” 这绝不是装模作样的公关辞令,而是一个亲眼目睹了造物主降临的人,在废墟上发出的本能悲叹。 过去几千年,人类之所以能昂着头自封为万物之灵,是因为我们坚信肉身大脑里蕴含着不可被公式化的神性与顿悟。 无论工具多么发达,那些最深奥的数学真理、物理规律和科学发现,永远是属于少数天选之子的至高王冠。 但在那个周末,这顶王冠被一排排冷冰冰的显卡机柜,以千万美元的算力成本,暴力且优雅地摘了下来。 最难的命题变成了算力的问题,一生的苦修变成了通宵跑完的代码, 它证明了人类科学探索的上限被无限拓宽,但也宣判了人类纯粹依靠肉身智慧傲视宇宙的时代,正式敲响了丧钟。 那个世界上离超级智能最近的人,手里握着远比展示给外界更多的真理,但他喝完酒后,听起来比任何人都悲伤😭
Body · AI translation
In a recent public interview, OpenAI CEO Sam Altman threw out a controversial business take that set the entire internet's geeks ablaze: he thinks it's a sheer waste for ordinary people to hoard large amounts of memory in their personal computers. Altman stated bluntly that users around the world hoarding massive amounts of memory in their PCs is extremely wasteful,
because most machines have a utilization rate of only five percent, and this precious high-bandwidth memory should be concentrated in cloud data centers rather than left locally to run models.
Everyone assumes by default that how much hardware to install in a personal computer is the consumer's freedom, and that running lightweight open-source models locally is a kind of compute equalization for developers.
But the way big tech giants view hardware resources has never been about user freedom; it's about the concentration of compute across the entire industrial system.
In the ledgers of top-tier model companies, global production capacity for advanced memory is an absolutely rigid physical bottleneck, and tens of millions of idle personal devices sitting on desks are every minute and every second diverting strategic materials that should be piled into data centers for round-the-clock, high-load operation.
It's like before centralized power supply became widespread, when factory owners saw every household setting up small generators in their yards and would surely have felt that this decentralized possession was a huge waste of limited steel and fuel.
For all the tech people tinkering with local models and building personal workstations, this is an extremely sobering reminder: cloud giants have never wanted you to hold all your compute in your own hands, and between on-device inference and centralized APIs, a covert war over the supply of physical hardware is breaking out.
Of course, this rhetoric from the perspective of compute utilization perfectly fits OpenAI's business stance of promoting cloud subscriptions; cloud efficiency does not equal users' privacy, security, and independent sovereignty, and moving memory into the cloud means your compute faucet is also completely handed over to someone else.
The core contest has never been about how high memory utilization actually is, but about when AI becomes the next electricity, whether you own your own generator or can only pay the cloud for water, electricity, and gas month by month.