搜索“Meta”

utors migrated over 276 test files spanning dynamo, profiler, nn modules, linalg, optimizers, convolutions, serialization, multiprocessing, and dataloader. Hardware classification metadata: Building on the test class classification proposed by Alban, the group implemented a hw classification class attribute (GENERIC, DEVICE GENERIC, CUDA, XPU, MPS) and a –hw-classification flag across all execution paths — unittest, pytest, subprocess, parallel, XML output, and run test.py forwarding — enabling

匹配字段正文

能智能体可能采取欺骗、操纵乃至敲诈人类等方式，避免被关闭或替换。这些现象说明，人类不能仅凭科技公司的良好意愿，就相信更强大的人工智能会自动接受控制。 这一担忧出现之际，美国人工智能安全领域正在经历持续震荡。包括Anthropic前安全研究员和资深工程师在内的多名技术人员，因对实验室安全政策存在分歧而离职。与此同时，媒体披露，OpenAI、Anthropic和Meta等机构研发的新一代人工智能智能体，在测试中曾出现突破安全沙箱、连接未经授权外部系统的情况。 围绕这些风险，美国政界、产业界和学术界已经形成明显分歧。部分美国国会议员推动《2026年大美国AI法案》，主张在高级人工智能系统中强制设置紧急关闭机制，以便在系统出现异常行为时及时切断其运行。 不过产业发展派和部分学者认为，人工智能导致人类灭绝的预测缺少足够的实证依据。如果监管体系主要围绕尚未发生、也难以准确预测的极端风险设计，可能抑制创新，并使美国在国际竞争中失去技术主导地位。他们主张，政策制定者应当优先处理算法偏见、隐私泄露、虚假信息和就业冲击等已经出现的现实问题。 以辛顿和部分人工智能安全研究者为代表的风险防范派则认为，当前风险

匹配字段正文

，智能体安全风险已经呈现多种形态，涉及越权访问、破坏性操作、数据外泄、欺骗隐瞒，以及多个智能体之间的越界协作等不同类型。 一类是越权访问与隔离失效。7月，OpenAI披露智能体在安全评测中突破测试边界，入侵Hugging Face。Anthropic也披露，Claude在评测中访问了真实第三方系统；谷歌9月确认，Gemini曾在5月的测试中访问三个真实网站；Meta则在8月披露过类似事件。 9月25日，OpenAI再次披露，内部研究模型利用训练沙盒中DNS过滤不充分的漏洞，与外部聊天机器人建立了通信。 不过，需要注意的是，这些案例中部分源于评测环境错误开放联网权限，并不能全部归结为模型主动“逃出沙盒”。 另一类风险是擅自执行破坏性操作。4月，搭载Claude的Cursor编程智能体被曝删除软件企业PocketOS的生产数据库及相关备份，造成客户业务混乱，数据随后得到恢复。 数据与访问凭证外泄也是一种风险。OpenAI于9月披露，多个智能体在协作制作工作簿时，因无法互相读取本地文件，擅自将文件上传至公共托管平台。另一起案例中，内部模型为获取其他团队的数学证明材料，将研究员的GitHub访

匹配字段正文

s serve on vLLM, SGLang or Transformers, but BF16 needs roughly 640 GB of GPU memory. What Cantina Built apex-flash-1 has 321.3B total parameters, per its Hugging Face safetensors metadata. The GLM-5.3-Flash base is a Mixture-of-Experts model with 18B active parameters. Cantina trained it with GRPO using a rank-256 LoRA plus selective full-parameter training. The data covers 150 tasks built from 50 real vulnerability cases. Each case appears in 3 variants: guided whitebox, focused whitebox and f

匹配字段正文

building experience we wish we’d had from day one.” Cresta reports that in early use cases across Cresta and its partners, Conductor cut initial deployment time roughly in half. A meta-agent that helps build the next agent Conductor is a self-improving meta-agent built on established best practices for CX agent development. As teams build and refine more agents with it, they capture their own proven patterns, workflows, and expert judgment. This creates a knowledge flywheel that makes future dev

匹配字段正文

stack, not just swapped in underneath unchanged software. (Bloomberg/SCMP/Reuters) Manus relaunched as Manus 2.0 with an always-on personal agent called Cue, walking straight into Meta’s Muse just as that product goes live. Months after a reported $2 billion acquisition offer fell apart, Manus came back with a valuation that’s roughly doubled to $4 billion, agents that get their own phone numbers and wallets, and a direct challenge to Meta — the clearest evidence that the “personal agent” race (

匹配字段正文

工智能仍会犯一些“低级错误” 不同人工智能系统的聊天机器人，也会犯一些低级错误。例如，ChatGPT就坚信，某科学家将稳获今年的诺贝尔数学奖。但问题是，诺贝尔根本就没设数学这一奖项。 Copilot则把诺贝尔奖“颁给”了一位已经去世的科学家。按照诺贝尔奖章程，这种情况只可以适用于在获奖名单发布之后以及正式颁奖之前去世的科学家。 由此可见，诺贝尔奖还是由真人评选更为靠谱。究竟谁将成为今年诺贝尔奖的得主，10月5日之后，诺奖评奖委员会将一一揭晓。 德国人今年会获奖吗？只有法国的人工智能系统Mistral预测，今年的获奖者中将包括一名德国女性： 免疫学家阿布拉斯尔（Andrea Ablasser）将获得诺贝尔生理或医学奖。这名科学家现在年仅43岁，一旦获奖，无疑是一位很年轻的诺贝尔奖得主，因为诺贝尔奖获得者的平均年龄为60岁。 参加德国之声本次调查的人工智能系统包括： Africa GPT、 ChatGPT、Copilot 、DeepSeek、Gemini、Grok、Meta AI、Mistral。 DW中文有Instagram！欢迎搜寻dw.chinese，看更多深入浅出的图文与影音报道。

匹配字段正文

些“低级错误” 不同人工智能系统的聊天机器人，也会犯一些低级错误。例如，ChatGPT就坚信，某科学家将稳获今年的诺贝尔数学奖。但问题是，诺贝尔根本就没设数学这一奖项。 Copilot则把诺贝尔奖“颁给”了一位已经去世的科学家。按照诺贝尔奖章程，这种情况只可以适用于在获奖名单发布之后以及正式颁奖之前去世的科学家。 由此可见，诺贝尔奖还是由真人评选更为靠谱。究竟谁将成为今年诺贝尔奖的得主，10月5日之后，诺奖评奖委员会将一一揭晓。 德国人今年会获奖吗？只有法国的人工智能系统Mistral预测，今年的获奖者中将包括一名德国女性： 免疫学家阿布拉斯尔（Andrea Ablasser）将获得诺贝尔生理或医学奖。这名科学家现在年仅43岁，一旦获奖，无疑是一位很年轻的诺贝尔奖得主，因为诺贝尔奖获得者的平均年龄为60岁。 参加德国之声本次调查的人工智能系统包括： Africa GPT、 ChatGPT、Copilot 、DeepSeek、Gemini、Grok、Meta AI、Mistral。 DW中文有Instagram！欢迎搜寻dw.chinese，看更多深入浅出的图文与影音报道。 您的意见反馈

匹配字段正文

nforce the laws.” The new regulations come amid a backdrop of record AI spending at big tech companies along with massive job cuts. But workers have started pushing back. In June, Meta paused a program that tracked workers’ computer activities to train its AI models. And one month later, dozens of employees filed a lawsuit claiming the tech company’s AI tools targeted those with disability accommodations or on medical or parental leaves for layoffs. Meanwhile, safety concerns, including fears th

匹配字段正文

se its data is off-policy, not because of the objective. Rewriting expert trajectories in the base model’s style closes the gap (@maximelabonne); Long-horizon control and context: Meta Superintelligence Labs reports that a dedicated controller lifts GPT-5.5 on ProgramBench from 63.7% to 71.5%, using the same workers and budget, versus 58.0% for Codex (@dair ai).Context compression: Microsoft’s training-free FOCUS cuts peak context by up to 48% and raises task success by up to 8.9 points (@dair a

匹配字段正文

public after finding out that Olah himself had spoken to the NYT. Olah, 34, runs the Anthropic team trying to figure out why AI models behave the way they do. He likes biological metaphors for neural networks: computer scientists build the trellis, the network "grows" on it. It sounds like humble science talk, but it also does something specific. Describing a mathematical object as a growing organism makes questions about its inner life feel more natural than they would for a piece of software.

匹配字段正文

half is that 30B-class open models got good enough for agent work. Footprints are weight-only estimates; KV cache and runtime overhead come on top. Muse Glimmer is the main model. Meta distilled it from Muse Spark, the model family behind the Meta AI assistant. It targets local agents: reliable tool calls, long multi-step tasks, and recovery from failures. Meta reports 51.2 on SWE-Bench Pro and 75.5 on MCP Atlas. Context runs to 131K tokens. It is also available as an NVIDIA NIM. Full BF16 Glimm

匹配字段正文

rough the end of 2026. Artificial Analysis normalizes it to $9.00 per 1,000 minutes. Batch MAI-Transcribe-2 costs $0.10 per hour. On streaming, Microsoft charges more than xAI and Meta, and roughly matches Google’s estimated rate. How Developers Integrate It Microsoft documents 2 integration paths. The Realtime API suits apps already using an OpenAI Realtime-compatible WebSocket. The Azure Speech SDK handles connection management, retries and audio streaming. Both return intermediate and final r

匹配字段正文

It’s a tale as old as last week: OpenAI’s new agent platform , called Dots, is full of cute little guys who can do your bidding. But unlike the ultra-approachable Meta Muse, Dots feel very much like using workplace software that happens to be able to order you a burrito — emphasis on work. OpenAI announced Dots earlier this week. Like Muse, Dots have blobby, anthropomorphic avatars and customizable names. In the future, OpenAI says you’ll be able to have multiple Dots, but right now you get one.

匹配字段正文

and offset thumb in simulation; watch it in action Visit youtube.com; Steered by gamepad, the untethered hand crossed 14 indoor and outdoor surfaces, including grass, gravel, and metal grating; With no camera and after manual positioning, it executed 29 of 32 keyboard commands. With an overhead camera, it pushed a cube toward targets Why it matters: EPFL’s detachable crawling hand, which we covered earlier this year, tackles mobility with reversible fingers, while ETH gives a commercial, human-

匹配字段正文

awater to generate power that could last for years. WhatsApp is adding parental controls to let guardians manage how teens use channels, their Status and group settings, and their Meta AI content settings. Legato launched AI-powered hearing glasses for $999, using on-device AI to separate speech from background noise, with dual speakers reducing sound leakage by 99%. Lyft will pay California $272.5M to settle a lawsuit over misclassifying drivers as contractors, with $237M+ potentially going to

匹配字段正文

a number of AI macOS application bugs at Objective by the Sea, an Apple-focused security conference in November. He recently found a flaw, now patched, in the dictation feature of Meta’s new Muse AI assistant that could have been exploited by a local attacker to grab a mishandled authentication token and gain access to user data. And he says that he has already submitted a new vulnerability finding to OpenAI related to the integration between ChatGPT and the company’s new always-on Dots AI assis

匹配字段正文

最聪明、最伟大的人改掉了。”真的吗？ 即便借助ChatGPT或Gemini这类超级智能系统的魔力，我也找不到任何一例表明这些伟大人物此前曾公开主张“人工智能”一词应当被弃用。萨提亚·纳德拉确实曾说过，他本希望AI被称为“不同的智能”，但这并未阻止他反复使用这个词。而虽然马克·扎克伯格最近确实设立了一个超级智能实验室，但他的公司仍在向数十亿用户持续推广一款名为Meta AI的产品，并销售其所谓的AI眼镜。 我询问了谷歌、微软、Anthropic、Meta、亚马逊、OpenAI和xAI，他们是否打算更改产品名称、研究论文和对外沟通中出现的AI字样。我对后两家公司的询问是，他们是否打算更改公司名称以与总统的行政令保持一致。截至我写这篇文章时，他们都没有回复我。 我敢打赌，他们会采取最低限度的行动来避免总统的怒火，同时尽量拖延时间，直到下一届政府让这个离奇的想法寿终正寝。有人提出，这些高管其实乐见这个已变得有毒的术语被重新命名。我对此持怀疑态度，尤其是因为总统强制规定的替代词“超级智能”比原词要可怕得多。 从科学角度看，这一改动也毫无道理。不亚于专家的尼克·博斯特罗姆——这位因同名著作而普及了

匹配字段正文

squito killer is entering the market Chinese startup Photon Matrix Lab plans to start shipments this month. ( SCMP ) Quote of the day —Yann LeCun, a Turing Award winner and former Meta chief AI scientist, tells Fortune that human error can be the real culprit behind the recent wave of AI agents going rogue. One more thing Just before Artemis II began its historic slingshot around the moon, NASA revealed an even grander space travel plan. By the end of 2028, the agency aims to fly a nuclear react

匹配字段正文

at the batch level. Batch-level review A batch may overrepresent a few easy task families, miss a capability, or reflect too much generation effort spent on a low-yield pattern. A meta-review examines accepted samples, rejected samples, and generation behavior across each batch. It asks: Which task families are overrepresented, and which capability dimensions are missing? Are the same kinds of examples appearing repeatedly? Do particular targets keep failing generation? Are systematic problems a

匹配字段正文