开源和开放权重的人工智能已经成熟为一股主导力量，从实验性尝试转变为全球数字基础设施的核心组成部分。到 2026 年中期，开放权重模型在编程、指令执行和通用知识等关键领域已与封闭的前沿模型达到实质性均势。尽管专有模型在复杂推理和长上下文检索方面仍占优势，开放模型却已占据大部分生产环境的 token 使用量。推动这一变化的是推理成本的大幅崩溃：过去三年下降了五十倍，使得对大多数企业而言，自主托管在财务上优于按量计费且由供应商控制的 API 。

尽管采用率很高，开源生态仍存在显著的运营缺口。虽然有 79% 的开发者在构建 AI 功能时使用开放模型，团队常常难以将原型推向生产。摩擦的根源并非模型能力不足，而是缺乏企业级工具、统一标准和可靠的维护。无论是小型组织还是大型企业，都将基础设施复杂性、安全与合规接入，以及维护定制化技术栈的难度列为开放 AI 部署的主要障碍。封闭提供方虽能提供"交钥匙"体验，但专有供应商锁定带来的运营成本——往往还隐藏在背后——正推动一波云回迁，企业希望收回对自身数据和流程的主权。

开放 AI 的战略重要性引发了全球性转向，超过 70 个国家正在制定强调主权与可选基础设施能力的 AI 政策。各国政府越来越把开放权重视作对冲外国出口管制和供应商停服等风险的手段。 China 尤其积极地将开源传播作为核心国家战略，用以规避半导体限制并加速本地创新。与此同时，像 European Union 这样的地区正把有利于主权化、开源化 AI 的要求制度化，确保国家数字基础设施保持在公共或本地控制之下，把 AI 问题从采购层面上升为国家政策问题。

随着产业演进，位于模型之上的"harness"——即编排循环、记忆与权限层——已成为新的控制战场。闭源实验室越来越多地把其专有的 harness 与模型捆绑，形成垂直一体化产品，构成实质性的护城河。这带来了"优化性锁定"的风险：harness 只有在提供方自己的权重上表现最佳。开源社区正在通过开发中立的框架与标准（例如 Model Context Protocol）予以回应，力求保持代理层的可互换性。目标是把模型保持为商品化、可替换的组成部分，同时为记忆与安全构建属于用户而非供应商的持久、可移植的系统。

归根结底，AI 的未来取决于社区能否解决所谓的"write surface"问题——即代理在现实世界执行动作的能力目前缺乏稳健且可移植的安全标准。鉴于当前对人工监督的依赖常因同意疲劳而失效，开源 AI 的下一次重大跃迁很可能来自能够强制执行有状态、基于策略治理的元级控制层（meta-harnesses）。通过对这些基础层——记忆、编排与权限标准——进行投资，开源运动可确保 AI 生态保持多元化，使构建者在一个供应商可控断开开关日益成为切实威胁的世界中，继续掌控其工具、成本与数据。 Open-source and open-weight artificial intelligence has matured into a dominant force, shifting from an experimental endeavor to a central component of global digital infrastructure. By mid-2026, open-weight models have reached effective parity with closed frontier models in critical areas such as coding, instruction-following, and general knowledge. While proprietary models still maintain an edge in complex reasoning and long-context retrieval, open models have captured a majority of production token volume. This surge is driven by a massive collapse in inference costs, which have fallen fiftyfold over the last three years, making self-hosting a financially superior alternative to metered, vendor-controlled APIs for most enterprises.

Despite this success in adoption, the open ecosystem faces a significant operational gap. While 79% of developers building AI functionality use open models, teams frequently struggle to move from prototype to production. This friction is not due to a lack of model capability, but rather a deficit in enterprise-grade tooling, standardization, and reliable maintenance. Smaller organizations and massive enterprises alike report that the primary hurdles to open AI deployment include infrastructure complexity, security and compliance integration, and the difficulty of maintaining custom stacks. While closed providers offer a "turnkey" experience, the operational, and often hidden, costs of proprietary vendor lock-in are driving a wave of cloud repatriation as companies seek to reclaim sovereignty over their own data and processes.

The strategic importance of open AI has led to a global shift, with over 70 nations developing AI policies that emphasize sovereignty and the ability to choose infrastructure. Governments are increasingly viewing open weights as a hedge against the volatility of foreign export controls and vendor shutdowns. China, in particular, has aggressively leveraged open-source dissemination as a core national strategy to bypass semiconductor restrictions and accelerate local innovation. Simultaneously, regions like the European Union are formalizing mandates that favor sovereign, open-source AI to ensure that national digital infrastructure remains under public or local control, moving the question of AI from a procurement issue to one of state policy.

As the industry evolves, the "harness"—the orchestration loop, memory, and permission layer sitting above the model—has become the new battleground for control. Closed-source labs are increasingly integrating their own proprietary harnesses with their models, creating a bundled, vertically integrated product that effectively serves as a moat. This creates a risk of "optimization lock-in," where the harness performs best only on the provider's own weights. The open-source community is responding by developing neutral frameworks and standards, such as the Model Context Protocol, to ensure that the agentic layer remains interchangeable. The goal is to keep the model as a commoditized, swappable component while building durable, portable systems for memory and security that belong to the user rather than the vendor.

Ultimately, the future of AI hinges on whether the community can solve the "write surface" problem, where an agent's ability to execute actions in the real world currently lacks a robust, portable security standard. With the current reliance on human oversight often failing due to consent fatigue, the next major leap in open-source AI will likely be the emergence of meta-harnesses that enforce stateful, policy-based governance. By investing in these foundational layers—memory, orchestration, and permission standards—the open-source movement can ensure that the AI ecosystem remains pluralistic, allowing builders to maintain control over their tools, costs, and data in a world where vendor-controlled off-switches are becoming an increasingly tangible threat.

Moonshot AI 推出了 Kimi K3，拥有 2.8 万亿参数，是首个达到 3T 级别并开放权重的模型，标志着重要里程碑。该模型已可通过其网站和 API 使用，开源权重计划于 2026 年 7 月 27 日发布。基准测试显示其表现令人印象深刻，常能与 Anthropic 、 OpenAI 等竞争对手的顶级产品相抗衡。 Artificial Analysis 报告称，Kimi K3 在 Elo 评分上较前代 Kimi K2.6 有显著提升，同时在定价上保持竞争力并提高了 token 效率。值得注意的是，此次发布也体现了 Moonshot AI 的战略转变：这是他们迄今为止定价最高的一款模型，价格与 Anthropic 的 Claude Sonnet 系列相当。

为评估新模型的能力，作者使用了长期沿用的个人基准测试——生成一幅 pelican 骑自行车的 SVG 。 Kimi K3 的运行凸显其对 reasoning tokens 的高度依赖，大部分输出用于内部推理。虽然生成过程非常成功，模型在图像分析方面也展现出强大的 vision 能力，但这也暴露了其密集推理带来的高昂成本。有趣的是，对 token 数量的分析显示存在一个隐藏的 system prompt；在被质询时，模型拒绝透露该提示的内容。

尽管 "pelican benchmark" 已沿用近两年，且无法覆盖 agentic tool-calling 或 long-context reliability 等关键现代需求，作者依然将其视为实用的非正式 "hello world" 测试。这个测试促使动手实践，有助于快速评估新模型的成本、几何感知能力和基本指令遵循情况。通过在不同版本中持续运行相同提示词，作者可以衡量模型家族的进展，并验证诸如 LLM CLI 等工具与流程是否已正确接入最新更新。

总之，pelican 测试虽非衡量 AI 专业效用的科学手段，但为对新发布进行初步审计提供了一种一致、透明且沿袭已久的方法。该练习能揭示模型的具体行为特征，例如在输入 token 与 reasoning tokens 之间如何权衡，并留下一件有形的产物，证明模型已被充分检验。对作者而言，它仍是快速发展的 AI 领域中重要的探索手段，在快速获得可操作洞见与维持可靠、可重复标准之间取得了平衡。 Moonshot AI has introduced Kimi K3, a model boasting 2.8 trillion parameters, marking a significant milestone as the first 3T-class open-weights model. Currently available through their website and API, with an open-weights release scheduled for July 27, 2026, the model demonstrates impressive benchmark performance, frequently rivaling top-tier offerings from competitors like Anthropic and OpenAI. Artificial Analysis reports that Kimi K3 shows a substantial improvement in Elo ratings compared to its predecessor, Kimi K2.6, while maintaining competitive pricing and improved token efficiency. Notably, this release represents a shift in strategy for Moonshot AI, as it is their most expensive model to date, priced on par with Anthropic's Claude Sonnet series.

To evaluate the new model's capabilities, the author employed a long-standing personal benchmark: generating an SVG of a pelican riding a bicycle. The Kimi K3 execution highlighted the model's heavy reliance on reasoning tokens, with a significant portion of its output dedicated to internal logic. While the generation process proved successful and even showcased capable vision features through image analysis, it also underscored the current high cost associated with the model's intensive reasoning processes. Interestingly, analysis of the token count suggests the presence of a hidden system prompt that the model remains protective of, refusing to disclose its contents when challenged.

Despite the "pelican benchmark" being nearly two years old and failing to address critical modern needs like agentic tool-calling or long-context reliability, the author maintains its utility as a informal "hello world" test. The practice serves as a forcing function for getting hands-on experience with new models, facilitating a quick assessment of cost, geometric awareness, and basic instruction following. By consistently running this specific prompt across various releases, the author can gauge progress within model families and verify that tools and pipelines, such as the LLM CLI, are correctly integrated with the latest updates.

Ultimately, while the pelican test is not a scientific measure of an AI's professional utility, it provides a consistent, transparent, and tradition-bound way to perform an initial audit of a new release. The exercise reveals specific behavioral characteristics, such as how a model balances input versus reasoning tokens, and provides a tangible artifact that demonstrates the model has been thoroughly vetted. For the author, it remains a valuable component of the discovery process in an rapidly evolving AI landscape, balancing the need for quick, actionable insights with a reliable, repeatable standard.

研究人员在太空探索上取得了重要突破：他们在一颗位于遥远恒星宜居带的岩石类类地行星周围探测到了大气层。该发现发表在 Science 期刊上，标志着科学家首次在太阳系外这类行星上成功检测到大气。该行星为 LHS 1140 b，绕行一颗距离地球约 48 光年的小型、低温红矮星。

虽然发现了大气，但团队识别出的主要成分是氦气。由于氦气本身不能维持生命，这并不意味着该行星存在生命。不过，科学家仍然乐观，认为在大气更深层可能存在其他有利于生命的气体，而这些层次尚未被充分探测和表征。

这一发现的重要性在于它贴近人类寻找太阳系外生命的目标。要具备潜在的生命承载力，行星必须处于 Goldilocks zone，即既不过热也不过冷、可能拥有液态水的轨道范围。虽然在这些宜居区域已发现数百颗行星，但同时具备小型岩石构成与大气的行星仍然罕见，因此这一进展尤为关键。

来自 Harvard University 的第一作者 Dr. Collin Cherubim 将此项发现称为重大成就，认为它让人类在回答"我们在宇宙中是否孤独"这一根本问题上更进一步。来自 Harvard 的 Dr. David Charbonneau 也指出，太阳系外出现一颗带大气的类地行星，本身就是寻找外星生命拼图中的重要一块。

这项研究丰富了关于 exoplanets 的研究成果，但要确认生命仍面临诸多挑战。此前对 K2-18b 和 TRAPPIST-1 系统的研究也各有争议，关于其他行星的大气特征曾多次遭到质疑或出现数据冲突。相比之下，对 LHS 1140 b 大气的探测为未来的观测和深入分析提供了一个明确的目标。 Researchers have reached a significant milestone in space exploration by identifying an atmosphere surrounding a rocky, Earth-like planet located in the habitable zone of a distant star. This discovery, published in the journal Science, marks the first time scientists have successfully detected an atmosphere on a planet of this type outside our solar system. The planet, known as LHS 1140 b, orbits a small, cool red star approximately 48 light-years away from Earth.

While the presence of an atmosphere is a breakthrough, the specific gas identified by the research team is helium. Because helium cannot support life on its own, the finding does not confirm that the planet is inhabited. However, scientists remain optimistic that other, life-sustaining gases could exist in the deeper layers of the atmosphere, which have not yet been fully characterized.

The importance of this discovery lies in its proximity to the scientific pursuit of finding life beyond our own solar system. For a planet to potentially support life, it must reside within the Goldilocks zone, a specific orbital distance where conditions are neither too hot nor too cold, allowing for the possibility of liquid water. While hundreds of planets have been spotted in these habitable regions, finding a small, rocky world that also possesses an atmosphere is a rare and vital development.

Lead author Dr. Collin Cherubim of Harvard University described the discovery as a significant achievement, noting that it brings humanity one step closer to answering fundamental questions about whether we are alone in the universe. Dr. David Charbonneau, also of Harvard, emphasized that the mere existence of an Earth-like planet with an atmosphere outside our solar system is a major piece of the puzzle in the search for extraterrestrial life.

This research adds to a growing body of work focused on exoplanets, though the path to confirming life remains challenging. Other planets, such as K2-18b and the TRAPPIST-1 system, have been heavily scrutinized in the past, with varying results. While previous claims regarding gas signatures on other planets have been met with skepticism or conflicting data, the detection of an atmosphere on LHS 1140 b provides a concrete target for future observations and analysis.

该仪表板展示了来自为安全研究、威胁情报和教学用途部署的 SSH 蜜罐的实时遥测数据。通过捕获入站连接尝试，系统记录了多种信息，包括源 IP 、尝试的用户名与密码、执行的命令以及客户端指纹等，为观察未经授权实体探测网络基础设施的手法提供了实时视角。

观测到的流量来源多样：受损主机、代理服务器、 VPN 、扫描器、云实例以及僵尸网络节点等。需要注意的是，源 IP 并不一定能识别发起攻击的个人，而只是表明用于执行这些自动化探测的基础设施。捕获的数据常包含恶意内容，如未授权命令、潜在的恶意软件投放尝试和不可信的凭据。

由于这些数据是原始攻击记录，不应被视为已证实的归因或可安全运行的代码。该仪表板主要作为透明度工具，用于观察常见攻击模式，例如使用不同凭据反复尝试登录。在连接成功的情况下，蜜罐还会记录后续交互，包括命令输入和文件下载，从而揭示攻击者常用的自动化工作流程。

我们建议用户和研究人员对所有显示的信息保持必要的谨慎。如任何具体数据点涉及隐私、安全或滥用问题，应向站点运营方报告以便审查并在必要时删除。通过持续监控和汇总这些事件，该项目有助于分析人员更好地了解威胁态势及针对开放 SSH 服务的自动化扫描的持续性。 This dashboard provides live telemetry from an SSH honeypot designed for security research, threat intelligence, and educational purposes. By capturing inbound connection attempts, the system logs a variety of data, including source IP addresses, attempted usernames and passwords, executed commands, and specific client fingerprints. This information offers a real-time window into the tactics used by unauthorized entities to probe network infrastructure.

The observed traffic originates from a diverse range of sources, including compromised hosts, proxy servers, VPNs, scanners, cloud instances, and botnet nodes. It is important to note that a source IP address does not necessarily identify the individual behind an attack, but rather highlights the infrastructure used to carry out these automated probes. The captured data frequently includes malicious content, such as unauthorized commands, potential malware delivery attempts, and untrusted credentials.

Because this data is raw and reflects actual attack attempts, it should not be considered verified attribution or safe-to-run code. The dashboard functions as a transparency tool for observing common attack patterns, such as repeated login failures with varying credentials. In instances where a connection is successful, the honeypot records subsequent interactions, including command inputs and file downloads, providing insight into the automated workflows typically deployed by attackers.

Users and researchers are encouraged to treat all displayed information with appropriate caution. The platform emphasizes that if any specific data points reveal privacy, security, or abuse concerns, they should be reported to the site operator for review and potential removal. By continuously monitoring and aggregating these events, the project helps analysts better understand the threat landscape and the persistent nature of automated scanning activity against open SSH services.

当顾问被引入一个组织时，通常是因为某个问题已经严重到值得投入成本和精力去解决。虽然解决问题是显而易见的目标，但人们面对挑战时往往会采取并不直接解决问题的应对方式。识别这些替代行为对任何试图推动组织变革的人都至关重要，因为这些模式并不一定出于恶意，而是系统运作的常态。

第一种常见的反应是把问题四处推诿。在许多企业环境中，这表现为局部优化：在某一环节改进流程，却在其他地方带来同等的或相应的问题。与其责怪那些通常只是在各自部门约束下争取利益的个人，不如把目光投向高层领导，调整激励机制并塑造更广阔的系统视角，这往往更为有效。

第二种反应是让问题得以保留，这就是所谓的 Shirky 原则。机构常常与其要解决的问题纠缠太深，以至于无意中维持这些问题来保障自身的存在。要想应对这一点，必须找出谁在从现状中获益。即便你不同意那些依赖问题持续存在的人，承认他们的利益也是制定有效策略时不可或缺的一环。

最后一种是引入新问题。每一个解决方案都会带来一系列后果，说明当你解决了当前的问题，下一个问题往往会随之而来。这种循环并不是要你放弃解决问题，而是要摒弃"工作终将完全结束"的幻觉。一位有经验的顾问明白，尽管解决问题能带来更好的境况，但在长期成功中，偶尔选择忽略某些问题的能力同样重要。 When consultants are brought into an organization, it is typically because a problem has become so burdensome that it is worth the cost and effort to resolve. While problem-solving is the obvious goal, people often respond to challenges in ways that do not involve fixing them. Recognizing these alternative behaviors is essential for anyone trying to navigate organizational change, as these patterns are not necessarily signs of malice, but rather the reality of how systems function.

The first common response is pushing problems around. In many corporate environments, this manifests as local optimization, where a process is improved in one area only to create a corresponding, often equivalent, issue elsewhere. Rather than blaming individuals, who are usually just trying to win the game within their own departmental constraints, it is more effective to look toward senior leadership to adjust the incentives and the broader system view.

The second response is the preservation of problems, a phenomenon known as the Shirky Principle. Institutions often become so deeply tied to the problems they were designed to solve that they inadvertently perpetuate them to ensure their own survival. Successfully managing this requires identifying who benefits from the status quo. Even if you do not agree with those who rely on the persistence of a problem, acknowledging their interests is a necessary part of any effective strategy.

Finally, there is the act of promoting new problems. Every solution introduces its own set of consequences, echoing the idea that once you resolve your primary issue, the next one simply takes its place. This cycle does not mean one should stop trying to solve problems, but rather that one must let go of the illusion that the work will ever be finished. A skilled consultant understands that while fixing issues leads to a better life, the ability to occasionally ignore problems is equally vital to long-term success.

Lisp 不是一种单一的语言，而是由多个编程方言组成的家族——它们共享基本语法，但在运算符、语义和标准库上差异很大。对初学者而言，选择哪个方言并不如掌握 Lisp 那种思维方式重要。因为核心概念在各方言间基本一致，学会一种后转到另一种通常相对容易。

Common Lisp 是最成熟、最全面也最稳定的方言。自 1994 年标准化以来仍广泛适用，几十年前的代码常常能在现代实现（如 SBCL）上直接运行。它开箱即用，功能强大，包括用于调试的 condition/restart 机制和支持多重分派的对象系统 CLOS 。虽然缺少诸如内置模式匹配的现代便捷功能，但能编译为原生代码，使其速度非常快，适合长期运行的服务、研究以及高性能应用。

Clojure 的设计旨在把 Lisp 带入现代生态：运行于 JVM，使开发者能利用现有的 Java 库和生态。它强调函数式编程风格，依靠纯函数和不可变数据结构来管理状态。凭借对并发的良好支持和全栈能力（包括用于浏览器的 ClojureScript），Clojure 可以说是最适合生产环境的 Lisp，广泛应用于金融和数据密集型领域。调试有时会被复杂的 JVM 堆栈跟踪所困扰，但其对简洁性和高产出的追求使其非常实用。

Racket 源自 Scheme 传统，更像是一个面向语言的开发环境。它允许开发者轻松创建新语言或领域专用语言，因此在学术研究和计算机科学教育中备受青睐。 Racket 附带丰富的库，用于构建 GUI 、 Web 服务器等，DrRacket IDE 为初学者提供了友好的入门体验。尽管在工业界的影响力可能不如 Clojure，但其卫生宏系统以及通过 Typed Racket 提供的可选静态类型支持，使其成为进行实验性编程的有力工具。

除了这些通用方言之外，Elisp 作为 Emacs 的扩展语言，承担着专业且非常实用的角色。它允许对编辑器进行实时定制，生动展示了基于 Lisp 的环境如何在运行时被操控。无论选择哪种方言，各有优势：Common Lisp 的原生性能、 Clojure 的现代生态、 Racket 的教学与实验性灵活性。最终该如何选择，取决于你的目标——是追求职业实用、深入系统理解，还是入门语言设计。 Lisp is not a single language but a diverse family of programming dialects that share fundamental syntax while differing significantly in their operators, semantics, and standard libraries. For beginners, the choice of dialect is less critical than the shift in thinking required to master Lisp. Because the core concepts remain consistent across the family, learning one makes it relatively easy to transition to another later on.

Common Lisp stands out as the most mature, comprehensive, and stable dialect. Standardized in 1994, it remains highly relevant today, with code written decades ago often running perfectly on modern implementations like SBCL. It offers a massive feature set out of the box, including powerful condition and restart systems for debugging and an object system called CLOS that supports multiple dispatch. While it lacks some modern conveniences like built-in pattern matching, its ability to compile to native code makes it exceptionally fast and suitable for long-running processes, research, and high-performance applications.

Clojure was designed to bring Lisp into the modern era by targeting the JVM, allowing developers to leverage existing Java libraries and ecosystems. It emphasizes a functional programming style, utilizing pure functions and immutable data structures to manage state effectively. Clojure is arguably the most production-ready Lisp, widely used in finance and data-heavy industries for its concurrency support and full-stack capabilities, including ClojureScript for browser development. While debugging can sometimes be hampered by complex JVM stack traces, its focus on simplicity and productivity makes it a highly practical choice.

Racket, which descended from the Scheme tradition, is uniquely defined as a language-oriented environment. It allows developers to easily create entirely new languages and domain-specific languages, making it a favorite in academic research and computer science education. It comes with a robust set of batteries-included libraries for tasks like building GUIs and web servers, and its DrRacket IDE provides a friendly entry point for beginners. Though it may not have the same industry footprint as Clojure, its hygienic macro system and support for optional static typing through Typed Racket make it a powerful tool for experimental programming.

Beyond these general-purpose dialects, Elisp serves a specialized but highly practical role as the extension language for the Emacs editor. It allows for the real-time customization of the editor itself, providing a unique look at how a Lisp-based environment can be manipulated while it runs. Regardless of which dialect you choose, each offers a distinct set of advantages, from the native performance of Common Lisp to the modern ecosystem of Clojure or the educational versatility of Racket. Ultimately, the best path forward depends on your specific goals, whether you are looking for professional utility, deep system insight, or an introduction to language design.

Apple 最近向 OpenAI 数十名在职员工发出法律信函，将双方的竞争进一步升级。这一事态表明，随着各大科技公司在迅速发展的 AI 领域争夺顶尖人才和知识产权，业内竞争正日益激烈。

此举正值 AI 行业对劳动实践与人才招揽审查日趋严格之际。通过向特定个人发出法律函件，Apple 旨在保护自身利益，同时加快其 AI 研发步伐。随着像 OpenAI 这样的实验室不断推出快速创新，老牌科技公司为跟上节奏而采取此类行动也变得愈加普遍。

业内人士认为，这些法律手段反映了生成式 AI 领域更广泛的主导权之争。 OpenAI 在面向公众的应用方面持续领先，其竞争对手则在想方设法壮大内部团队，以缓解高速增长的科技环境中常见的人才流失。

此事清楚表明，各公司在这场持续的 AI 主导权竞赛中，为保持领先地位承受着巨大的压力。 Apple has recently escalated its competition with OpenAI by issuing legal letters to dozens of current employees at the artificial intelligence firm. This development signals a tightening of the professional rivalry as major technology companies scramble to secure top-tier talent and intellectual property in the rapidly evolving AI sector.

The move comes at a time of heightened scrutiny regarding labor practices and talent acquisition within the AI industry. By targeting specific individuals with legal communication, Apple is positioning itself to protect its interests as it intensifies its own AI development efforts. Such actions are becoming increasingly common as established tech giants work to keep pace with the swift innovations coming out of labs like OpenAI.

Industry observers suggest that these legal maneuvers reflect a broader battle for dominance in the generative AI space. As OpenAI continues to lead in public-facing applications, its competitors are looking for ways to bolster their own internal teams and mitigate the impact of the brain drain that often characterizes high-growth tech environments. The situation remains a clear sign of the immense pressure companies face to maintain their edge in the ongoing race for AI supremacy.

Measuring Progress Toward AGI: Cognitive Abilities hackathon 最近落下帷幕，Google DeepMind 向开发出旨在测试人工智能超越简单信息检索能力的创新基准的团队颁发了奖项。比赛聚焦五大认知方向：执行功能、学习、元认知、社会认知和注意力，其目标是超越传统静态数据集，评估前沿模型在动态或新颖环境中的推理、行动与判断能力。

四个大奖项目脱颖而出。 MEDLEY-BENCH 通过引入社会压力来测试模型是否能识别自身不确定性，探讨行为层面的元认知。 LearningBench 专注于上下文学习，要求模型从零开始掌握基于文本的新系统。 GAUGE 利用元认知阶梯测试模型的置信度与弃权决策，以评估知识与行为之间的差距。最后，Metaproteus 通过让模型预测自身输出分布与采样倾向来考察其自知能力。

此外，各认知方向还评出了专项奖项，比如用于评估执行功能的 Turn Bench 与 SecureExec-Bench，以及用于创意学习评估的 GrammarGym 和 EphLangBench，它们通过合成或短时语言防止模型记忆化。其他获奖项目关注元认知灵活性、社会认知与选择性注意机制，凸显认知能力通常依赖情境而非一成不变。

公告发布后，部分参赛者对评估流程的透明度表示强烈不满，指出缺乏完整排行榜和详细评分披露，阻碍团队从结果中学习。一些人质疑大奖评选的有效性，认为获奖方法存在逻辑漏洞且结果不够公开，评审标准与最终结果之间可能存在偏差。

批评者还指出某些获胜方案在数据解读与可重复性方面的问题，要求组织方公开评分数据并说明为何选择特定基准而非那些宣称具有更高技术可扩展性或更贴合竞赛宗旨的方案。他们强调，既然比赛以衡量认知能力为核心，评估过程本身应保持高度可读性与学术诚实。

讨论也延伸到人工通用智能的更广范畴。一些参赛者认为当前评估框架过于偏重文本与符号推理，建议未来应将物理智能纳入考量，如力、扭矩与材料交互等，以更全面地覆盖具身智能。尽管结果存在争议，许多参赛者仍认可此次比赛在推动更严谨、更有洞察力的基准设计方面的价值。 The Measuring Progress Toward AGI: Cognitive Abilities hackathon recently concluded, with Google DeepMind awarding prizes to teams that developed innovative benchmarks designed to test artificial intelligence beyond mere information recall. The competition focused on five distinct cognitive tracks. Executive functions, learning, metacognition, social cognition, and attention. The goal was to move past traditional static datasets and instead evaluate how frontier models reason, act, and make judgments in dynamic or novel environments.

Four grand prize winners were selected for their contributions. MEDLEY-BENCH explores behavioral metacognition by introducing social pressure to test if models can identify their own uncertainty. LearningBench focuses on in-context learning by forcing models to master novel, text-based systems from scratch. GAUGE evaluates the gap between model knowledge and action by utilizing a metacognitive staircase to test confidence and the decision to abstain. Finally, Metaproteus examines a model's self-knowledge by testing its ability to predict its own output distributions and sampling tendencies.

In addition to the grand prize winners, track-specific prizes were awarded to projects addressing a variety of cognitive challenges. These included benchmarks for executive function, such as Turn Bench and SecureExec-Bench, as well as creative learning assessments like GrammarGym and EphLangBench, which use synthetic or ephemeral languages to prevent memorization. Other notable winners addressed metacognitive flexibility, social cognition, and selective attention mechanisms, highlighting that cognitive capabilities are often context-dependent rather than monolithic.

Following the announcement, a segment of the participant community expressed significant frustration regarding the transparency of the evaluation process. Several entrants raised concerns about the lack of a comprehensive leaderboard or detailed scoring disclosures, which they argued hindered the ability of teams to learn from the results. Some participants specifically challenged the validity of the grand prize selections, citing perceived logical flaws in the winners' methodologies, opaque results, and a potential misalignment between the stated judging criteria and the final outcomes.

Critics specifically pointed to discrepancies in data interpretation and reproducibility within certain winning writeups. There was a notable demand for the organizers to release scoring data and provide clearer justifications for why specific benchmarks were chosen over others that claimed higher technical scalability or adherence to the competition's core mandates. These participants emphasized that in a competition centered on measuring cognitive abilities, the evaluation process itself should be held to high standards of legibility and intellectual honesty.

The discussion also touched on the broader scope of Artificial General Intelligence, with some participants noting that the current evaluation frameworks are heavily weighted toward textual and symbolic reasoning. There were suggestions that future iterations of such challenges should expand to include physical intelligence, such as force, torque, and material interaction, to better capture the full spectrum of embodied intelligence. Despite the controversy over the results, many participants acknowledged the value of the competition in pushing the field toward more rigorous and insightful benchmark design.

为了弄清人类如何在复杂听觉环境中定位目标，研究者最近考察了大脑在竞争性语音流之间切换注意力的能力。研究在多说话人场景中对参与者进行 EEG 记录，监测他们在两位说话者之间转移注意力时的大脑反应。结果表明，注意力切换既非瞬时完成，也不是干净利落的过渡；大脑表现出一种明显的不对称性：在完全脱离前一语音流之前，就已经开始对新的目标语音流产生响应。这就产生了一个短暂的过渡期，此时两条语音流会同时在皮层中被表征。

研究还通过追踪α频段活动，将这些神经动态与认知负荷联系起来。结果发现，注意力切换后 EEG 的 alpha 波功率显著下降，且当对新目标的完全投入建立起来时，alpha 功率降至最低。这说明在重新定向注意力的过程中，大脑需要投入大量听觉努力，而只有在成功跟踪到新说话者后，这种努力才开始减弱。

除了时间维度之外，研究还探讨了大脑在切换过程中如何处理语言语境。研究者利用大型语言模型比较了不同的语境累积策略，检验大脑是保留过去的信息还是重置语言预期。结果支持重置机制，表明在切换注意力时，听者会清除先前的词汇语境，以优先处理新说话者的语言线索。这暗示大脑有一套动态且灵活的语义先验更新机制，有助于在变化的环境中更好地理解语义信息。

总体而言，这项工作更清晰地描绘了大脑在同时应对多个声音源时如何在稳定与灵活之间取得平衡。通过将注意力切换分解为独立的投入与脱离阶段，研究提出了理解注意力神经生理学的新框架。这些发现不仅阐明了听觉感知的基本机制，也对开发在嘈杂现实社交环境中更好支持用户的先进助听技术具有实际意义。 To understand how humans navigate complex auditory environments, researchers recently investigated the brain's ability to switch attention between competing speech streams. Using EEG recordings, the study monitored participants as they shifted focus between two different talkers in a multi-talker scenario. The results confirm that attention switching is not a clean, instantaneous transition. Instead, the brain exhibits a distinct asymmetry, where it begins engaging with a new target stream before it has finished disengaging from the previous one. This creates a brief, transient period where both speech streams are simultaneously represented in the human cortex.

The research also linked these neural dynamics to cognitive effort by tracking alpha-band activity in the brain. It was observed that EEG alpha power drops significantly after an attention switch, with the minimum level of alpha power occurring precisely when engagement with the new target is fully established. This suggests that the brain exerts substantial listening effort during the reorientation process, which begins to subside only once the new speaker is successfully tracked.

Beyond temporal dynamics, the study explored how the brain manages linguistic context during these shifts. By comparing different context-accumulation strategies using Large Language Models, the researchers tested whether the brain maintains past information or resets its linguistic expectations. The findings favor a reset mechanism, suggesting that upon switching attention, listeners essentially clear their previous lexical context to prioritize the new speaker. This points to a dynamic, flexible system for updating semantic priors, which facilitates better comprehension in changing environments.

Ultimately, this work provides a clearer picture of how our brains maintain both stability and flexibility when juggling multiple audio sources. By dissecting the process into discrete engagement and disengagement components, the study offers a new framework for understanding the neurophysiology of attention. These findings not only illuminate the basic mechanics of auditory perception but also have practical implications for developing advanced hearing technologies that can better support individuals in noisy, real-world social settings.

Pebble 在完成 Pebble Time 2 预购订单方面取得了显著进展，自 3 月以来已生产超过 23,000 台。交付已完成超过 80%，公司预计到 7 月底所有颜色款式都将有现货。随着团队逐步摆脱对预购的依赖，他们也在扩展配件线，包括充电器和即将推出的表带。

软件开发团队产出颇丰，重点放在功耗优化上。 Pebble 2 Duo 的续航已提升到 30 天以上，Pebble Time 2 目前平均约 21 天。开发者发布了多项重要 API 更新，加入了对触摸屏、扬声器和 RGB 背光的支持。 Index 01 功能已完整集成到移动应用中，为用户提供开源的任务管理、日历同步和云端加密备份工具。接下来路线图将优先通过切换到新的蓝牙通信协议来改善 iOS 通知处理，最终实现对通知的直接回复。

针对硬件反馈，Pebble 保持透明并积极应对制造缺陷。尽管量产仍然靠人工，难免会出现个别问题，公司承诺对报告的缺陷（如屏幕裂纹、按键故障或高耗电）提供免费更换，无论是否在保修期内。组装线上已实施更严格的测试流程，以尽量减少此类问题。

Pebble Round 2 在解决了不锈钢表壳的外观制造问题后，已进入下一发布阶段。 Beta 测试正在进行，量产计划于 7 月最后一周启动。公司目标是在 9 月底前完成全部 14,000 份现有预购订单，并将在发货前两周通过电子邮件通知客户，确认信息并最终确定配件选择。

最后，Index 01 ring 已正式进入量产，预计大部分预购将在 8 月底前发货。鉴于用户反馈戒指可能偏紧，公司强烈建议客户重新确认尺寸；如果不确定，建议选大一号，因为调整偏松的戒指比扩大太小的戒指要容易得多。 Pebble is making significant progress in fulfilling pre-orders for the Pebble Time 2, having manufactured over 23,000 units since March. With fulfillment now more than 80 percent complete, the company expects to reach in-stock status for all color variants by the end of July. As the team transitions away from pre-order dependency, they are also expanding their accessory offerings, including chargers and upcoming strap options.

The software development team has been highly productive, focusing heavily on power optimization. Battery life for the Pebble 2 Duo has surged to over 30 days, while the Pebble Time 2 currently averages 21 days. Developers have released several key API updates, including support for touch screens, speakers, and RGB backlights. Furthermore, the Index 01 feature is now fully integrated into the mobile app, providing users with open-source tools for task management, calendar syncing, and cloud-encrypted data backup. Looking ahead, the roadmap prioritizes improving iOS notification handling by transitioning to a new Bluetooth communication protocol, which will eventually enable direct replies to notifications.

Addressing hardware feedback, Pebble maintains a transparent and proactive stance regarding manufacturing imperfections. While mass production remains a human-intensive endeavor prone to occasional errors, the company is committed to providing free replacements for reported defects, such as screen cracks, button issues, or high power consumption, regardless of warranty status. Stringent new testing procedures on the assembly line have been implemented to mitigate these issues as production continues.

The Pebble Round 2 is entering the next phase of its launch following the resolution of a cosmetic manufacturing hurdle involving the stainless steel case. With beta testing currently underway, mass production is slated to begin in the final week of July. The company aims to fulfill all 14,000 existing pre-orders by the end of September and will notify customers via email two weeks prior to their shipping date to confirm details and finalize accessory selections.

Finally, the Index 01 ring has officially entered mass production, with shipping for the majority of pre-orders expected by the end of August. Because user feedback suggests the rings may fit slightly tighter than expected, the company is strongly encouraging customers to re-verify their sizes. If a user is uncertain, they are advised to opt for a larger size, as it is easier to adjust a loose fit than to modify a ring that is too small.