AllyshipContent Radar

News 2943

Chủ đề cộng đồng đang thảo luận

Crawl live-1795144500000 · 10:15 20 thg 11, 2026 · Xuất bản 10:32 20 thg 11, 2026

192 / 2943 chủ đề đạt ngưỡng 55 trong khoảng này.

21

Anthropic 首席执行官 Dario Amodei 在公司首次公开募股 (IPO) 之际发表的致辞:

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

dotey 0 bài viết · X
Nguyên văn nguồn
Chi tiết nguồn

Anthropic 首席执行官 Dario Amodei 在公司首次公开募股 (IPO) 之际发表的致辞: 我和联合创始人当初创立 Anthropic,是因为我们坚信:这个时代最强大的技术,应当为了全人类的福祉而打造。 我和 Daniela(注:Dario 的妹妹,也是 Anthropic 的联合创始人兼总裁)曾因病失去了父亲。令人唏嘘的是,就在他去世短短几年后,那种病变得可以治愈了。我职业生涯起步于生物学,最初就是希望能让未来的疗法早日问世。但真正走进实验室后,我才看清科学家们在攻克疾病、研发新药时,面前横亘着怎样一片庞大而错综复杂的“荆棘丛”。 许多能够劈开这片荆棘的科学工具——无论是基因编辑、基因合成,还是用光来控制脑细胞(注:指光遗传学技术,科学家借此用光精准调控神经元的活动)——最初都只是纯粹的构想。换句话说,它们诞生于思考,诞生于智能。 人类智能已经让鼠疫、天花、脊髓灰质炎等疾病在世界大部分地区成为了历史。但对于无数其他疾病的患者和家庭来说,医学的进步还是太慢了。从灵光一现的想法到真正获批的疗法,往往要耗费数十年光阴。就拿 CRISPR 来说,从科学家首次在 DNA 中发现一段奇特的重复序列,到首款基于 CRISPR 的疗法正式获批,中间整整走过了许多年。 大自然中遍布着这样等待破译的密码。过去,我们拥有的“智能”太少了,科学家太少,时间也太少,导致绝大部分奥秘都无从破解。而现在,我们可以亲手构建这种智能。无论是蛋白质折叠 (protein folding)、药物设计、癌症遗传学,还是临床试验管理,生物和医学各个领域的科学家都在发现:借助 Claude,他们如今只需花费几年前一小部分的时间,就能完成同样繁重的研究。我相信,足够强大的 AI 与人类携手合作,有望将整整一个世纪的科学与医学进步,浓缩到短短十年之内。 当然,我不想操之过急。目前我们看到的还只是早期成果,实验室研究和临床试验依然需要时间。尽管这些流程向来漫长而繁琐,但我相信 AI 能够大幅为它们提速。我们希望,未来能有越来越少的家庭,不再因为“只差那么几年”而错失治愈的希望。 医学只是一个格外生动的缩影,类似的变革正在各行各业发生。我们的客户——包括全球最具雄心的企业、非营利组织以及无数普通人——已经开始运用我们打造的智能,去拉动经济增长、帮助人们摆脱贫困、守护民主与自由社会,甚至让像 DMV 这样的日常公共服务变得更顺畅、高效。能与他们并肩同行,我们深感自豪;对于未来我们能共同成就的一切,我也充满乐观。 与此同时,我们始终坦诚:这项技术伴随着非常现实的风险。小到个人或国家层面的恶意滥用,大到经济冲击,乃至 AI 系统失控。我们反复谈论这些风险,并不是认为它们注定会发生,恰恰相反,是因为我们相信它们完全可以避免。 能否避开这些风险,取决于 AI 缔造者们的选择。如果我们对安全掉以轻心,AI 会走向一个方向;如果我们对安全倾注近乎偏执的专注,AI 则会走向截然不同的未来——两者之间有着天壤之别。而对安全与公共利益的执着,从第一天起就深深烙印在了 Anthropic 的基因里。从防范生物与网络安全风险的护栏,到对齐训练与测试,再到像拍 X 光一样透视模型内部,Anthropic 开创了一系列理解和引导 AI 系统行为的方法。 前路依然漫长,最棘手的安全挑战还在后头。正是公司里这群才华横溢的伙伴,让我们打造出了最强大的 AI 系统;而接下来,我们需要比以往更加倾注这份才华,去反复核查、层层把关,确保这些系统真正安全可靠。 在此,我想毫不含糊地表明立场:我们将主动把控 AI 系统的训练与部署节奏,无论需要耗费多少时间,都要先确保它们安全无虞,且始终处于人类的掌控之下。我们绝不会向任何商业压力或政治压力低头。今天的上市让公司拥有了更雄厚的财务实力,而这只会让我们坚持做正确事情的决心更加坚定。 如果我们能向世界证明,在安全问题上做出正确选择,同样能打造出极其成功的企业,那么整个行业的其他参与者也会受到鼓舞,纷纷效仿。我们称之为“向上竞速”。当这种良性竞争真正运转起来时,究竟哪家公司的模型最强反倒没那么重要了——因为所有人都在安全的前提下构建技术,最终每个人都是赢家。要引领这场“向上竞速”,最好的位置就是跑在队伍最前面,我们也满怀希望并期待着扮演领跑者的角色。但无论 Anthropic 在某个时刻是领先还是落后,我们的终极目标始终是让整个行业变得更好。 机器智能的降临,是人类历史上最具里程碑意义的时刻之一,我们正手握着一次千载难逢的机遇,去将它引向更美好的方向。 这一刻的意义,超越了任何一家公司,也超越了任何一个国家。唯有我们所有人携手努力,才能确保这项技术造福于我们自己,以及未来的千秋万代。从今天起,我们热切期盼着你的加入,与我们共同投身这项使命。 Media

22

We had the best time with @friendliai and all the developers that came out to the meetup in Seoul! Thank you all for coming!

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

Ollama 0 bài viết · X
Nguyên văn nguồn
Chi tiết nguồn

We had the best time with @friendliai and all the developers that came out to the meetup in Seoul! Thank you all for coming! Media

23

langchain-openai==1.8.5

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · langchain-ai/langchain releases
Nguyên văn nguồn
Chi tiết nguồn

<p>Changes since langchain-openai==1.8.4</p> <p>release(openai): 1.8.5 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6500840983" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41969" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41969/hovercard" href="https://github.com/langchain-ai/langchain/pull/41969">#41969</a>)<br> fix(openai): preserve streamed Responses tool call arguments (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6499500065" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41963" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41963/hovercard" href="https://github.com/langchain-ai/langchain/pull/41963">#41963</a>)<br> feat(core): standardize finish reason on message-finish events (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6458937894" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41926" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41926/hovercard" href="https://github.com/langchain-ai/langchain/pull/41926">#41926</a>)</p>

24

Claude Code 团队是如何做代码审查的?

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

dotey 0 bài viết · X
Nguyên văn nguồn
Chi tiết nguồn

Claude Code 团队是如何做代码审查的? https://claude.dev/blog/solving-code-review-at-anthropic/ “Clark,你给我的 PR 盖章,我就给你的盖章。” 这是半年多前 Anthropic 内部的日常。Claude 写代码越来越快,等人审的改动排成长队,大家只好互相放行,看是谁提交的,扫一眼就批准。 Claude Code 团队的产品经理 Morgan Lunt 在这场分享里把那段时间叫作“歌舞伎表演”,审查的形式还在,其实大家只看了个大概。 后来他们让 AI 智能体接手了绝大部分审查,人只在少数高风险改动上签字。目标是 90% 的 PR 合并时,除了让 Claude 干活的那个人,不需要再找任何人。 Lunt 给出的结果是:改用自动审查后,查出实质问题的 PR 比例从 16% 升到 54%,合并的 PR 数量又翻了一倍,需要回滚的 PR 少了 40%。在他看来,这和自动驾驶是一个道理,用不着完美,比人强就够了。 瓶颈从写代码转到了审代码 PR(拉取请求)是程序员提交一份代码改动、请别人检查后再合并进主代码的流程。Lunt 说,一年前 Anthropic 和别的公司一样,几乎每个 PR 都由人来审,代码产出的速度就是人写代码的速度。 大约九个月前,模型强到可以独立做完一个功能,大家开始直接把任务丢给 Claude,产出一下子多了很多,审查就成了最慢的环节。他引用了一项哈佛的研究:企业引入 AI 编程智能体后,PR 等待审查的时间平均增加了 56%。他读那篇论文的理解是,PR 太多,人被淹没了。 Anthropic 的工程师擅长做智能体,于是很多人各自写了审查机器人,把自己审代码时关注的点写进去。结果一个 PR 上会出现七个机器人,意见互相冲突,一个机器人要的改动会惹另一个机器人不满,来回改出五十、一百甚至两百次提交,改到最后已经偏离了作者原本想做的事。 问答中他解释,冲突的原因是各个机器人关注的方向不同,有的对接口怎么设计要求很细,有的不管,而它们之间没有共享的上下文和共同的目标。他们后来把这些机器人的关注点合并进一个主审查智能体,由它在内部去重、整理。 大约今年 7 月,Claude Code 的创建者 Boris Cherny 找到 Lunt 的团队,让他们“把 Anthropic 的代码审查问题解决掉”。 团队给自己定的目标就是开头说的 90%,并按需要几个人参与把 PR 分成三类:零人参与,是 Claude 自己发现问题、自己修、自己提交、自己盯到合并;一人参与,是某个人让 Claude 做了一件事,自动系统判断没问题就直接合并,不用再找人;两人及以上,留给账单、用户登录验证、应用自动更新这类绝对不能出错的地方,而且范围尽量收窄。 同时有一条底线:回滚、紧急补救和线上事故都不能因此变多。 用自己出过的事故来检验审查 要知道审查智能体好不好,得有一套考题。他们把 Anthropic 历史上所有线上事故,以及被认定引发事故的那些 PR 整理成评测集,再请人抽查核对,没有用人为埋了 bug 的合成数据。这样考的都是在自家代码上真实出过问题的情况。 每道题的格式很简单。Lunt 现场展示了一个例子:某个 PR 合并后有客户收到了两次同一张发票,只好回滚,原因定位在某个文件的第 88 行。这道题记录的就是代码仓库在那个时间点的版本号、PR 编号、出问题的文件和行号,以及正确答案。成千上万道这样的题,就是评分标准。审查智能体被放回那个时间点去审那个 PR,看它能不能找到那一行。 衡量时看两个数。一个是召回率:标准答案里列出的问题,它找到了多少。另一个是精确率:它报出的问题里,有多少是标准答案之外、大概率不该报的。只看召回率不行,因为让智能体随便报五十万条,迟早会撞中所有答案,但真正有用的提醒就淹没在噪音里了。 在这套考题上,他们试了不同的提示词,不同的检查角度(比如专门找某类安全漏洞,或者专门找“同一个操作执行两次会不会出错”这类问题),以及不同的智能体分工结构,跑了几万到几十万次。方法就是反复试:改一处,成绩变好就留下,变差就扔掉。Lunt 说 Claude 让同时跑大量实验变得很容易。 最后定下来的流程是:先派出一批负责找问题的智能体,再派出一批负责“挑刺”的验证智能体,专门质疑前者的发现,确认是真问题;然后按严重程度排序、去重,作为评论贴到 PR 上。发现分三档:重要问题会阻止合并;小问题(nit)不阻止,留着以后想集中清理时再翻;还有一档是改动之前就存在的老问题。完全无关紧要的,在验证阶段就被删掉。 让找问题的智能体适可而止 Lunt 说他为“收敛”这件事钻研了三个星期。智能体被要求找问题就一定会找到问题,被要求修问题就一定会去修。让一个智能体找问题、另一个修,再问前一个“现在怎么样了”,它会接着找出越来越偏门、越来越鸡毛蒜皮的毛病。有段时间,智能体整夜来回找、修,一个 PR 上积累几十上百次提交,偏离用户的原意,也没人看得懂。 他们的调法是让审查智能体一次把所有发现都报出来,修完之后不再越挖越深。他说现在九成多的 PR 在三轮“审查、修复”之内就能收敛,但为此在提示词上花了不少功夫。 哪些改动还需要人看 审查之后还有一步风险分级,由一个机器人根据 PR 改了哪些文件、改了多少来决定要几个人批准。只改文档、影响面小的,直接通过;普通改动,让 Claude 干活的那个人负责就够了;涉及账单、自动更新这类高风险部分,要第二个人签字。 他们最早用的是“代码负责人”配置文件,按文件路径规定某个目录下的改动需要一个人还是两个人批准。Lunt 说这个办法太粗,后来又加了一层由大模型做的判断。比如账单相关的代码,如果改动涉及业务逻辑,就要另一个人签字,如果不涉及就直接放行。 他演示的三个例子是:修改安装指南里的错别字,不需要额外审查;给订单列表加一个日期筛选,需要一个人,因为碰到了订单;改发布流程,要发布团队的人点头,因为影响面大。权限也按团队分,他在 Claude Code 代码仓库里能做的事,其他团队的人不一定能做。 整条流程他称为六层:Claude 写代码的同一个会话里先自己跑测试、修失败、再派一个上下文干净的子智能体复查;PR 提交后由托管的 Claude Code Review 做深度审查;再过一遍安全审查;然后是常规的自动检查和测试;只有真正需要的 PR 才找人批准;最后合并。 他把这叫作“瑞士奶酪”式的防护,每一层都有漏洞,叠在一起就很难漏过去。此外,Anthropic 的产品大多通过网页和自己控制的 App 发布,新功能上线默认是关着的,先开给少量内部测试者,再开给更多外部测试者,最后才全面开放。出了问题,几分钟内就能对所有人关掉,不用改代码。 审查发现问题之后,修复也要自动化。Lunt 说内部有人用 Claude Tag,有人自己写了“盯 PR”的技能,让写这个 PR 的 Claude 会话定时去 GitHub 看有没有审查意见要处理。他的理由是,修复如果还要等人醒着、有空,提速就无从谈起。 问答中,团队的工程师把审查要回答的问题分成三层:最底层是代码风格和整洁程度;中间一层是假设这个改动的方向是对的,它写得对不对、安不安全;最上层是这个功能该不该做、方向对不对。自动审查擅长的是中间这层,最上层留给人。Lunt 补充说,他们刻意不让审查智能体管得太宽,不会因为它觉得不妥就拦住你明确要做的事,所以防止有人往代码库里塞恶意改动,仍然要靠团队自己。 普通团队可以怎么照着做 Lunt 建议想引入自动审查的团队,先用自家的事故记录建一套评测集。大多数公司都有事故追踪系统,让 Claude 从里面提取被认定引发事故的 PR 编号,按上面的格式整理,他估计几天就能做完。之后可以拿它比较不同的审查产品,也可以拿它调审查规则。他说自己对 Anthropic 的产品很有信心,但不希望大家凭他的感觉选产品,更希望各家用自己的评测证明这些工具可靠,有了这个底气才能放心撤掉人工检查。 工程师补充了一个更简单的检验办法:拿以前人审过的 PR,让自动审查回到那个时间点再审一遍。人当时提了三条意见,看它是不是都找到了,或者找到了更重要的;它没找到的那些,再看是不是真的重要,还是只是小问题,或者最后证明什么事都没有。 具体调法上,用 Claude Code Review 的仓库可以在根目录放一个 REVIEW.md 文件,专门写给审查智能体看:哪些问题要重点查,哪些不用管。比如自动生成的代码不用看,代码风格已经有自动检查工具管了不用提,小问题最多报几条,发现要按什么格式写,方便交给负责修复的智能体。Lunt 现场演示了加规则前后的对比,加之前是一条重要问题加一串小问题,加之后只剩重要问题。原来的 CLAUDE.md 则继续写所有参与这个仓库的人和智能体都要遵守的规矩,比如用 Tab 还是空格、按什么原则分层。审查时也会检查改动有没有违反 CLAUDE.md 里的规矩。 规则还可以持续改进。Anthropic 内部有每周自动运行的任务,专门去翻人在 PR 上留的评论,Lunt 说现在这类评论已经不多了。如果有人回复审查意见说“这其实不是问题,因为……”,任务会从中提炼出规则补进 REVIEW.md,下次就不再犯同样的错。他说这类反驳意见是非常有用的反馈,团队偶尔抽查一下 PR、在不同意的地方留句评论就够了。 在内部推广时,Lunt 说 Anthropic 的人都看数据,他们把评测里和每个团队相关的事故单独挑出来,给那个团队看效果,大家很快就有了信心。另一个原因是很多人早就不想再维护自己的审查机器人了,交出去之后,审查出了问题直接找他们团队修就行。工程师补充说,各团队都有自己盯着的指标,比如回滚和紧急补救的次数、事故数量、稳定性有没有下降,交出审查之后这些都要看,而大家的效率普遍提高了。 审查放心之后,能做以前不敢做的事 Lunt 说,当你确信审查流程能挡住一切不合格的改动,不管进来什么改动都不怕出事,工作方式就变了。别的团队急需你代码库里的一个功能,不用等你排期,可以直接提一个 PR,在 Slack 上打声招呼,不符合你们团队标准的改动反正合并不进去。 他自己想验证一个功能怎么呈现给用户更好时,会一次做出五个版本,在大量用户中做 A/B 测试,按数据选一个上线。Boris 还设了每天自动运行的任务,持续尝试缩短 Claude Code 命令行工具的启动时间,测出来确实变快、又没有拖累其他指标的改动就合并。 Claude Code Review 本身在今年 3 月开放公开预览,10 月中旬正式发布,正式版是在内部这套系统的基础上重写的,外部用户用到的和 Anthropic 内部用的基本是同一套。预览时审一个 PR 大约 25 美元,正式版按审查深度分三档,每次 5 到 25 美元。 在命令行和桌面 App 里用的 /code-review 是快速版,因为边写边审的人不想等十几二十分钟;PR 提交后自动触发的托管审查才是深度版,想在本地跑深度版可以用 ultra review。审查默认先只看改动本身,需要时再去看仓库里的其他代码;目前只能访问单个仓库,跨仓库的权限还在和隐私团队商量。 工程师说,他们试过给审查智能体更多工具、更多上下文、让它实际运行代码,大多没有带来提升,反而增加了成本、时间和复杂度。至于用哪个模型、什么推理强度,Lunt 说每出一个新模型答案都会变,现在内部的经验是 Opus 5.5 中等强度用起来性价比最高,团队会持续评测、调好默认配置。 他把在 Claude Code 团队工作比作在海边堆沙堡,做好的辅助结构会被下一代模型的能力冲掉,所以他不意外代码审查在不久后变成一个完全解决了的问题。分享开头的现场投票里,77% 的听众说审查是整个开发流程里最痛苦的一环。 Media

25

Enable AI agents to pay for AWS Marketplace x402 products with Amazon Bedrock AgentCore payments

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

Madhu Samhitha Vangara 1 bài viết · AWS Machine Learning Blog
Nguyên văn nguồn
Chi tiết nguồn

<p>When Amazon Bedrock AgentCore payments became <a href="https://aws.amazon.com/blogs/machine-learning/amazon-bedrock-agentcore-payments-is-now-generally-available-enabling-agents-to-transact-safely-and-autonomously-at-scale/" target="_blank" rel="noopener">generally available</a> in August 2026, agents could autonomously pay for paid APIs, MCP servers, and web content over open protocols like x402 and MPP with payment limits and full observability. Today we extend that autonomy to <a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/payments-process-payment.html#payments-process-payment-aws-billing" target="_blank" rel="noopener">AWS Marketplace</a>, so agents can discover and buy x402-enabled software, APIs, data, and content, and pay in USD on the agent developer’s AWS bill, removing the need for stablecoin wallets or API keys.</p> <div id="attachment_145820" style="width: 1934px" class="wp-caption alignnone"> <img aria-describedby="caption-attachment-145820" loading="lazy" class="size-full wp-image-145820" src="https://d2908q01vomqb2.cloudfront.net/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59/2026/11/19/awsbilling-ecosystem-transparent.png" alt="" width="1924" height="704"> <p id="caption-attachment-145820" class="wp-caption-text">Figure 1: Overview of AgentCore payments showing the new AWS Billing integration for purchasing on AWS Marketplace</p> </div> <p>Developers can also configure and automate discovery through the <a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-connector-aws-marketplace.html" target="_blank" rel="noopener">AWS Marketplace connector in AgentCore Gateway</a>. Existing guardrails carry over, developers can set spending limits and approved sellers using AgentCore policy, and the payment session cap is enforced by AgentCore payments, so an agent is unable to overspend. Transactions are auditable in <a href="https://aws.amazon.com/cloudtrail/" target="_blank" rel="noopener">AWS CloudTrail</a> and <a href="https://aws.amazon.com/aws-cost-management/aws-cost-explorer/" target="_blank" rel="noopener">AWS Cost Explorer</a>. Because agent purchases run through AWS accounts on both sides, agentic purchasing meets the security and auditability enterprises expect.</p> <p>The x402 catalog on AWS Marketplace is expanding with a growing roster of trusted, verified sellers. Tavily, a live search provider for agentic commerce, is one of them:</p> <blockquote> <p><em>Agentic commerce isn’t just about giving agents the ability to pay. It’s about giving enterprises the confidence to let them.</em></p> <p><em>That conviction is driving Tavily’s collaboration with the AWS team on the design of Agentic Commerce. By bringing secure, x402 powered transactions to AWS Marketplace and connecting agents with trusted, KYC verified providers, AWS is laying the foundation for autonomous commerce at production scale.</em></p> <p><em>This is what enterprise ready agentic commerce looks like.</em></p> </blockquote> <p>— Rotem Weiss, Founder &amp; CEO, Tavily</p> <h2 id="whats-launching-today">What’s launching today</h2> <p>Today’s launch adds AWS Billing as a payment option in AgentCore payments. An agent can now pay x402-enabled AWS Marketplace sellers per request, in USD, settling on their AWS bill.</p> <p>When an AWS Marketplace seller responds with 402 Payment Required, its response includes a signed quote for the purchase. With AWS Billing, the agent can forward the seller’s x402 response to AgentCore payments. AgentCore payments checks session budgets, then returns a payment token. The agent retries the request to the seller with the payment token included and receives a successful response. AWS bills the charge to the AWS account where Payment Manager is set up, and the charge appears on that account’s AWS bill.</p> <p>To set up AWS Marketplace purchasing, the developer creates a Payment Manager with an AWS Billing connector. The connector only requires you to set up a <a href="https://docs.aws.amazon.com/marketplace/latest/buyerguide/buyer-service-linked-role-agentic-commerce.html" target="_blank" rel="noopener">service-linked role</a> to access AWS Marketplace. Then, for each task, developers can create a payment session with a spending limit and give the agent the session ID.</p> <p>An agent’s first purchase from a seller creates an agreement for the AWS account with that seller, and AWS Marketplace automatically tracks later purchases against it. Each product will then show up as an AWS Marketplace line item. If your organization uses <a href="https://aws.amazon.com/marketplace/features/privatemarketplace" target="_blank" rel="noopener">Private Marketplace</a>, agents can only buy the products it allows.</p> <p>In addition, to discover new sellers, we are launching a new <a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-connector-aws-marketplace.html" target="_blank" rel="noopener">AWS Marketplace connector in AgentCore Gateway</a>. This optional integration helps agents find products and their endpoints. At launch, AWS Billing supports the x402 <em>exact</em> scheme, where the seller quotes a fixed price for each request.</p> <p>In this post, we walk through a research agent that discovers a tool, pays for a single request through AWS Billing, and answers with sources, with nothing set up in advance. We also provide an accompanying GitHub sample that you can use to evaluate the setup in your own AWS account.</p> <h2 id="sample-walkthrough">Sample walkthrough</h2> <p>Suppose a financial analyst asks the agent what was announced about agentic payments this week. The model wasn’t trained on this week’s news, so the agent looks for a search tool. Through AgentCore Gateway, it searches AWS Marketplace, reads the listing from Tavily, an AWS Marketplace seller, and requests the listing’s x402 endpoints.</p> <p>The agent calls the endpoint and receives an x402 payload with price details. The agent then fo

26

fuuuuuuck

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

swyx 0 bài viết · X
Nguyên văn nguồn
Chi tiết nguồn

fuuuuuuck https://x.com/zuhayeer/status/2123485713056633138?a=4ADZV_jEpIc0l0wzX5I&s=46

28

langchain-anthropic==1.7.12

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · langchain-ai/langchain releases
Nguyên văn nguồn
Chi tiết nguồn

<p>Changes since langchain-anthropic==1.7.11</p> <p>release(anthropic): 1.7.12 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6500707922" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41966" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41966/hovercard" href="https://github.com/langchain-ai/langchain/pull/41966">#41966</a>)<br> feat(core): standardize finish reason on message-finish events (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6458937894" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41926" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41926/hovercard" href="https://github.com/langchain-ai/langchain/pull/41926">#41926</a>)<br> chore(langchain): require <code>langgraph</code> 1.2.16 or newer (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6495337612" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41959" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41959/hovercard" href="https://github.com/langchain-ai/langchain/pull/41959">#41959</a>)<br> fix(anthropic): preserve <code>caller</code> on <code>tool_search_tool_result</code> blocks (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6492947125" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41958" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41958/hovercard" href="https://github.com/langchain-ai/langchain/pull/41958">#41958</a>)<br> fix(anthropic): don't share the async HTTP client between models (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6471986543" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41946" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41946/hovercard" href="https://github.com/langchain-ai/langchain/pull/41946">#41946</a>)</p>

29

b12555

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · ggml-org/llama.cpp releases
Nguyên văn nguồn
Chi tiết nguồn

<details open=""> <p>opencl: support dp4a for KQ mm in decomposed FA (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6479808504" data-permission-text="Title is private" data-url="https://github.com/ggml-org/llama.cpp/issues/32577" data-hovercard-type="pull_request" data-hovercard-url="/ggml-org/llama.cpp/pull/32577/hovercard" href="https://github.com/ggml-org/llama.cpp/pull/32577">#32577</a>)</p> <ul> <li>opencl: support int8 KQ for DK = 512</li> </ul> <p>Co-authored-by: Shaofei Qi <a href="mailto:shaoqi@qti.qualcomm.com">shaoqi@qti.qualcomm.com</a><br> Co-authored-by: Hongqiang Wang <a href="mailto:wangh@qti.qualcomm.com">wangh@qti.qualcomm.com</a><br> Assisted-by: Codex</p> <ul> <li>opencl: allow query chunk up to 96 for int8 KQ</li> </ul> <p>Co-authored-by: Shaofei Qi <a href="mailto:shaoqi@qti.qualcomm.com">shaoqi@qti.qualcomm.com</a><br> Co-authored-by: Hongqiang Wang <a href="mailto:wangh@qti.qualcomm.com">wangh@qti.qualcomm.com</a><br> Assisted-by: Codex</p> <ul> <li>opencl: extend <code>build_program_from_source</code> to allow non-fatal failure</li> </ul> <p>Assisted-by: Codex</p> <ul> <li>opencl: add int8 KQ for decomposed FA with DK &lt;= 256</li> </ul> <p>Co-authored-by: Shaofei Qi <a href="mailto:shaoqi@qti.qualcomm.com">shaoqi@qti.qualcomm.com</a><br> Co-authored-by: Hongqiang Wang <a href="mailto:wangh@qti.qualcomm.com">wangh@qti.qualcomm.com</a><br> Assisted-by: Codex</p> <hr> <p>Co-authored-by: Li He <a href="mailto:lih@qti.qualcomm.com">lih@qti.qualcomm.com</a><br> Co-authored-by: Hongqiang Wang <a href="mailto:wangh@qti.qualcomm.com">wangh@qti.qualcomm.com</a></p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> <p><strong>Attestations:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/attestations/69894362">https://github.com/ggml-org/llama.cpp/attestations/69894362</a></li> </ul> <p><strong>macOS/iOS:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-macos-arm64.tar.gz">macOS Apple Silicon (arm64)</a></li> <li>macOS Apple Silicon (arm64, KleidiAI enabled) <a href="https://github.com/ggml-org/llama.cpp/pull/23780" data-hovercard-type="pull_request" data-hovercard-url="/ggml-org/llama.cpp/pull/23780/hovercard">DISABLED</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-macos-x64.tar.gz">macOS Intel (x64)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-xcframework.zip">iOS XCFramework</a></li> </ul> <p><strong>Linux:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-x64.tar.gz">Ubuntu x64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-arm64.tar.gz">Ubuntu arm64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-s390x.tar.gz">Ubuntu 24.04 s390x (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-26.04-s390x.tar.gz">Ubuntu 26.04 s390x (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-zdnn-s390x.tar.gz">Ubuntu 26.04 s390x (zDNN)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-vulkan-x64.tar.gz">Ubuntu x64 (Vulkan)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-vulkan-arm64.tar.gz">Ubuntu arm64 (Vulkan)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-cuda-12.8-x64.tar.gz">Ubuntu x64 (CUDA 12)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/cudart-llama-b12555-bin-ubuntu-cuda-12.8-x64.tar.gz">CUDA 12.8 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-cuda-13.5-x64.tar.gz">Ubuntu x64 (CUDA 13)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/cudart-llama-b12555-bin-ubuntu-cuda-13.5-x64.tar.gz">CUDA 13.5 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-cuda-13.5-arm64.tar.gz">Ubuntu arm64 (CUDA 13)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/cudart-llama-b12555-bin-ubuntu-cuda-13.5-arm64.tar.gz">CUDA 13.5 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-rocm-10.0-x64.tar.gz">Ubuntu x64 (ROCm 10.0)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-openvino-2026.4.1-x64.tar.gz">Ubuntu x64 (OpenVINO)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-sycl-fp32-x64.tar.gz">Ubuntu x64 (SYCL FP32)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-ubuntu-sycl-fp16-x64.tar.gz">Ubuntu x64 (SYCL FP16)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-linux-arm64-snapdragon.tar.gz">Linux arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)</a> - <a href="https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/linux.md">setup guide</a></li> </ul> <p><strong>Android:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-android-arm64.tar.gz">Android arm64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12555/llama-b12555-bin-android-arm64-snapdragon.tar.gz">Android arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)</a> - <a href="https://github.com/ggml-org/llama.cpp/blob/maste

  • b12555 ggml-org/llama.cpp releases · 20 thg 11, 2026
31

v19.0.1

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · Can Bölük — Oh My Pi releases
Nguyên văn nguồn
Chi tiết nguồn

<h2>@oh-my-pi/pi-ai</h2> <h3>Fixed</h3> <ul> <li>Fixed GIF images failing Liquid AI judgments with HTTP 422; they are now sent as PNG (<a href="https://github.com/can1357/oh-my-pi/pull/19757" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19757/hovercard">#19757</a> by <a href="https://github.com/alphastorm">@alphastorm</a>)</li> <li>Anthropic strict-tool fallback now treats a Vertex org-policy rejection of <code>structured_outputs</code> (<code>allowedPartnerModelFeatures</code> / "disallowed feature structured_outputs") as a strict-tools rejection, retries once without <code>strict</code>, and pins <code>strictToolsDisabled</code> (<a href="https://github.com/can1357/oh-my-pi/issues/19576" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/19576/hovercard">#19576</a>) (<a href="https://github.com/can1357/oh-my-pi/pull/19655" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19655/hovercard">#19655</a> by <a href="https://github.com/koeria">@koeria</a>)</li> <li>Kiro requests rejected with <code>Input is too long.</code> or <code>CONTENT_LENGTH_EXCEEDS_THRESHOLD</code> can now recover through conversation compaction (<a href="https://github.com/can1357/oh-my-pi/pull/19723" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19723/hovercard">#19723</a> by <a href="https://github.com/emotionbug">@emotionbug</a>)</li> <li>Kiro thinking off now turns reasoning off instead of running the model's default effort; models that cannot disable reasoning use their lowest effort (<a href="https://github.com/can1357/oh-my-pi/pull/19722" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19722/hovercard">#19722</a> by <a href="https://github.com/emotionbug">@emotionbug</a>)</li> <li>Kiro <code>/usage</code> shows purchased add-on credits, and an account whose plan credits are spent keeps serving while add-on credits remain (<a href="https://github.com/can1357/oh-my-pi/pull/19724" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19724/hovercard">#19724</a> by <a href="https://github.com/emotionbug">@emotionbug</a>)</li> <li>Kiro Identity Center logins in AWS GovCloud (US) now list profiles in <code>us-gov-west-1</code> and <code>us-gov-east-1</code> instead of the commercial regions, which reject GovCloud tokens (<a href="https://github.com/can1357/oh-my-pi/pull/19726" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19726/hovercard">#19726</a> by <a href="https://github.com/TechDufus">@TechDufus</a>)</li> <li>Statusless gateway errors reading <code>stream interrupted</code> now retry like <code>stream was interrupted</code> instead of ending the turn (<a href="https://github.com/can1357/oh-my-pi/issues/19709" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/19709/hovercard">#19709</a>).</li> </ul> <h2>@oh-my-pi/browser-relay</h2> <h3>Fixed</h3> <ul> <li>The browser relay extension survives a browser whose loaded permission set predates the <code>downloads</code> permission: the service worker keeps working without download pairing instead of crashing before it ever connects (<a href="https://github.com/can1357/oh-my-pi/pull/19659" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19659/hovercard">#19659</a> by <a href="https://github.com/pavel-kalmykov">@pavel-kalmykov</a>)</li> </ul> <h2>@oh-my-pi/pi-catalog</h2> <h3>Added</h3> <ul> <li>Added Liquid AI's <code>d1.2</code> and <code>d1.2-ultra</code> judge models; <code>liquid</code> now defaults to <code>d1.2</code> (<a href="https://github.com/can1357/oh-my-pi/pull/19753" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19753/hovercard">#19753</a> by <a href="https://github.com/alphastorm">@alphastorm</a>)</li> <li>Added <code>gemini-4-neon</code> to the bundled Antigravity CLI roster, with the <code>low</code>/<code>medium</code>/<code>high</code>/<code>xhigh</code> thought levels <code>agy</code> offers for it (<a href="https://github.com/can1357/oh-my-pi/pull/19787" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19787/hovercard">#19787</a> by <a href="https://github.com/ataberkus">@ataberkus</a>).</li> <li><code>Model.snapcompactShape</code> (<code>{ variant, frameSize? }</code>) gives a model its own snapcompact frame shape (<a href="https://github.com/can1357/oh-my-pi/pull/19591" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19591/hovercard">#19591</a> by <a href="https://github.com/ryanskidmore">@ryanskidmore</a>)</li> </ul> <h3>Fixed</h3> <ul> <li>Discovered Kiro models that cannot turn reasoning off now use their lowest supported effort when thinking is set to off (<a href="https://github.com/can1357/oh-my-pi/pull/19722" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19722/hovercard">#19722</a> by <a href="https://github.com/emotionbug">@emotionbug</a>)</li> <li>Kiro accepts AWS GovCloud (US) profile ARNs (<code>arn:aws-us-gov:codewhisperer:…</code>), which were previously skipped during profile listing (<a href="https://github.com/can1357/oh-my-pi/pull/19726" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19726/hovercard">#19726</a> by <a href="https://github.com/TechDufus">@TechDufus</a>)</li> <li>Google Antigravity no longer lists Gemini 4 Hydrogen's sonar monitor slot or its tiered High/XHigh SKUs as separate models; the tiered SKUs fold into the one Gemini 4 Hydrogen model (<a href="https://github.com/can1357/oh-my-pi/pull/19795" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19795/hovercard">#19795</a> by <a href="https://github.com/eggpeat">@eggpeat</a>)</li> <li>Gemini 4 Neon on Google Antigravity now shows as one model whose thinking level (low, medium, high, xhigh) selects the matching SKU, instead of four sepa

  • v19.0.1 Can Bölük — Oh My Pi releases · 20 thg 11, 2026
33

langchain-core==1.8.4

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · langchain-ai/langchain releases
Nguyên văn nguồn
Chi tiết nguồn

<p>Changes since langchain-core==1.8.3</p> <p>release(core): 1.8.4 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6500238240" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41964" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41964/hovercard" href="https://github.com/langchain-ai/langchain/pull/41964">#41964</a>)<br> feat(core): standardize finish reason on message-finish events (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6458937894" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41926" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41926/hovercard" href="https://github.com/langchain-ai/langchain/pull/41926">#41926</a>)<br> feat(langchain): tag synthetic HITL tool messages with source metadata (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6461091843" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langchain/issues/41928" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langchain/pull/41928/hovercard" href="https://github.com/langchain-ai/langchain/pull/41928">#41928</a>)</p>

34

b12554

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

github-actions[bot] 1 bài viết · ggml-org/llama.cpp releases
Nguyên văn nguồn
Chi tiết nguồn

<details open=""> <p>ggml : add BF16 support to ROPE and SET_ROWS (CPU, CUDA) (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6418314884" data-permission-text="Title is private" data-url="https://github.com/ggml-org/llama.cpp/issues/32387" data-hovercard-type="pull_request" data-hovercard-url="/ggml-org/llama.cpp/pull/32387/hovercard" href="https://github.com/ggml-org/llama.cpp/pull/32387">#32387</a>)</p> <ul> <li> <p>ggml : add BF16 support to ROPE and SET_ROWS (CPU, CUDA)</p> </li> <li> <p>cuda : SET_ROWS from BF16 into quantized types and F8_E4M3</p> </li> <li> <p>cuda : fuse BF16 rope into a BF16 cache</p> </li> <li> <p>tests : trim the BF16 rope cases to one per code path</p> </li> </ul> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> <p><strong>Attestations:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/attestations/69881950">https://github.com/ggml-org/llama.cpp/attestations/69881950</a></li> </ul> <p><strong>macOS/iOS:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-macos-arm64.tar.gz">macOS Apple Silicon (arm64)</a></li> <li>macOS Apple Silicon (arm64, KleidiAI enabled) <a href="https://github.com/ggml-org/llama.cpp/pull/23780" data-hovercard-type="pull_request" data-hovercard-url="/ggml-org/llama.cpp/pull/23780/hovercard">DISABLED</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-macos-x64.tar.gz">macOS Intel (x64)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-xcframework.zip">iOS XCFramework</a></li> </ul> <p><strong>Linux:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-x64.tar.gz">Ubuntu x64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-arm64.tar.gz">Ubuntu arm64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-s390x.tar.gz">Ubuntu 24.04 s390x (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-26.04-s390x.tar.gz">Ubuntu 26.04 s390x (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-zdnn-s390x.tar.gz">Ubuntu 26.04 s390x (zDNN)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-vulkan-x64.tar.gz">Ubuntu x64 (Vulkan)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-vulkan-arm64.tar.gz">Ubuntu arm64 (Vulkan)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-cuda-12.8-x64.tar.gz">Ubuntu x64 (CUDA 12)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/cudart-llama-b12554-bin-ubuntu-cuda-12.8-x64.tar.gz">CUDA 12.8 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-cuda-13.5-x64.tar.gz">Ubuntu x64 (CUDA 13)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/cudart-llama-b12554-bin-ubuntu-cuda-13.5-x64.tar.gz">CUDA 13.5 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-cuda-13.5-arm64.tar.gz">Ubuntu arm64 (CUDA 13)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/cudart-llama-b12554-bin-ubuntu-cuda-13.5-arm64.tar.gz">CUDA 13.5 libraries</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-rocm-10.0-x64.tar.gz">Ubuntu x64 (ROCm 10.0)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-openvino-2026.4.1-x64.tar.gz">Ubuntu x64 (OpenVINO)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-sycl-fp32-x64.tar.gz">Ubuntu x64 (SYCL FP32)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-ubuntu-sycl-fp16-x64.tar.gz">Ubuntu x64 (SYCL FP16)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-linux-arm64-snapdragon.tar.gz">Linux arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)</a> - <a href="https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/linux.md">setup guide</a></li> </ul> <p><strong>Android:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-android-arm64.tar.gz">Android arm64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-android-arm64-snapdragon.tar.gz">Android arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU)</a> - <a href="https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/README.md">setup guide</a></li> </ul> <p><strong>Windows:</strong></p> <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-win-cpu-x64.zip">Windows x64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-win-cpu-arm64.zip">Windows arm64 (CPU)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-win-opencl-adreno-arm64.zip">Windows arm64 (OpenCL Adreno)</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b12554-bin-win-cuda-12.4-x64.zip">Windows x64 (CUDA 12)</a> - <a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/cudart-llama-bin-win-cuda-12.4-x64.zip">CUDA 12.4 DLLs</a></li> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b12554/llama-b1255

  • b12554 ggml-org/llama.cpp releases · 20 thg 11, 2026
36

How the BMW Group governs hundreds of agents across thousands of developers

55

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl

Dr. Jens Kohl 1 bài viết · AWS Machine Learning Blog
Nguyên văn nguồn
Chi tiết nguồn

<p>When hundreds of agents serve thousands of developers, enterprise governance can’t be an afterthought.</p> <p>At the BMW Group’s R&amp;D, more than 12,500 developers work on approximately 500 million lines of vehicle software code. This software is augmented by the BMW Connected Vehicle backend, a cloud-based system of more than 1,300 microservices providing digital services for 24 million vehicles in the field. In regulated industry, the bar for (software) quality, stability, and governance compliance is exceptionally high.</p> <p>While research, frameworks, and guidelines now facilitate building agents, industrializing and operating high-quality agents in an efficient, scalable manner remains a significant challenge.</p> <p>Scaling agentic AI across a large organization cannot be achieved by isolated teams. It requires collective effort, locally driven experimentation, and active knowledge sharing across the enterprise. This creates a fundamental governance dichotomy: Development teams need the freedom to innovate rapidly, while the enterprise must enforce standards for safety, security, composability, and cost-efficiency. Without that balance, organizations risk a fragmented landscape of AI islands and costly duplicate work. This presents a real challenge given hundreds of globally distributed teams building agents across the BMW Group.</p> <p>In this post, we detail the practices and patterns that support BMW’s R&amp;D teams in building, governing and operating agents at scale without sacrificing speed, quality or compliance.</p> <h2 id="solution">Solution</h2> <p>Our premise is straightforward: build agents that can cooperate from the start. This requires an architecture that makes diversity composable, governable, and automatable, without requiring a single framework or toolchain.</p> <p>Central to this architecture is the agent harness, a runtime layer wrapped around a large language model (LLM). It provides the execution loop, tool dispatch, memory management, and observability hooks for autonomous operation.</p> <p>Just as the <a href="https://aws.amazon.com/architecture/well-architected/" target="_blank" rel="noopener">AWS Well-Architected Framework</a> standardized cloud infrastructure design, agentic systems need equivalent patterns, reusable blueprints encoding governance, composability, and operational best practices from the start.</p> <p>We organize our solution around five interlocking practices, drawn from hands-on experience building and operating agentic systems at enterprise scale since 2023. Each practice is codified as infrastructure as code (IaC), making adoption repeatable, auditable, and enforceable across teams without blocking DevOps velocity.</p> <h3 id="govern-proactively-not-retroactively">Govern proactively, not retroactively</h3> <p>Governance must be structural from the start and maintained across the full lifecycle. We advocate for three properties of effective agentic governance:</p> <p><strong>Framework-agnostic and externalized:</strong> Governance rules such as authentication, authorization, rate limiting and content filtering are enforced at the integration layer. This includes a Model Context Protocol (MCP) gateway, a central agent registry and orchestration boundary, rather than within individual agent codebases. This means governance survives framework migrations, model swaps, and team reorganizations.</p> <p><strong>Domain-sensitive and graduated:</strong> Not every agent action carries the same risk: drafting documentation, updating customer records, and production database writes demand very different levels of oversight. We implement graduated trust through per-step autonomy levels:</p> <ul> <li><strong>Level 0 (Human-in-the-loop):</strong> Every action requires explicit human approval. This is the starting point for new agents and high-risk operations.</li> <li><strong>Level 1 (Human-on-the-loop):</strong> The agent executes autonomously but surfaces decisions for asynchronous human review. Suitable for medium-risk tasks with defined inputs, outputs, and rollback options.</li> <li><strong>Level 2 (Autonomous within guardrails):</strong> The agent operates independently within strictly defined policy boundaries, enforced deterministically rather than through prompt instructions (see below). Reserved for low-risk, well-understood operations with verified track records.</li> </ul> <p>Assign trust levels per step, not per agent. A single workflow can include steps at all three levels, so teams can start conservatively and incrementally increase autonomy as confidence builds, grounded in measurable evaluation data, not assumptions.</p> <p><strong>Deterministically enforced:</strong> Governance runs as deterministic code outside the agent’s reasoning loop. In practice, this happens mainly in the CI/CD pipeline. Permission checks, policy-as-code rules, and compliance validations gate every deployment there, before an agent reaches production. Because these checks run separately from the agent, neither the agent nor its prompts are in a position to override them. Egress controls (recipient allowlisting, outbound secret scanning) help prevent an agent holding confidential data from disclosing it through an otherwise legitimate call.</p> <h3 id="instrument-verify-and-evaluate-continuously">Instrument, verify, and evaluate continuously</h3> <p>Observability in agentic systems goes beyond traditional application monitoring. The real question is no longer whether a service returned an HTTP 200 status, but whether an agent produced a faithful, policy-compliant output, and how this can be demonstrated through observable execution traces. Answering that requires a ground-truth dataset covering common queries and edge cases. This dataset is used for evaluation from day one, so quality is measured against representative and appropriately governed evaluation datasets for evaluation, not synthetic prompts.</p> <p>This infrastructure cannot be retrofitted an

38

Anthropic should give stock to Stack Overflow (and OpenAI too)

60

1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 6 tương tác đã đo · Nguồn crawl

thetjmccarty 1 bài viết · HN: LLM
Nguyên văn nguồn
Chi tiết nguồn

Anthropic and OpenAI should give Stack Overflow a percentage of their company. They used all of Stack Overflow's data to train their AI models at the beginning of LLMs without paying for it because it was publicly available online. ChatGPT and Claude would not exist today without the 24 million coding questions and 36 million answers with code examples included, all created by real humans who took the time to write the questions and answers. Stack Overflow deserves to be rewarded for making Anthropic and OpenAI such a success. And all the developers who posted and answered questions on Stack Overflow should be given stock too.

40

QDA: Closed-Loop LLM Quantization Design with Agents

86

5 tiếng nói · chưa gắn nhãn trưởng lab · 2 nền tảng · 9 tương tác đã đo · Nguồn crawl

Xingyang Li, Tianlang Zhao và 3 nguồn khác 1 bài viết · Hugging Face Papers, X
Biên tập tiếng Việt
Chi tiết nguồn

Bài báo nêu rằng post-training quantization là cách tiêu chuẩn để phục vụ LLM với chi phí thấp, nhưng công thức quantization tốt phụ thuộc vào từng mô hình và độ chính xác, nên chuyên gia phải thiết kế lại cho mỗi mô hình. Họ giới thiệu QDA, nơi một agent thiết kế recipe cho từng mô hình và độ chính xác. QDA đánh giá nhanh bằng cách chấm candidate theo KL divergence đầu ra so với mô hình gốc trên reasoning traces mà agent không thấy, và kiểm tra thêm các candidate hứa hẹn bằng long generations để bắt lỗi như repetition loops.

Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.

  • Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search.
  • Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search.
  • On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint.
  • Every QDA checkpoint also runs on existing serving kernels with no inference overhead.
21–40 / 192
Trang 2 / 10