Here, he’s been consistent; in March 2024 Karp told a CNBC reporter that ‘if you have a position that does not cost you ever to lose an employee, it’s not a position’
This statement by Alex Karp suggests a focus on employee turnover as a measure of company health, which may require further analysis of his management style.
The message received more than 50 ‘+1’ emojis
The popularity of this message among employees suggests a significant level of concern or agreement with the sentiment expressed.
The post—which includes many of Karp’s long-standing beliefs on how Silicon Valley could better serve US national interests—goes as far as suggesting that the US should consider reinstating the draft
This statement from a Palantir post suggests a strong political stance that may have influenced employee morale and perceptions of the company.
Karp gave an interview to CNBC claiming that AI could undermine the power of ‘humanities-trained—largely Democratic—voters’ and increase the power of working-class male voters
This statement by Alex Karp is a non-consensus view on the impact of AI, which may require further analysis of its implications and potential biases.
We hire the best and brightest talent to help defend America and its allies and to build and deploy our software to help governments and businesses around the world.
This statement from a Palantir spokesperson presents the company's mission, which contrasts with employee concerns and may require further analysis.
Employees could accept the intense external criticism and awkward conversations with family and friends about working for a company named after J. R. R. Tolkien’s corrupting all-seeing orb
This quote highlights a cultural perspective on Palantir that may have influenced employee morale and actions.
Last fall, Palantir seemed to become the technological backbone of Trump’s immigration enforcement machinery, providing software identifying, tracking, and helping deport immigrants on behalf of the Department of Homeland Security
This statement suggests a significant role of Palantir in immigration enforcement, which may need to be verified for accuracy and context.
At one point during the call, one of the employees tried to level with the group, explaining that Palantir’s work with ICE was a priority for Karp and something that likely wouldn’t change any time soon.
This statement indicates a high priority given to Palantir's work with ICE by the CEO, which may be a point of contention among employees.
Around this time, Palantir started wiping Slack conversations after seven days in at least one channel where most of the internal debate takes place, #palantir-in-the-news.
The deletion of Slack conversations could indicate a desire to suppress internal debate, which may be worth investigating further.
We were supposed to be the ones who were preventing a lot of these abuses. Now we're not preventing them. We seem to be enabling them.
This quote reflects a significant internal conflict within the company and may require further evidence to support the claim of enabling abuses.
Palantir was founded—with initial venture capital investment from the CIA—at a moment of national consensus following the September 11, 2001, attacks
The mention of CIA investment may raise questions about the company's initial intentions and potential biases in its operations.
Interviews with current and former Palantir employees, along with internal Slack messages obtained by WIRED, suggest a workforce in turmoil.
The claim of a workforce in turmoil is based on interviews and internal messages, which may not represent the entire employee base and could be biased.
Last fall, Palantir seemed to become the technological backbone of Trump’s immigration enforcement machinery, providing software identifying, tracking, and helping deport immigrants on behalf of the Department of Homeland Security
This statement suggests a significant role of Palantir in Trump's immigration enforcement, which may require further verification of the extent and nature of their involvement.
The company has also had to make major investments in its AI efforts in order to keep up with competitors in the space — earlier this month, it debuted a completely overhauled AI product called Muse Spark.
这里提到了 Meta 在 AI 领域的投资，需要探究这些投资的具体内容和回报，以及它们如何影响公司的整体战略。
Meta spent tens of billions on its metaverse efforts, which largely failed.
这是一个值得深入了解的背景信息，需要探究 Meta 在元宇宙上的具体投资和失败的原因，以及这些投资如何导致裁员。
This is not an easy tradeoff and it will mean letting go of people who have made meaningful contributions to Meta during their time here.
这句话可能带有一定的主观色彩，需要进一步了解 Meta 高管对于这次裁员的看法，以及他们对受影响员工的态度。
The cuts will begin on May 20.
这是一个具体的时间节点，值得关注 Meta 是否按时开始裁员，以及裁员的具体实施情况。
Meta also will not hire for 6,000 roles that are currently open.
这是一个重要的数据点，表明 Meta 不仅计划裁员，还将暂停招聘，这可能会对公司的长期招聘和扩张策略产生影响。
Meta is planning to cut 10% of its workforce, amounting to 8,000 employees, according to a report from Bloomberg.
需要核查的是，Meta 是否真的计划裁减 10% 的员工，即 8,000 人。这可能涉及到 Meta 的官方声明和相关的内部文件。
What used to take reps 5-6 hours a week now runs automatically in the background on every deal.
这是一个具体的效率提升数据，显示工作空间代理可以将销售代表每周5-6小时的工作自动化。这相当于每周节省约12.5%-15%的工作时间，是一个显著的效率提升，特别是在销售团队中。
Workspace agents will be free until May 6, 2026, with credit-based pricing starting on that date.
这是一个明确的时间节点和定价策略，表明OpenAI计划在2026年5月6日开始实施基于信用的收费模式。这个时间点距离发布日期(2026年4月22日)仅两周，可能是为了鼓励早期采用。
Workspace agents are available in research preview in ChatGPT Business, Enterprise, Edu, and Teachers plans.
这表明工作空间代理目前处于研究预览阶段，仅限于特定的商业和企业计划，尚未对所有用户开放。这种限制可能是为了控制测试范围和收集反馈，但也反映了产品仍处于早期发展阶段。
There has never been a more important time for us to stand up and show why science matters. I hope you'll support us in that mission.
这句话包含历史性断言'never been a more important time'，但缺乏量化数据支持。这种表述反映了当前对科学重要性的普遍认知，但需要具体指标如科学预算、政策变化或全球挑战的严重程度数据来验证这一历史性判断。
Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history.
180年的机构历史提供了重要背景，但'most critical moment'的主观判断缺乏量化依据。这种表述反映了媒体对当前科学重要性的强调，但需要具体数据支持这一历史性断言，例如科学资金、论文数量或政策变化的量化指标。
Lichtman is hopeful because ChatGPT's discovery validates a sense he's had since graduate school. 'I had the intuition that these problems were kind of clustered together and they had some kind of unifying feel to them,' he says.
这里提供了专业数学家的直觉判断，但缺乏量化数据支持。'clustered together'和'unifying feel'是模糊表述，无法验证。这反映了数学研究中直觉的重要性，同时也显示了当前AI辅助研究在提供可验证证据方面的局限性。
The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.
这里暗示了AI的创新性在于跨领域应用已知公式，而非创造全新数学。'well known'的表述表明这不是突破性发现，而是应用方式的创新。这种'组合创新'可能是AI在数学领域的主要贡献方式，需要更多关于具体公式和应用案例的数据支持。
The duo had jump-started the AI-for-Erdős craze late last year by prompting a free version of ChatGPT with open problems chosen at random from the Erdős problems website.
时间点'late last year'表明这种现象已持续数月，不是一时兴起。'随机选择'的方法暗示了大规模AI辅助数学探索的潜力，但文章未提供具体解决了多少问题或成功率，这些数据缺失限制了我们对AI数学能力的全面评估。
Erdős also noticed that the score drops if all of a set's numbers are large—the larger the numbers, the less large the score could become. He guessed that as the set's numbers approached infinity, the maximum score would drop to exactly one.
这个数据点提供了具体的数学预测值'1'，这是一个精确的量化结果。当数字趋近于无穷大时，分数降至1的预测展示了数学中的极限概念，这是AI可能帮助验证的精确数学命题。'exactly one'的表述强调了数学的精确性。
Erdős also came up with the Erdős sum, a 'score' you can calculate for any primitive set. He showed that the sum had a maximum possible value—and conjectured that this value must hold only for the set of all prime numbers.
这里提供了数学概念的具体量化指标。'最大可能值'的表述暗示了有明确的数学界限，但文章未提供具体数值。这反映了数学中某些概念虽然可量化，但具体数值可能需要更专业的数学背景才能理解，体现了数学研究的抽象性。
Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.
这个数据点突显了问题的难度和解决者的背景反差。60年的未解问题表明其复杂性，而23岁无高级数学训练的业余爱好者解决它，暗示AI可能正在改变数学研究的门槛和方式。这个年龄和背景信息增强了故事的戏剧性，但也需要更多关于Price教育背景的细节来全面评估。
They range dramatically in both significance and difficulty, and many AI solutions have turned out to be less original than they appeared.
大多数人认为AI在数学领域的突破都是具有高度原创性的，但作者指出许多AI解决方案实际上不如看起来那么原创，这挑战了我们对AI创新能力的过高期待。
Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve.
大多数人认为解决长期未解的数学难题需要顶尖数学家的专业知识和多年研究，但作者认为一个业余爱好者通过AI就做到了，这挑战了数学专业壁垒的传统观念。
I had the intuition that these problems were kind of clustered together and they had some kind of unifying feel to them. And this new method is really confirming that intuition.
大多数人认为数学问题之间通常是独立且需要不同方法解决的，但作者认为这些问题实际上是相互关联的，有统一的方法可以解决，这挑战了我们对数学问题多样性的传统认知。
The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.
大多数人认为数学突破需要全新的理论或方法，但作者认为AI只是应用了一个已知但未被想到应用于此问题的公式，这挑战了数学创新必须依赖全新方法的传统观念。
The question Price solved—or prompted ChatGPT to solve—concerns special sets of whole numbers, where no number in the set can be evenly divided by any other.
大多数人认为解决复杂的数学问题需要深入的专业知识和复杂的推理过程，但作者表明一个简单的概念（不能互相整除的数字集合）可以构成一个60年未解决的难题，挑战了人们对数学问题复杂性的认知。
But experts have warned that these problems are an imperfect benchmark of artificial intelligence's mathematical prowess. They range dramatically in both significance and difficulty, and many AI solutions have turned out to be less original than they appeared.
大多数人认为AI解决数学问题是其能力的有力证明，但作者认为这些问题作为AI数学能力的衡量标准是有缺陷的，挑战了人们对AI数学成就评估的普遍标准。
An AI researcher subsequently gifted them each a ChatGPT Pro subscription to encourage their 'vibe mathing.'
大多数人认为严肃的数学研究需要严谨的方法和深厚的专业知识，但作者使用'vibe mathing'这种非正式术语描述这种研究方式，挑战了学术研究方法论的传统规范。
We have discovered a new way to think about large numbers and their anatomy. It's a nice achievement. I think the jury is still out on the long-term significance.
大多数人认为AI的数学突破具有重大意义，但作者认为其长期意义尚不确定，这挑战了人们对AI数学成就重要性的普遍预期，暗示技术突破不一定等同于长期价值。
I had the intuition that these problems were kind of clustered together and they had some kind of unifying feel to them. And this new method is really confirming that intuition.
大多数人认为数学问题各自独立，需要不同的方法解决，但作者认为这些问题实际上有某种统一性，挑战了数学问题多样性和独立性的传统认知。
The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.
大多数人认为数学突破需要全新的理论或方法，但作者认为AI只是将已知公式应用到新领域就能取得突破，这挑战了人们对数学创新本质的理解，暗示创新有时来自于跨领域应用而非全新创造。
Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.
大多数人认为解决重大数学问题需要深厚的专业训练和多年经验，但作者认为一个23岁没有高级数学训练的业余人士也能解决60年悬而未决的问题，这挑战了学术界对专业资质的传统认知。
Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.
大多数人认为解决复杂的数学问题需要深厚的专业训练和多年经验，但作者认为一个没有高级数学训练的23岁年轻人仅凭AI工具就能解决困扰顶级数学家60年的问题，这挑战了数学领域的专业壁垒认知。
What he does have is a ChatGPT Pro subscription, which gives him access to the latest large language models from OpenAI.
大多数人认为数学成就主要依赖于个人智力和训练，但Price的成功关键是他拥有AI工具访问权限，这暗示在未来的数学领域，技术资源可能比个人能力更重要，挑战了传统天才观念。
Lichtman tried to prove this, too, but got stuck like everyone else before him.
大多数人认为数学突破来自于持续不断的努力和渐进式改进，但Lichtman和其他专家的失败表明，有时问题不在于努力程度而在于思维方式的局限，这挑战了我们对数学进步过程的认知。
An AI researcher subsequently gifted them each a ChatGPT Pro subscription to encourage their 'vibe mathing.'
大多数人认为严肃的数学研究需要严谨的方法和深厚的理论基础，但研究人员用'vibe mathing'这种非正式方式描述他们的工作，暗示数学发现可能源于看似随性的探索而非严格的规划。
I had the intuition that these problems were kind of clustered together and they had some kind of unifying feel to them. And this new method is really confirming that intuition.
大多数人认为数学问题是孤立的，需要不同的方法解决，但Lichtman的直觉表明这些问题可能有内在联系，AI的发现证实了这一观点，暗示数学领域可能存在尚未被发现的深层统一性。
The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.
大多数人认为数学突破需要全新的理论或方法，但AI的解决方案使用了已知公式只是应用到了新领域，这表明创新可能更多来自于跨领域应用而非全新发明，挑战了我们对数学创新本质的理解。
Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.
大多数人认为解决复杂的数学难题需要深厚的专业训练和多年经验，但这个案例表明，一个没有高级数学训练的23岁年轻人仅通过AI工具就解决了困扰顶尖数学家60年的问题，挑战了专业知识在数学突破中的必要性。
Resolution increases make them more expensive, then efficiency gains reduce costs - a sawtooth pattern.
大多数人可能认为AI成本会呈现单调下降或上升的趋势，但作者提出'锯齿状'模式，即精度提升导致成本上升，然后效率提升又降低成本。这种波动性挑战了人们对技术成本发展的常规预期。
Will smarter models be increasingly expensive because of greater accuracy or less expensive because they're smarter?
作者提出一个非共识的二分法：大多数人认为AI模型要么因更精确而更贵，要么因更智能而更便宜。但作者暗示这两种趋势可能同时存在，形成锯齿状的成本模式，这挑战了人们对技术成本发展的线性预期。
Smaller pieces force the model to pay closer attention to each word, like reading a contract word by word instead of skimming paragraphs.
大多数人认为更智能的AI会以更高效的方式处理信息，但作者指出，为了提高精确度，先进模型实际上需要更细致地处理每个词单元，这违背了人们对'智能'通常意味着'更高效率'的直觉认知。
Then Opus 4.7 shipped & the smarter model became much more expensive. The cause : a new tokenizer
大多数人认为AI模型变贵主要是因为能力提升，但作者揭示了一个反直觉的原因：更精确的分词器(tokenizer)导致需要处理更多token，从而使更智能的模型反而变得更贵。这挑战了'能力提升导致成本上升'的简单归因。
Opus 4.5 costs 67% more than Sonnet. But Opus 4.5 used 76% fewer tokens to reach the same outcome.
大多数人认为单位成本更高的模型总使用成本也会更高，但作者通过具体数据展示，尽管Opus 4.5的单token成本高出67%，但由于其效率大幅提升，实际完成任务的总成本反而降低了60%。这挑战了简单的线性成本思维。
When Anthropic launched Opus 4.5 in November 2025, the bigger, more expensive model was actually cheaper to use.
大多数人认为更先进的AI模型必然更昂贵，但作者指出Claude Opus 4.5作为更大、更先进的模型实际上使用成本更低。这挑战了'先进=昂贵'的普遍认知，展示了AI效率提升可能带来的成本反直觉现象。
The agent interprets new information and adapts the logic. The engine applies that logic continuously and emits precise updates.
大多数人认为AI代理应该完全负责从数据收集到决策执行的整个流程。但作者提出颠覆性的观点：AI应该专注于逻辑解释和适应，而将执行和持续评估交给专门的数据库引擎。这种分工模式挑战了当前AI代理应该全能化的主流认知。
Agents and CDC streams are powerful together because they split the work well.
大多数人可能认为AI代理应该独立完成所有任务，包括数据获取和处理。但作者提出反直觉的分工模式：AI专注于逻辑解释和适应，而数据库引擎专注于持续评估和精确更新。这种分工挑战了当前AI代理应该端到端处理所有任务的主流观点。
The fix is not smarter prompts. It is software built to meet agents halfway.
大多数人认为提高AI性能的关键在于更好的提示工程或更智能的模型。但作者认为解决方案在于重新设计软件架构，使其与AI代理更好地协作，而不是继续改进AI本身。这是一个颠覆性的观点，挑战了当前AI开发的主流方向。
Humans are not a good target for calm technology.
大多数人认为技术的目标应该是让人类更容易使用和理解。但作者提出相反观点：人类不适合作为'平静技术'的目标，因为当前的AI设计要求人类持续监督和互动，这与平静技术的本质相悖。
Today's agents, the copilots, the chatbots are designed to be human like.
大多数人认为AI助手应该模仿人类的交流方式，以便更好地与人类协作。但作者认为这种设计是错误的，因为它增加了认知负荷，违背了'平静技术'的理念。作者暗示AI应该更像是背景工具，而不是虚拟同事。
More than 3,000 forensic engines run in parallel on every submitted sample, covering signal, prosody, articulation, codec, and provenance domains.
3,000多个法证引擎并行运行展示了深度伪造检测的复杂性。这个数字表明检测系统需要从多个维度分析音频样本，才能准确识别合成语音。这也反映了随着AI技术的发展，检测技术也在不断进步和复杂化。
The FBI Internet Crime Complaint Center logged 2.3 billion dollars in losses for victims aged 60 and over in calendar year 2026.
60岁以上受害者在2026年损失高达23亿美元，这是一个惊人的数字。这表明老年群体是语音合成攻击的主要目标，他们可能更容易被紧急冒充电话所欺骗。这一数据强调了针对特定人群的网络安全教育的必要性。
Pindrop reported a 475 percent year-over-year increase in synthetic voice attacks against insurance call centers across 2025.
475%的年增长率表明语音合成攻击呈爆炸性增长。这一惊人的数字反映了AI语音技术的普及和攻击者利用这些技术的速度。保险公司成为主要目标是因为理赔主要通过电话处理，这使得语音验证成为关键安全环节。
The Wall Street Journal reported in February 2026 that high-quality voice cloning now requires roughly fifteen seconds of clean reference audio for tools available off the shelf.
15秒的干净参考音频是高质量语音克隆的门槛，而Mercor泄露的数据平均每个承包商有2-5分钟的录音，远超过这一阈值。这意味着攻击者可以使用泄露的数据创建非常逼真的语音克隆，大大增加了数据被滥用的风险。
According to the leaked sample index, the archive covers more than 40,000 contractors who signed up to label data, record reading passages, and run through verification calls for AI training.
40,000名承包商受到影响，这是一个相当大的数字。考虑到每个承包商提供了2-5分钟的录音，总录音时长可能达到80,000-200,000分钟，即约1,333-3,333小时。这个规模的数据泄露可能影响数百万最终使用这些AI系统的用户。
The dump is reported at roughly four terabytes and bundles a payload that breach analysts have been warning about for two years: voice biometrics paired with the same person's government-issued identity document.
4TB的数据量表明这是一个大规模的数据泄露事件，相当于约100万首歌曲的音频数据。将语音生物识别与政府签发的身份文件配对是特别危险的组合，因为攻击者可以同时获得声音克隆的素材和身份验证的凭证。这种组合大大增加了数据被武器化的可能性。
Meanwhile, in reality, the only 'official' MeshCore is the github repo. It's the source of truth in terms of what is MeshCore, and Andy has never contributed to that.
大多数人认为拥有商标或域名的人自然拥有项目的'官方'地位，但作者坚持只有GitHub仓库才是真正的'官方'来源，这挑战了知识产权与项目官方身份之间的常规认知。
Since inception, the MeshCore development team have been working hard to build MeshCore. We've released more than 85 versions of the MeshCore Companion, Repeater and Room Server firmwares with support for more than 75 hardware variants. All of this has been hand crafted, by humans.
在当今AI辅助编程盛行的时代，大多数人认为利用AI工具加速开发是理所当然的，但MeshCore团队坚持所有代码都是手工编写，这挑战了软件开发行业的效率优先共识。
Andy Kirby did do an amazing job helping to promote the MeshCore project on his personal YouTube, but only promotes his own products now.
大多数人认为项目贡献者应该持续推广整个项目生态系统，但作者暗示Andy从推广整个项目转向仅推广自己的产品，这种转变在开源社区中是罕见的，通常不被视为最佳实践。
We have always been wary of AI generated code, but felt everyone is free to do what they want and experiment, etc.
大多数人认为在软件开发中使用AI工具是提高效率和创新的合理方式，但作者团队明确表示他们一直对AI生成的代码持谨慎态度，这反映了在开源社区中对AI代码质量控制的非主流立场。
This ultimately also leads to false positives, but my manual QA run verified it's maybe 5-10%.
大多数人认为AI检测系统应该追求零错误，但作者接受5-10%的误报率，这挑战了技术检测的完美主义标准。这种务实态度暗示在AI识别领域，准确率和实用性之间需要权衡，而非盲目追求完美。
LLM tend to use certain font combos like Space Grotesk, Instrument Serif and Geist
大多数人认为AI能模仿任何设计风格，但作者指出AI实际上有特定的字体偏好，这揭示了AI设计的局限性而非无限可能性。这一发现挑战了我们对AI设计能力的认知，表明AI可能只是复制而非真正创新。
Claude Code has led to a large increase in Show HN projects. So much, that the moderators of HN had to restrict Show HN submissions for new accounts.
大多数人认为AI工具提高了生产力，但作者将其与内容泛滥和平台限制直接关联，暗示AI不仅提高了数量还可能损害了社区质量。这种观点挑战了'AI总是进步'的乐观叙事，提出了技术应用的负面后果。
I guess people will get back to crafting beautiful designs to stand out from the slop. On the other hand, I'm not sure how much design will still matter once AI agents are the primary users of the web.
大多数人认为设计始终对用户体验至关重要，但作者质疑当AI成为主要网络用户时设计的重要性，这挑战了设计行业的核心假设。这一观点暗示设计可能从面向人类转向面向AI，彻底改变设计价值链。
Is this bad? Not really, just uninspired. After all, validating a business idea was never about fancy design, and before the AI era, everything looked like Bootstrap.
大多数人认为AI生成的设计是'坏的设计'，但作者认为这只是'缺乏灵感'，将其与Bootstrap时代相提并论，暗示这种设计平庸化是技术发展的自然循环而非灾难性退步。这种观点挑战了我们对设计价值的传统认知。
A designer recently told me that 'colored left borders are almost as reliable a sign of AI-generated design as em-dashes for text'
大多数人认为AI设计难以识别，但作者认为简单的视觉元素如彩色边框就能可靠地识别AI生成的设计，这挑战了我们对AI设计复杂性的认知。这种观点暗示AI设计实际上有可预测的模式，而非完全无法捉摸。
The good world is where everyone has AI, and not as a revokable privilege through an API, but through hard possession.
大多数人可能认为通过API访问AI是民主化和可扩展的方式，但作者认为真正的AI民主化应该是通过硬所有权（hard possession），挑战了当前AI服务的主流商业模式。
It works for Mars. I think there's so much value in colonizing Mars, and it's sad to me to see SpaceX diluting the mission buying up random AI bubble crap.
大多数人可能认为AI和太空探索都是值得追求的目标，但作者认为这两者存在冲突，暗示SpaceX在AI领域的投资分散了其火星殖民的核心使命，挑战了科技多元化发展的共识。
How does a normal person fit into Elon's world? What institutions will Elon leave behind? Is there any value in that society to art and culture?
大多数人认为马斯克的愿景（如火星殖民）是积极和令人向往的，但作者质疑这种社会对普通人和文化艺术的价值，暗示马斯克的愿景可能创造一个缺乏人文关怀的社会。
I can hear the rabid Elon fan defending him about Tesla patents or the Twitter algorithm or something, but those are not serious open source projects.
大多数人认为埃隆·马斯克的开源贡献（如特斯拉专利）是值得称赞的，但作者认为这些并非真正的开源项目，暗示马斯克的开源承诺是表面性的，与真正的开源精神（如Linux和Kubernetes）有本质区别。
Even the ideal version, industrial megaprojects at hyperhuman scale while constantly being out over your skis with leverage sounds hellish.
大多数人认为大型AI项目和工业规模的发展是进步和繁荣的象征，但作者认为这种超人类规模的项目听起来像是地狱般的体验，因为它可能导致过度杠杆化和不可持续的压力。
Commoditizing complements doesn't always work because focus is scarce even for the largest, fastest growing businesses.
大多数人认为科技巨头拥有无限资源实施各种战略，但作者指出即使是最大、增长最快的企业也面临注意力稀缺问题。这一观点挑战了规模经济理论，暗示过度扩张可能导致核心竞争力的稀释。
But plenty of categories survived through specialization or direct competition : cloud, travel, domain registration, social networking.
大多数人认为免费化策略会摧毁所有竞争领域，但作者认为通过专业化或直接竞争，某些领域如云计算、旅行等依然能够生存。这一观点挑战了技术决定论，强调了人类专业知识和差异化价值在AI时代的重要性。
Some categories never developed a competitive response to this strategy : email, advertising infrastructure, user-generated video.
大多数人认为所有商业领域都有能力应对颠覆性竞争，但作者指出某些类别如电子邮件、广告基础设施等从未找到有效的竞争对策。这暗示了某些市场结构可能存在根本性弱点，无法通过传统竞争策略应对免费化浪潮。
The risk of this strategy to the ecosystem is that it makes previously attractive categories no longer viable. Commoditizing the complement does not demand a best-in-class replacement.
大多数人认为市场竞争会推动产品持续创新和改进，但作者认为免费化策略实际上降低了市场对卓越产品的需求，因为'足够好'的免费产品就能改变市场动态。这一观点挑战了传统创新经济学理论，暗示市场可能因免费化而停滞。
Several correlated but not strictly identical changes happened over the same few months: scaling inference compute, heavier use of RL in post-training, and models producing reasoning tokens.
大多数人可能将AI能力加速归因于单一因素（如模型规模增大），但作者指出这是多种因素共同作用的结果，包括推理计算扩展、强化学习在训练后阶段的使用增加以及模型生成推理标记的能力。这一多元归因挑战了单一因素决定论。
Tasks where correctness is harder to verify may not have seen the same speedup, so the acceleration we document here may not be as general as the headline numbers suggest.
大多数人可能被媒体报道的AI加速数据所影响，认为所有AI任务都在加速，但作者明确指出，那些正确性难以验证的任务可能没有相同的加速速度。这一观点挑战了人们对AI能力普遍加速的乐观预期。
The three metrics where we find acceleration are concentrated in programming and mathematics. These are areas that labs have explicitly targeted for improvement, and they share an important property: correctness is easy to verify automatically.
大多数人可能认为AI能力的加速是跨领域普遍发生的，但作者指出加速主要集中在编程和数学领域，因为这些领域正确性容易自动验证。这一发现挑战了人们对AI能力普遍提升的假设，暗示加速可能是有选择性的。
Our fourth metric, an index constructed from WeirdML V2 results, showed no sign of acceleration. A single global linear trend fit the data best.
大多数人可能认为所有AI能力指标都应该同步加速，但作者发现WeirdML V2指标没有显示出任何加速迹象，最佳拟合仍是简单的全局线性趋势。这一发现表明AI能力的加速并不是普遍现象，而是特定于某些任务领域。
Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.
大多数人认为不同AI模型之间的性能差异是渐进式的，但作者发现推理模型不仅一次性实现了性能跃升，而且以比非推理模型快2-3倍的速度持续进步。这一发现挑战了人们对AI模型性能提升方式的常规理解。
Three of the four metrics (ECI, log METR 50% time horizon, and a math-focused index we constructed from several math benchmarks) show strong evidence that progress has sped up relative to a global linear trend fit to data from 2023 onward.
大多数人认为AI能力提升是渐进式的线性发展，但作者通过数据分析发现，在三个关键指标上，AI能力实际上已经加速，这挑战了人们对AI发展速度的普遍认知。这种加速现象发生在2023年之后，与推理模型的发布时间点吻合。
Each cell shows how often a given curve fit is not significantly worse than the fit with the best cross-validation accuracy.
研究使用交叉验证来评估不同曲线拟合的优劣，每个单元格显示给定曲线拟合与最佳拟合相比不显著差于的频率。这种方法提供了更稳健的统计评估，减少了过拟合风险。
We examine whether AI capabilities are accelerating by fitting statistical models to benchmark performance over time, and comparing their predictive accuracies.
研究方法基于统计模型拟合和预测准确度比较，这是一种严谨的方法论。通过比较不同曲线拟合的预测能力，可以更客观地判断是否存在加速趋势，而非仅凭直观观察。
Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.
推理模型性能提升速度是非推理模型的2-3倍，这是一个显著的增长率差异。这个倍数关系表明推理模型确实带来了质的飞跃，但需要考虑这是否反映了模型架构的根本改进，还是仅仅由于更多计算资源的投入。
Three of four metrics show strong evidence of acceleration, driven by reasoning models.
文章核心发现，75%的指标显示AI能力正在加速，且主要由推理模型驱动。这是一个明确的量化结论，但需要关注的是，仅基于4个指标就得出'加速'的结论可能存在样本偏差，特别是这些指标主要集中在数学和编程领域。
Our fourth metric, an index constructed from WeirdML V2 results, showed no sign of acceleration. A single global linear trend fit the data best.
这个25%的指标没有显示出加速趋势，提供了一个重要的对比案例。作者推测这可能是因为WeirdML V2设置了资源限制环境(模型只有5次提交代码的机会，无法使用外部工具)，这与当前RL训练的重点不符。这表明AI进步可能高度依赖于测试环境和评估标准。
We have been calling this the 'reasoning' / 'non-reasoning' split, but this is not a perfectly clean dichotomy. Several correlated but not strictly identical changes happened over the same few months: scaling inference compute, heavier use of RL in post-training, and models producing reasoning tokens.
这里承认了分类方法的局限性，指出2024年左右的AI能力加速可能是由多个因素共同作用的结果，而非仅仅是推理能力的提升。这表明文章作者对数据的复杂性有清醒认识，但缺乏对这些因素相对重要性的量化分析。
The best-performing model across these three metrics was a pair of independent linear trends: one for reasoning models and one for non-reasoning models.
这个模型选择结果(100%的三个指标)表明将模型分为推理和非推理两类是最优预测模型。这提供了强有力的统计证据，支持推理能力可能是AI加速发展的关键因素。然而，文章没有详细说明如何定义推理模型，这可能影响结果的可靠性。
Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.
这是一个重要的性能对比数据，表明推理模型比非推理模型的进步速度快2-3倍。这是一个显著的加速比率，暗示推理能力的突破可能代表了AI发展的一个转折点。然而，文章没有提供具体的基准测试数据来支持这一倍数关系，需要谨慎对待。
Three of the four metrics (ECI, log METR 50% time horizon, and a math-focused index we constructed from several math benchmarks) show strong evidence that progress has sped up relative to a global linear trend fit to data from 2023 onward.
这是一个关键的统计数据，表明75%的AI能力指标显示出加速趋势。文章使用2023年后的数据进行线性拟合，发现三个指标偏离了线性趋势。这个比例相当高，但值得注意的是，样本量较小(n=4)，可能影响统计显著性。需要更多指标来验证这一发现。
Our website uses cookies to enhance your browsing experience and analyze site traffic.
网站提到使用cookies分析流量，但没有提供具体的流量数据、用户会话数或页面浏览量等关键指标，无法进行量化分析。
Have a question? Noticed something wrong? Let us know.
网站提供了反馈表单，但没有提供任何关于反馈数量、响应时间或用户满意度的具体数据，此处缺乏量化依据。
Subscribe
页面中只有一个订阅按钮，但没有提供具体的订阅数据、用户数量或转化率，无法进行任何有意义的量化分析。
Get the latest from Epoch AI in your inbox
网站提供了一个订阅选项，但没有提供具体的订阅者数量或增长率数据，此处缺乏量化依据。
© 2026 Epoch AI
页面显示的版权日期为2026年，这表明页面可能被预发布或是一个占位符。当前实际年份是2023年，这个时间跨度暗示网站可能被错误配置。
Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.
大多数人认为软件供应链攻击通常是针对特定生态系统或缓慢扩散的，但作者展示了跨生态系统的快速级联攻击。这种攻击速度和范围远超传统认知，表明现代软件供应链的脆弱性被严重低估。
Modern-day security tooling looks for the wrong things. Most software composition analysis tools work by checking your dependencies against a database of known vulnerabilities – CVEs. But a deliberately planted backdoor doesn't have a CVE.
大多数安全团队依赖CVE数据库来评估风险，但作者指出这种方法对故意植入的后门完全无效。这一观点挑战了行业共识，暗示现有安全工具在新型供应链攻击面前已经过时，需要转向行为分析等新方法。
The result is a mismatch that should terrify anyone building software: the attack surface is expanding faster than any human can monitor, and the entities making dependency decisions are increasingly not human.
大多数人认为安全问题可以通过增加人力监控和审查来解决，但作者认为在AI时代，攻击面扩展速度已经超过了人类监控能力，且依赖决策越来越由AI而非人类做出。这一观点挑战了传统安全理念，暗示需要全新的自动化防御机制。
Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication. That's roughly 63,000 times faster than the industry average.
令人惊讶的是：Socket公司在Axios攻击发布后仅6分钟就检测到恶意依赖，这比行业平均水平快约63,000倍。这种速度差异凸显了传统安全工具与新型行为检测方法之间的巨大鸿沟，也展示了早期检测在防止供应链攻击中的关键作用。
Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.
令人惊讶的是：一个单一的访问令牌可以在短短八天内横跨五个主要生态系统（GitHub Actions、Docker Hub、npm、PyPI和VS Code扩展市场），自动传播恶意代码，影响数千个组织。这种级联供应链攻击展示了现代软件生态系统的脆弱性。
The industry average time to detect a supply chain breach is 267 days. SolarWinds went undetected for 14 months. XZ Utils took two years to surface.
令人惊讶的是：软件供应链漏洞的平均检测时间长达267天，有些攻击如XZ Utils甚至需要两年才被发现。这意味着攻击者有充足的时间在系统中潜伏并造成广泛损害，而组织往往在损害发生后才意识到问题。
You can open the Threads Sidebar from the icon in the bottom left, or via the keybinding option-cmd-j on macOS and ctrl-option-j on Linux and Windows.
文章提供了具体的键盘快捷键信息，这是一个具体的技术细节。option-cmd-j和ctrl-option-j是跨平台的快捷键组合，表明设计考虑了不同操作系统的用户习惯。这些具体的技术细节增加了文章的实用性，但缺乏关于这些快捷键的使用频率或用户满意度数据。
Ask ten different programmers how they use AI, and you can get ten different answers.
文章使用'十个程序员'的例子来说明AI使用方式的多样性，这是一个具体的样本数量。这个数字虽然小，但有效地说明了开发社区对AI工具的态度差异。这种表述方式简洁有力，但缺乏更大规模的调研数据来支持这一观察。
It took us longer, and we won't lie, it drove us a little crazy.
文章提到开发过程'花费了更长时间'，这是一个时间跨度的定性描述。虽然缺乏具体的时间数据，但这句话暗示了开发过程的复杂性和挑战性。这种表述增加了文章的人性化色彩，但缺乏具体的时间节点或与其他项目开发周期的对比数据。
We spent days loading the system with hundreds of threads, refining rough edges and polishing corners that developers may never see.
文章提到团队使用'数百个线程'进行了数天的压力测试，这是一个具体的工作量指标。'数百个'虽然不是精确数字，但表明系统设计考虑了大规模并发场景。这种大规模测试表明开发团队对系统稳定性的重视程度，但缺乏具体的线程数量上限和性能指标数据。
All of this runs at Zed's famously buttery-smooth 120 fps
文章声称Zed以120fps的流畅度运行，这是一个非常具体的技术性能指标。120fps远高于大多数编辑器的60fps标准，表明Zed在处理多代理任务时仍能保持极高的渲染性能。这个数据点对于评估Zed作为开发工具的响应能力具有重要意义，但文章未提供基准测试数据来支持这一说法。
Elevate your brand to the forefront of conversation around emerging technologies
这是一个营销声明，但缺乏具体数据支持。没有提供广告效果、转化率或投资回报率等关键指标。这种表述过于笼统，无法评估其广告服务的实际价值和效果。
Founded at the Massachusetts Institute of Technology in 1899
这个时间点与当前日期(2026年)相比，意味着该机构已经运营了127年。这使其成为美国历史最悠久的科技媒体之一，经历了从电力时代到数字时代的多次技术变革，积累了丰富的行业洞察。
an unmatched audience of technology and business elite
这是一个定性描述而非量化数据。虽然暗示了读者群体的高质量，但没有提供具体用户数量、人口统计特征或与竞争对手的对比数据。这种表述缺乏可验证性，难以评估其市场定位的准确性。
From event sponsorships to custom content to visually arresting video storytelling
这里列举了三种广告形式，但没有提供具体数据或比例。这是一个缺乏量化依据的描述，无法评估各种广告形式的商业价值或受众覆盖率。对于广告效果分析，需要更具体的投入产出比数据。
We weren't able to find the page you were looking for.
这是一个404错误页面的标准提示，表明请求的URL不存在。虽然这不是文章内容，但作为网页错误信息，它反映了链接失效的问题，可能意味着原文章已被删除或URL结构发生变化。
Founded at the Massachusetts Institute of Technology in 1899
这个数据点表明MIT Technology Review有着127年的历史，是一家具有悠久传统的科技媒体。这个时间跨度意味着该机构经历了多次技术革命，其历史积淀为其内容提供了独特的视角和权威性。
delivering meaningful compute in the next three months and nearly 1GW in total before the end of the year
未来三个月内将提供有意义的计算能力，到今年年底前总计近1GW，这一时间表和规模显示了Anthropic应对当前需求压力的具体计划。1GW的规模虽然远低于5GW的总承诺，但代表了短期内显著的容量增加。这一数据点反映了AI基础设施需求与供应之间的紧张关系，以及公司对快速扩展能力的重视。
Significant Trainium2 capacity is coming online in Q2 and scaled Trainium3 capacity is expected to come online later this year
明确提到Trainium2芯片将在第二季度上线，而Trainium3芯片将在今年晚些时候上线，提供了具体的时间节点。这一数据点显示了芯片技术迭代的快速节奏，以及Anthropic与AWS在硬件路线图上的紧密合作。这种快速迭代能力对于保持AI模型的竞争力至关重要，但也带来了基础设施规划和成本控制的挑战。
run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025
年收入从2025年底的约90亿美元增长到超过300亿美元，增长率超过233%，这是一个惊人的增长速度。这一数据表明AI服务市场的爆发式增长，以及Anthropic在商业化方面的显著进展。然而，如此高的增长率是否可持续存疑，且300亿美元的年收入对于一家成立不久的AI公司来说相当惊人，需要更多财务细节来验证。
Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future
亚马逊对Anthropic的50亿美元投资（加上潜在的额外200亿）是AI领域最大的战略投资之一。这一数据点不仅反映了亚马逊对Anthropic技术的信心，也表明了云服务提供商与AI公司之间日益紧密的合作关系。与之前亚马逊已投资的80亿美元相比，这一新增投资显示了亚马逊对Anthropic未来发展的长期看好。
committing more than $100 billion over the next ten years to AWS technologies
未来十年投入超过1000亿美元用于AWS技术，这是一个惊人的数字，远超大多数科技公司的年度资本支出。这一长期承诺显示了Anthropic对AWS基础设施的深度依赖，以及他们对未来AI发展所需计算资源的巨大预期。这一投入规模也暗示了AI基础设施成本将持续上升。
over one million Trainium2 chips to train and serve Claude
使用超过100万颗Trainium2芯片的数据，展示了Anthropic在AI硬件部署上的巨大规模。这一数字不仅反映了计算能力的投入，也显示了与AWS在芯片定制上的深度合作。对于AI模型训练而言，百万级芯片的部署规模是行业顶尖水平，表明Claude可能需要大量计算资源进行训练和推理。
over 100,000 customers now run Claude on Amazon Bedrock
10万客户使用Claude在Amazon Bedrock上的数据，表明Anthropic的企业客户基础已经相当庞大。这一数字不仅反映了市场接受度，也验证了Claude作为企业级AI工具的商业价值。与OpenAI的GPT系列相比，这一客户量级显示出Anthropic在企业市场已取得显著进展。
up to 5 gigawatts (GW) of capacity for training and deploying Claude
5GW的算力规模是惊人的，相当于一个小型国家的电力消耗。这一数据表明Anthropic正在为AI模型训练和部署投入前所未有的基础设施资源，反映了大语言模型对计算资源需求的指数级增长。这一规模超过了大多数AI公司的基础设施投入，显示出Anthropic在AI基础设施竞争中的野心。
Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future. This builds on the $8 billion Amazon has previously invested.
大多数人认为科技巨头对AI公司的投资通常在数亿级别，但Amazon对Anthropic的总投资可能高达330亿美元，这远超行业共识。这种规模的投资表明科技巨头对AI基础设施的重视程度和投入规模正在以前所未有的方式增长，可能重塑AI行业的资本结构和竞争动态。
Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: AWS (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).
大多数人认为AI模型通常会与单一云平台深度绑定，形成生态系统锁定，但Claude同时出现在三大云平台上，这挑战了AI行业平台绑定策略的主流认知。这种多平台策略可能预示着AI模型提供商正寻求更大的市场覆盖和避免单一平台依赖，改变行业竞争格局。
Anthropic will also use incremental capacity for Claude in Amazon Bedrock. The agreement includes expansion of inference in Asia and Europe to better serve Claude's growing international customer base.
大多数人认为AI模型主要在美国市场发展，但Anthropic明确表示正在大力扩展亚洲和欧洲市场，这挑战了AI服务主要集中在美国的共识。这种全球扩张速度表明AI市场的地理分布正在迅速多元化，可能重塑全球AI产业格局。
Our run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025.
大多数人认为AI公司仍处于烧钱阶段，难以实现盈利，但Anthropic的收入在短短几个月内增长了三倍多，达到300亿美元的年化收入。这一惊人的增长速度挑战了AI行业普遍亏损的共识，表明AI模型商业化可能比预期更快、规模更大。
We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude
大多数人认为AI公司主要依赖通用GPU芯片训练模型，但Anthropic与Amazon的合作表明他们正大规模采用专用AI芯片(Trainium)，这挑战了行业对通用芯片依赖的主流认知。5GW的容量远超大多数AI公司的规模，反映了专用芯片在AI训练中的经济性和效率优势正在被重新评估。
This card was updated on April 24, 2026, to include additional information about safeguards for the deployment of GPT‑5.5 and GPT‑5.5 Pro in the API.
大多数人认为系统卡应该在发布时包含所有相关信息，不需要后续更新，但OpenAI在发布后仅一天就更新了系统卡以增加API部署的安全措施信息。这挑战了科技产品文档管理的常规做法，暗示AI安全措施是动态发展的，需要持续调整，这违背了传统软件发布中'文档一次性完成'的共识。
We separately evaluate GPT‑5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
大多数人认为如果两个模型使用相同的基础架构，它们的风险和安全需求应该相似，但OpenAI明确表示GPT-5.5 Pro需要单独评估，因为'设置可能显著影响相关风险或适当的安全措施立场'。这挑战了AI评估领域普遍认为的'相同基础模型的安全特性一致'的共识，暗示即使是微小的设置变化也可能导致显著不同的风险特征。
We are releasing GPT‑5.5 with our strongest set of safeguards to date, designed to reduce misuse while preserving legitimate, beneficial uses of advanced capabilities.
大多数人认为更强的安全限制会不可避免地限制AI的功能和实用性，但OpenAI声称他们能够同时实现'减少滥用'和'保留合法、有益的高级功能使用'。这挑战了AI安全领域普遍存在的'安全与功能之间存在权衡'的共识，暗示他们已经找到了一种创新的方法，可以在不牺牲功能的情况下增强安全性。
GPT‑5.5 understands the task earlier, asks for less guidance, uses tools more effectively, checks it work and keeps going until it's done.
大多数人认为AI模型需要持续的人工指导和监督才能完成复杂任务，但作者声称GPT-5.5能够'理解任务更早，要求更少的指导，更有效地使用工具，检查工作并持续进行直到完成'。这挑战了AI领域普遍认为的'当前AI系统仍需大量人类监督'的共识，暗示GPT-5.5已经实现了更高程度的自主性。
We subjected the model to our full suite of predeployment safety evaluations and our Preparedness Framework, including targeted red-teaming for advanced cybersecurity and biology capabilities
大多数人认为AI安全评估主要集中在防止直接有害输出，但OpenAI特别强调了对'高级网络生物学能力'的针对性红队测试。这暗示GPT-5.5可能具有比预期更强大的生物相关能力，这违背了AI领域普遍认为的'语言模型主要处理文本信息'的共识，表明AI已经深入到专业科学领域。
That momentum is starting to extend beyond engineering. Teams are using Codex to pull together context from different tools, reason through what matters, and turn scattered information into useful work - like briefs, plans, checklists, drafts, and follow-ups.
文章提到Codex的使用范围正在从工程扩展到其他领域，但未提供具体的使用案例数据或采用率。此处缺乏量化依据，无法评估Codex在企业非工程团队中的实际应用程度和价值。
Our professionals are using Codex to move from static requirements to working solutions in hours, not weeks. It's enabling rapid prototyping, real-time workflow redesign, and faster iteration across the development lifecycle.
Accenture首席AI官声称将开发时间从'周'缩短到'小时'，这是一个显著的效率提升声明，但缺乏具体数据支持。此处缺乏量化依据，无法验证这一断言的真实性或普遍适用性。
Today, those partners include Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, and Tata Consultancy Services (TCS).
文章列出了7家全球系统整合合作伙伴(GSIs)，这些都是大型IT咨询和系统集成公司。这一合作策略表明OpenAI正在通过这些拥有丰富企业客户资源的合作伙伴来加速Codex在企业市场的渗透，但未提供这些合作伙伴的客户覆盖范围或预期增长数据。
Companies are using Codex across the software development lifecycle. Virgin Atlantic is using it to increase test coverage and increase team velocity - reducing technical debt and improving performance.
虽然文章提到了Virgin Atlantic使用Codex的具体应用场景，但没有提供任何量化数据来衡量其效果。此处缺乏量化依据，无法评估Codex实际带来的性能提升或技术债务减少程度。
In early April, we shared that more than 3 million developers were using Codex every week. Just two weeks later, that number has grown to more than 4 million.
这表明Codex的开发者采用率在两周内增长了33.3%（从300万增加到400万），这是一个惊人的增长率。这种快速增长反映了开发者对AI编程工具的强烈需求，也暗示了Codex可能正在经历病毒式传播或企业快速采用阶段。
Accepted applicants and collaborators must have existing ChatGPT accounts to apply
大多数人认为安全测试应独立于产品生态系统，但OpenAI要求申请者必须是现有ChatGPT用户，这打破了传统安全测试的独立性原则，表明他们认为平台内测试能提供更相关和实用的安全洞察。
Once selected, successful applicants will be onboarded to the bio bug bounty platform
大多数人认为AI安全测试应是开放和民主化的，但OpenAI采用邀请制并限制在'受信任的红色团队测试者'范围内，这与众包安全测试的主流趋势相悖，表明他们认为生物安全领域需要更严格的准入控制。
All prompts, completions, findings, and communications are covered by NDA
大多数人认为安全漏洞信息应公开以促进集体防御，但OpenAI要求所有发现都受保密协议保护，这与开源安全理念相悖，表明他们认为生物安全领域的特殊性质需要不同于传统网络安全的信息控制。
Smaller awards may be granted for partial wins at our discretion
大多数人认为安全测试要么成功要么失败，不应有'部分成功'的概念，但OpenAI明确表示会为'部分胜利'提供奖励，这打破了传统二元思维，表明他们重视渐进式安全改进而非仅追求完美解决方案。
$25,000 to the first true universal jailbreak to clear all five questions
大多数人认为AI安全漏洞不应被奖励，而应被消除，但OpenAI设立高额奖金鼓励研究人员寻找'通用越狱方法'，这挑战了传统安全观念，表明他们认为有价值的安全测试需要经济激励。
Testing universal jailbreaks for biorisks in GPT‑5.5
大多数人认为AI安全测试应专注于防止有害内容生成，但OpenAI主动邀请研究人员寻找'通用越狱方法'来突破生物安全限制，这挑战了传统安全思维，表明他们认为主动寻找漏洞比被动防御更有效。
🔹 **Rich World Knowledge:** Leads all current open models, trailing only Gemini-3.1-Pro.
这里提供了模型知识能力的相对排名：领先所有当前开源模型，但仅落后于Gemini-3.1-Pro。这是一个相对定位而非绝对性能数据。这种表述暗示DeepSeek-V4-Pro在知识广度上达到了接近顶级闭源模型的水平，这对需要广泛知识的应用场景具有重要意义。然而，缺乏具体的评估指标和分数，难以准确量化这一差距。
🔹 **Enhanced Agentic Capabilities:** Open-source SOTA in Agentic Coding benchmarks.
虽然文中没有提供具体的基准测试数据，但声称在代理编程基准测试中达到开源SOTA(最先进水平)。这是一个重要断言，但缺乏具体量化指标。如果属实，这将代表DeepSeek在AI代理能力方面的重大突破，特别是在代码生成和执行任务上。需要查看技术报告中的具体基准测试数据来验证这一声明。
⚠️ Note: deepseek-chat & deepseek-reasoner will be fully retired and inaccessible after Jul 24th, 2026, 15:59 (UTC Time).
这里明确指出了旧模型退役的具体时间节点：2026年7月24日15:59 UTC。这是一个精确的时间点，表明公司正在进行产品线更新换代。从发布日期(2026年4月24日)到退役日期，只有约3个月过渡期，用户需要尽快迁移到新模型，这可能反映了公司对新产品性能的高度自信。
🔹 **1M Standard:** 1M context is now the default across all official DeepSeek services.
DeepSeek V4将上下文长度提升到100万token，成为行业新标准。这一数据点意义重大，相比行业常见的32K-128K上下文窗口，提升了约8-31倍，能处理更长文档和复杂任务。这需要创新的注意力机制和内存管理技术支撑，文中提到的'Novel Attention: Token-wise compression + DSA'可能是实现这一突破的关键。
🔹 **DeepSeek-V4-Flash:** 284B total / 13B active params. Your fast, efficient, and economical choice.
DeepSeek-V4-Flash的参数规模明显小于Pro版本：总参数2840亿，活跃参数130亿。参数效率比约为4.6%，略高于Pro版本。这种参数设计使其在保持性能的同时实现更快响应和更低成本，适合需要快速响应的应用场景。
🔹 **DeepSeek-V4-Pro:** 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
这里提供了DeepSeek-V4-Pro的具体参数数据：总参数1.6万亿，活跃参数490亿。这种参数规模远超大多数开源模型，接近顶级闭源模型。参数效率比(活跃参数/总参数)约为3%，表明采用了稀疏激活技术，这可能是其性能与效率平衡的关键。
Ubuntu 26.04 LTS provides the strongest foundation for our confidential computing stack. It allows us to deploy a single securely designed image for all our verifiably private AI workloads across Intel, AMD, and NVIDIA hardware, with no platform-specific changes required.
引用自Tinfoil联合创始人，强调了Ubuntu 26.04 LTS在机密计算方面的优势，支持Intel、AMD和NVIDIA硬件上的单一安全镜像。这表明Ubuntu在跨平台机密计算方面的领先地位，为AI工作loads提供了统一的安全基础，减少了平台特定配置的需求。
Ubuntu now fully supports RVA23, the baseline standard for RISC-V. This ensures that teams innovating on RISC-V can take full advantage of the platform, including in mixed-architecture environments.
文章指出Ubuntu现在完全支持RISC-V的RVA23标准，这反映了Ubuntu对新兴架构的前瞻性支持。RISC-V作为一种开放指令集架构，正逐渐获得关注。Ubuntu的支持将促进RISC-V生态系统的成熟，特别是在混合架构环境中的应用。
TPM-backed full-disk encryption is now generally available in the Ubuntu installer.
文章提到TPM支持的全盘加密功能现在已在Ubuntu安装程序中普遍可用。这一安全功能将加密绑定到特定设备的TPM芯片上，大大提高了物理访问攻击的门槛。相比其他Linux发行版，Ubuntu将此功能集成到安装程序中，简化了企业部署安全系统的过程。
Ubuntu 26.04 LTS is the first LTS to expand the number of memory safe system components. In practice, this means new kernel drivers and subsystems written in Rust, as well as `sudo-rs` and `uutils``coreutils` bringing memory-safe reimplementations of foundational system tools such as `sudo`, `ls`, `cp`, and `mv`.
文章强调Ubuntu 26.04 LTS是首个增加内存安全系统组件的LTS版本，包括Rust编写的内核驱动和子系统，以及sudo-rs和uutils coreutils等内存安全的基础系统工具重实现。这一举措显著提高了系统的安全性，减少内存相关漏洞的风险，展示了Ubuntu在内存安全方面的领先地位。
Canonical Livepatch now extends its rebootless kernel patching capability to Arm64 for the first time.
这标志着Canonical Livepatch技术的重要里程碑，首次扩展到Arm64架构。对于运行Ubuntu的Arm64服务器和边缘设备，这意味着无需重启即可应用关键内核补丁，大大提高了系统可用性。这一功能的扩展反映了Ubuntu对ARM生态系统的持续投入。
IgH Master driver brings microsecond-level timing precision natively into the OS, removing a significant integration burden for engineers building motion control systems, robotics platforms, or complex factory automation.
文章提到EtherCAT驱动提供微秒级(10^-6秒)的时间精度，这对工业自动化应用至关重要。这种高精度时间同步能力是Ubuntu在工业领域的一个关键优势，相比其他通用操作系统，Ubuntu在实时性方面的改进使其更适合工业物联网和自动化场景。
Ubuntu 26.04 LTS is built on Linux 7.0, continuing Canonical's commitment to shipping the latest upstream kernels at the time of release.
文章明确指出Ubuntu 26.04 LTS基于Linux 7.0内核，这表明Canonical坚持使用最新上游内核的策略。相比其他可能使用更保守内核版本的Linux发行版，Ubuntu的这一策略确保了用户能够获得最新的硬件支持和性能改进。
With optimized images across AWS, Azure, Google Cloud, IBM Cloud and Oracle Cloud, developers and enterprises can rely on Ubuntu 26.04 LTS for their most demanding public cloud workloads.
文章提到Ubuntu 26.04 LTS支持5大主流云平台(AWS, Azure, Google Cloud, IBM Cloud, Oracle Cloud)，这反映了Ubuntu在云环境中的广泛兼容性。相比其他Linux发行版，Ubuntu在多云支持方面表现出色，这增强了其作为企业级操作系统的竞争力。
Ubuntu powers millions of PCs and laptops around the world.
这是一个模糊的数量描述，'millions'没有提供具体数字，无法确定Ubuntu的确切用户规模。相比其他Linux发行版如Red Hat或SUSE，Ubuntu确实拥有更广泛的桌面用户基础，但缺乏精确的市场份额数据支持这一说法。
The 11th long-term supported release of Ubuntu delivers deep silicon optimization and state-of-the-art security for enterprise workloads.
这表明Ubuntu 26.04是第11个LTS版本，按照Ubuntu每两年发布一个LTS版本的规律，这与Ubuntu的历史发展时间线一致。作为第11个LTS版本，它代表了Canonical在长期支持方面的成熟经验，为企业和用户提供稳定可靠的选择。
_Self-reported score with custom Anthropic scaffold._ SWEPro were evaluated with the mini-swe-agent scaffold. However, we use the scores reported by Anthropic for Opus with the max thinking efforts due to frequent timeouts during our evaluation trials.
脚注2揭示了重要数据点：Opus 4.6的53.4分是Anthropic的自报分数，因为作者在评估过程中频繁遇到超时问题，无法自行验证。这表明性能比较中存在数据可靠性问题，特别是对于Opus的评估依赖于厂商自报数据，可能存在偏差。
The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass.
文章描述了一种递归推理机制，称小模型通过自我迭代可以达到单次推理无法达到的结果，但未提供具体的性能提升数据或实验证据。这一断言缺乏量化依据，需要更多实验数据支持。
Sakana Fugu models are based on our ICLR 2026 papers (**Trinity** and **Conductor**), and we have substantially further improved the methods to increase the performance and user experience
文章提到模型基于ICLR 2026论文，并已大幅改进方法和用户体验，但没有具体说明改进的幅度或基准数据。此处缺乏量化依据，无法评估从研究原型到商业产品的改进程度。
Two variants are available: **Sakana Fugu Mini 🐟**, optimized with latency in mind, and **Sakana Fugu Ultra 🐡**, the full orchestration system, optimized for performance for demanding tasks.
文章提到有两种变体：Mini（延迟优化）和Ultra（性能优化），但未提供具体的性能指标差异，如延迟降低百分比或吞吐量提升数据。这种缺乏具体量化参数的描述难以评估两种变体在实际应用中的性能差异。
GPQAD | 94.4 | 90.9 | 92.7 | 92.4 | **95.1** | LCBv6 | 90.3 | 92.1 | 92.4 | 90.4 | **93.2** | SWEPro | 48.4 | 51.2 | _53.4_ | 51.3 | **54.2**
性能对比表格显示，Sakana Fugu Ultra在三个基准测试中均优于竞争对手：GPQAD上达95.1%（超越Gemini 3.1的94.4%），LCBv6上达93.2%（超越GPT 5.4的92.1%），SWEPro上达54.2%（超越Opus 4.6的53.4%）。这些数据表明其多模型协调策略确实带来了性能提升，特别是在科学推理任务上优势明显。
Initially, our Sakana Fugu model will be available as an **API**, where it has served as a key internal tool for our own researchers and engineers
这里提到Sakana Fugu模型将作为API提供，且已作为内部工具使用，但没有具体说明内部使用的时间跨度或用户数量。此数据点缺乏具体量化依据，无法评估其内部应用的规模和成熟度。
We believe this is what drove the separate reports of usage limits draining faster than expected.
大多数人会直接将API使用量异常归因于用户行为或模型本身，但作者揭示了一个技术实现细节（缓存bug）如何间接导致使用量异常。这挑战了常规问题归因逻辑，展示了系统组件间的意外互动如何产生看似无关的问题表象。
As part of this investigation, we ran more ablations (removing lines from the system prompt to understand the impact of each line) using a broader set of evaluations. One of these evaluations showed a 3% drop for both Opus 4.6 and 4.7.
大多数人认为微小的系统提示变更只会带来微不足道的影响，但作者展示了一个看似微不足道的提示变更（限制字数）却导致了3%的性能下降。这挑战了'小变更小影响'的直觉认知，揭示了AI系统中微小变化可能带来的非线性影响。
After multiple weeks of internal testing and no regressions in the set of evaluations we ran, we felt confident about the change and shipped it alongside Opus 4.7 on April 16.
大多数人认为充分的内部测试可以预防产品发布后的重大问题，但作者展示了一个经过数周内部测试且没有发现问题的系统提示变更却导致了明显的质量下降。这挑战了'测试覆盖率等于产品质量'的传统观念，暗示了评估指标与实际用户体验之间可能存在巨大鸿沟。
Two unrelated experiments made it challenging for us to reproduce the issue at first: an internal-only server-side experiment related to message queuing; and an orthogonal change in how we display thinking suppressed this bug in most CLI sessions
大多数人认为复杂的系统测试流程应该能够发现大多数关键缺陷，但作者展示了即使有多重测试机制，两个看似无关的实验如何协同掩盖了一个严重bug。这挑战了'全面测试能保证产品质量'的传统认知，揭示了系统复杂性带来的意外风险。
In our internal evals and testing, medium effort achieved slightly lower intelligence with significantly less latency for the majority of tasks.
大多数人认为内部评估和测试足以代表用户真实体验，但作者承认他们的内部测试未能准确捕捉到用户对AI智能度的实际感知差异。这暗示了实验室环境与实际使用场景之间存在根本性脱节，挑战了传统产品测试方法论的有效性。
We reverted this change on April 7 after users told us they'd prefer to default to higher intelligence and opt into lower effort for simple tasks.
大多数人认为AI系统应该优化速度和效率，但作者认为用户更愿意默认选择更高智能而非更低延迟，这挑战了产品优化的常规思维。用户宁愿忍受偶尔的延迟也要换取更高的代码质量，这违背了大多数科技公司追求'更快更省'的常规做法。
The products will need to get worse, more expensive, or both if VCs are to get their money back.
主流观点认为科技公司会通过产品创新和改进来提高价值，但作者直言AI公司可能需要让产品变得更差或更昂贵才能满足投资者回报要求，这挑战了科技行业进步的叙事，揭示了资本压力与产品价值之间的潜在冲突。
Anthropic made fun of this idea during the last Super Bowl.
大多数人认为广告是AI公司实现盈利的可行途径，特别是考虑到免费服务的模式，但作者指出Anthropic公开嘲笑广告模式，暗示AI行业内部对商业模式存在根本性分歧，挑战了广告作为AI盈利解决方案的主流观点。
Open weight (read: free) models are widely available and good enough that most people probably couldn't tell the difference.
主流观点认为付费的云端LLM服务在质量上显著优于免费开源模型，但作者声称开源模型已经好到大多数用户无法分辨差异，这挑战了付费服务价值主张的核心，暗示AI行业可能面临价值重估。
the system achieved this training result more than 20 times faster than conventional synchronization methods.
大多数人认为分布式训练由于需要同步和通信，必然比单机训练慢，但作者认为Decoupled DiLoCo比传统同步方法快20倍以上，这挑战了人们对分布式训练速度的固有认知，展示了异步计算的潜力。
chips from different generations running at different speeds still matched the ML performance of single-chip-type training runs, ensuring that even older hardware can meaningfully accelerate AI training.
大多数人认为混合不同代际的硬件进行训练会降低性能或效率，但作者认为即使不同代际、不同速度的芯片混合使用，仍能达到与单一芯片类型训练相同的机器学习性能，这挑战了硬件必须同质化的行业共识。
With increasing levels of hardware failure, Decoupled DiLoCo continues to deliver a high level of 'goodput', or useful training, while that of other approaches nosedives.
大多数人认为硬件故障会显著降低分布式训练的效率和性能，但作者认为即使在硬件故障率极高的环境下，Decoupled DiLoCo仍能保持88%的有效训练率，而传统方法则暴跌至27%，这挑战了人们对故障容忍能力的传统认知。
By dividing large training runs across decoupled 'islands' of compute, with asynchronous data flowing between them, this architecture isolates local disruptions so that other parts of the system can keep learning efficiently.
大多数人认为分布式AI训练需要高度同步和紧密耦合的系统才能保证效率，但作者认为通过解耦的'计算岛屿'架构，即使局部硬件故障，系统其他部分仍能高效学习，因为故障被隔离了。这挑战了传统分布式训练必须保持同步的主流认知。
The Prompt API uses the Gemini Nano model in Chrome. While the API is built into Chrome, the model is downloaded separately the first time an origin uses the API.
大多数人认为内置API应该包含所有必要组件，无需额外下载，但作者明确指出模型需要单独下载。这与人们对'内置'API应该即开即用的普遍认知相悖，暗示用户首次使用时可能会面临显著的下载时间和存储压力。
The Prompt API for the web is still being developed. While we build this API, refer to our best practices on session management for optimal performance.
大多数人认为浏览器AI功能应该是成熟且生产就绪的，但作者明确表示该API仍在开发中。这与人们对Chrome作为成熟浏览器应该提供稳定可靠功能的认知相悖，暗示AI功能可能还不够稳定，需要开发者额外注意性能优化。
The network requirement is only for the initial download of the model. Subsequent use of the model does not require a network connection. No data is sent to Google or any third party when using the model.
大多数人认为使用Google的AI模型必然会涉及数据传输和隐私问题，但作者强调模型完全在设备上运行且不向Google发送数据。这与人们对大型科技公司AI服务通常涉及数据收集的普遍认知相悖，暗示Chrome的AI功能可能比想象的更加注重隐私保护。
The Prompt API isn't available in Web Workers for now, due to the complexity of establishing a responsible document for each worker in order to check the permissions policy status.
大多数人认为现代浏览器API应该支持Web Workers以实现并行处理，但作者明确表示Prompt API不支持Web Workers。这与人们对浏览器API应该全面支持现代Web开发模式的认知相悖，限制了开发者在后台线程中使用AI的能力。
Microsoft continues to participate directly in OpenAI's growth as a major shareholder.
大多数人认为在修改了合作协议后，微软可能会减少其在OpenAI的股权投资，但作者认为微软仍然是OpenAI的主要股东，这表明尽管合作关系有所调整，但双方仍然保持着深度的利益绑定，这可能是一种非传统的长期战略伙伴关系模式。
Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI's technology progress, at the same percentage but subject to a total cap.
大多数人认为随着OpenAI技术的发展，其对微软的支付可能会增加或调整，但作者认为OpenAI对微软的支付将保持固定比例且有上限，这表明OpenAI正在寻求更可预测的财务安排，不受技术进步的影响，这可能是一种反直觉的风险管理策略。
Microsoft will continue to have a license to OpenAI IP for models and products through 2032. Microsoft's license will now be non-exclusive.
大多数人认为微软会寻求对OpenAI技术的独家使用权，以保持其在AI领域的竞争优势，但作者认为微软的许可权变为非独家，这打破了传统科技合作中的排他性模式，暗示OpenAI正在向更开放的合作方式转变，可能为其他合作伙伴铺平道路。
Microsoft will no longer pay a revenue share to OpenAI.
大多数人认为微软作为OpenAI的主要投资者和合作伙伴，会继续通过收入分成来支持OpenAI的发展，但作者认为微软已经改变了这一模式，这可能表明微软认为OpenAI的技术已经足够成熟，不再需要这种财务激励，或者微软有其他方式从合作中获益。
OpenAI can now serve all its products to customers across any cloud provider.
大多数人认为OpenAI会完全依赖微软Azure云服务，因为微软是其主要投资者和合作伙伴，但作者认为OpenAI现在拥有了多云策略的灵活性，这打破了科技巨头间典型的排他性合作模式，暗示OpenAI正在寻求更大的自主权和市场机会。
this means that existing estimates overstate the returns to software R&D, and makes the software intelligence explosion seem much less likely.
R&D Returns Overstated
Accounting for compute bottlenecks suggests that returns to software R&D may be lower than previously estimated, reducing explosion likelihood.
But I think we have enough evidence to think that software progress might really be several times a year, and to make a best guess contextualized with a lot of uncertainty.
Progress Estimation
Despite uncertainties, evidence suggests software progresses at several times per year, with estimates ranging from 2-50x annually.
gpt-oss-20b does substantially better than GPT-3 on MMLU, despite using the same amount of training compute.
Real-World Progress Example
Comparing models with same compute but different performance (like GPT-3 vs gpt-oss-20b) provides concrete evidence of software progress.
This means that almost all existing estimates of software progress were misleading.
Measurement Problems
Existing software progress estimates are misleading due to data quality improvements and scale-dependence factors not properly accounted for.
these estimates rely on an overly conservative estimate of software progress of 3× per year
Progress Underestimation
Existing software intelligence explosion models may use conservative progress estimates, potentially underestimating explosion likelihood.
Synthetic data can help push beyond this — a good example that Millidge raises is the Phi series of models.
Synthetic Data Impact
Synthetic data generation techniques like Phi models can dramatically improve efficiency beyond traditional distillation methods.