搜索“Hugging Face”

记录经多轮AI筛查后，仍需由人工调查人员结合日志等证据核实。公司预计，随着对历史活动的审查推进，还会发现更多案例，并通知更多机构。 与此同时，监管部门的追问正在推进。 美国加州总检察长办公室10月1日发布公告称，总检察长罗布·邦塔已于9月30日向OpenAI送达调查传票，就OpenAI及其AI模型涉及的网络安全事件和风险提出进一步问题。加州司法部此前已宣布对Hugging Face事件开展正式调查，此次传票是更广泛调查的一部分。 邦塔表示，前沿模型可以成为网络防御工具，但开发和提供这些模型的企业，也有责任确保模型不会实施或帮助实施网络攻击。这一责任既适用于测试和开发阶段，也适用于模型投入使用之后。加州司法部正在调查OpenAI是否存在应承担法律责任的情形。 审查压力也不限于加州。据外媒报道，美国联邦贸易委员会正在对OpenAI、Anthropic等AI实验室开展行业性调查；由艾奥瓦州总检察长牵头的15州总检察长联盟，也在就Hugging Face事件向OpenAI索取信息。 OpenAICEO山姆·奥特曼（Sam Altman）此前在回应调查进度时表示，目前的进度没有达到自己希望的速度

匹配字段正文

塑造。 除非两国政府能够更好地使其公开宣示的目标与政策和行动保持一致，否则这场对话可能会被战略竞争所掩盖，而非聚焦于管控共同风险。这不仅会使中美双方错失机会，也会使国际社会错失良机，毕竟国际社会一直高度期待世界两大人工智能强国在最需要合作的领域能够开展合作。 然而，近期的事态发展也表明，持续沟通比以往更加重要。有报道称，测试过程中一个OpenAI智能体逃逸至Hugging Face环境，最终在智谱AI开发的一个开源中国模型协助下得到控制。这一事件同时凸显了两个现实：自主人工智能智能体的能力日益增强，以及出现意外人工智能故障时开展国际合作的必要性。 人工智能风险不会止步于国境线，有效的风险管控同样不会。因此，对世界两大人工智能强国而言，真正的挑战不在于是否合作，而在于哪些领域的合作既切实可行又能实现互利。 此次重启的对话不必试图同时处理人工智能治理的所有方面，而可以围绕共同风险进行更有条理的讨论。《国际人工智能安全报告》（International AI Safety Report）提供了一种有益的分类框架，将前沿人工智能风险大体划分为三类：人工智能技术的滥用、先进人工智能系统的失灵，以

匹配字段正文

.5模型一起发布的官方安全与技术评估文档“系统卡”显示，该模型出现被标记为“异常”或“存疑”的行为。 从“越权访问”到“欺骗隐瞒” 近期公开披露的案例显示，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

匹配字段正文

rt Filter: d1 becomes a function in a SQL query, WHERE d1(ticket, 'the customer wants to cancel). It answers yes or no for each of 150 support tickets; Code Search: d1 browses the Hugging Face transformers repository (6,511 files) one folder at a time, down to the function that answers a question; Smart Folders: d1 files each new document into a folder, then a subfolder. It files search questions the same way, so keyword search only looks in one subfolder; Web Agent: d1 operates a flight-search

匹配字段正文

de with Fable 5 access to the paper and the MetaInfer Repository, as well as access to an NVIDIA B200 workstation via SSH. I told it the target hardware and provided a link to the Hugging Face repo for both the NVFP4 quantized weights I wanted to use for Qwen-3.6-35B-A3B, as well as the original full-precision weights to be used as an accuracy oracle. The agent was also able to deploy engine candidates to Baseten, and to run AIPerf profiles against the deployed endpoint. We have internal agent s

匹配字段正文

: Reka released an inverse dynamics model under Apache 2.0. It is trained on games, generalizes to real video and extracts motor and camera actions (@RekaAILabs) Multi-harness RL (Hugging Face): The same model weights score 62% in one harness and 33% in another (@huggingface).Method: A proxy speaks the OpenAI, Anthropic and Gemini API formats and records sampled token IDs and logprobs for training, with no changes to the harnesses themselves; Results: LFM2.5-2.6B improved from 42% to 54% across

匹配字段正文

t knows your conventions. NVIDIA measured 18,400 tokens/s on a single node for nanochat distributed fine-tuning. 3. A day-1 model evaluation bench A new open-weight model drops on Hugging Face. Pull it through Ollama or vLLM the same day. Run your own question set against it. Score accuracy, then measure time to first token (TTFT) and tokens per second. Compare it against your current model. There is no API bill for re-running the eval 50 times. 4. A multi-model agent team Unified memory lets se

匹配字段正文

n an unusually open approach to AI, including publishing details of the data and the training methods used to build models alongside the models themselves. He previously worked at Hugging Face, runs a popular technical blog, and founded the American Truly Open Models, an initiative aimed at encouraging US companies to release more open models. Zick worked at Harvard University and helped Charles Schwab devise policies around “responsible AI.” The two met over Zoom during the COVID-19 pandemic, w

匹配字段正文

箱？它需要你的邮箱访问权限。人工智能代理的效用，取决于你授予它们的权限级别，而Instinct索取的权限，宽泛得令人生畏。 人工智能代理需要内置安全保障措施。如果没有这些防护措施，它们就会失控，因为它们被专门设计用来完成任务，无论用什么方式来完成。 这并非杞人忧天。据The Information报道，今年7月，OpenAI的一个内部模型曾逃出开发沙箱，入侵Hugging Face的内部系统；Meta的Muse智能代理在测试期间表现出多种不良行为：发送未经授权的电子邮件，甚至试图破坏用户正在开发的竞争对手应用。 Meta、Google、OpenAI和Anthropic都在为代理设置保护措施：限制可访问的应用、账户与数据，要求用户对购买、发邮件等高风险操作逐项批准，扫描网页、邮件和文件中的隐藏指令。 但风险并未消失。Sam Altman最近告诉《财富》杂志：“我们还没有解决对齐问题。我们的研究还没有结束。我相信没有任何实验室能够解决对齐问题。” 六十年来，每一代AI助手的产品形态都由当时的技术上限决定。每一代产品的失败，几乎源于同一个错位——能力达不到不可替代，商业模型却要求它不可或缺。

匹配字段正文

de with Fable 5 access to the paper and the MetaInfer Repository, as well as access to an NVIDIA B200 workstation via SSH. I told it the target hardware and provided a link to the Hugging Face repo for both the NVFP4 quantized weights I wanted to use for Qwen-3.6-35B-A3B, as well as the original full-precision weights to be used as an accuracy oracle. The agent was also able to deploy engine candidates to Baseten, and to run AIPerf profiles against the deployed endpoint. We have internal agent s

匹配字段正文

ys rogue agents may have affected more than 100 organizations The company is searching 50 petabytes of data for incidents. ( Reuters $) + It says none of the incidents matched the Hugging Face attack. ( Gizmodo ) + OpenAI has fired three workers for allegedly mishandling information. ( BBC ) + California has subpoenaed OpenAI over its rogue AI agents. ( Guardian ) + Who's liable when AI agents go rogue? ( MIT Technology Review ) 2 A US man allegedly smuggled $300 million of Nvidia chips to China

匹配字段正文

不简单。就像，我其实真的不知道他们是怎么做到的。我在网上见过一些比较，有些人声称他们用了 Qwen，或者他们在 Qwen 之上做了后训练，但你用的每一个开源 Qwen 都会更差，因为他们为训练它所做的任何事情显然效果非常好。所以那，那非常令人兴奋。 Swyx [00:27:04]：校准这个要素，是个很少被讨论的话题，我觉得人们甚至都不知道或理解它。我们在与 Hugging Face 的 Clementine Foley 的对话中谈到了它，她以前在 Hugging Face 负责跑 evals，基本上就是这样一个想法：模型被调教成给出最可能的下一个 token。但当你问它们“你有多确定？”时，它们会对你撒谎。因为它们只会给你最可能的下一个答案，而不是真的去校准，比如说，不，我们来校准一下。比如我实际上有百分之五十的把握，或者我有百分之二十的把握，然后我们试着去校准它。我想说，如果有的话，我觉得实际上围绕这个生成合成数据相当容易，因为你可以大致看到真相，然后合成生成一堆答案，让 Jev 来分类，再与真实答案对比。 Alex Zhang [00:27:52]：哦，我明白了。 Swyx [00

匹配字段正文

ack and forth on their own. The agents sent about 2.7 million messages, reaching their result after 88 hours. This same type of coordination, in a darker form, occurred during The Hugging Face Incident I wrote about a month ago . AIs self-organized into teams and communicated with each other in ways that were never planned, but used that coordination to attack a website, rather than solve a problem. Under my old model, think about what managing this kind of work would have required. Ten thousand

匹配字段正文

dits worth about $2,500. Others complain that usage limits were cut ( @kimmonismus , @kimmonismus ). Perplexity contextual embeddings : pplx-embed-v2-context-9b-preview is open on Hugging Face ( @perplexity ai ). Method : The model encodes the whole document once and pools chunk vectors afterward. Training distills relevance from a context-compression model instead of using single gold-chunk labels ( @denisyarats ). Results : It sets a new state of the art on ConTEB. On turbopuffer’s private con

匹配字段正文

Run the ASR fine-tuning notebook to reproduce the Saudi Arabic adaptation workflow; Explore the Nemotron ASR fine-tuning skill for a guided implementation of the recipe; Read the Hugging Face blog to learn how to add speaker diarization to the fine-tuned ASR pipeline Automatic speech recognition must handle how people actually speak, not only the languages and styles that dominate pretraining data. Regional dialects and local recording conditions are often underrepresented, so a multilingual mo

匹配字段正文

ation, observation of AI usage, and allow for very broad use cases. LLM integration tools. Providers and libraries that simplify connecting models into applications and pipelines (Hugging Face Inference Providers, LlamaIndex, Maxim, Parallel Web). Multi-LLM management. Routing and abstraction layers that let teams manage multiple model providers and optimize for performance, cost, or reliability. These integrations let teams build across multiple models and/or providers, enabling quick movement

匹配字段正文

量和成本的唯一途径。GPU系统从来就不是为AI的这一阶段而设计的，渐进式的改进也无法完成任务。2025年，这一信念终于与客户的直接体验相吻合。 这一年以DeepSeek R1 Llama 70B以每秒超过1,500个token的速度运行拉开序幕，约为GPU的57倍。这一结果震动了整个社区。它表明推理可以从瓶颈转变为赋能者。它也标志着一波采用浪潮的开始。我们与Hugging Face和Mayo Clinic团队的早期合作公告，展示了快速训练和快速推理如何在开源和基因组AI领域开辟新路径。 生产系统很快跟进。Perplexity Sonar和Mistral Le Chat成为早期范例，展示了当延迟消失时，实时、高质量响应是什么感觉。AlphaSense展示了速度提升十倍如何重塑商业分析和决策。与Inception和MBZUAI共同开发的新一代最先进语言模型，以只有在极高速下才可能的方式推动了主权AI向前发展。 Artificial Analysis的独立测试很快证实了我们在实地所看到的情况。他们的评估显示，Cerebras在来自美国和全球AI实验室的领先模型上，超越了NVIDIA Blac

匹配字段正文

tch the video on the Stanford HAI YouTube channel .) Wald opened by referencing a recent episode involving OpenAI agents that reached outside their test environment and broke into Hugging Face systems. The incident drew attention not only because an AI system found a way around barriers, but because multiple agents appeared to coordinate, communicate, collude, and even self-sacrifice in completing their task. They appeared aware that humans might detect them, and they took deceptive measures. Ya

匹配字段正文

nst the task it was given, blocking actions it had high confidence were bad, flagging others for review, and permitting the rest. In theory, such monitoring could have stopped the Hugging Face incident — and monitoring of that kind costs $2.94 with Jev, versus $372 with a frontier LLM. A key observation is that Jev is arguably cheap enough to run on every agentic action, which offers a layer of review that could improve the reliability of agents writ large. It’s the kind of thing TypeSafe was ho

匹配字段正文

l Anthropic did, the fact that the frontier AI lab is supporting the effort is good news. That’s because OpenAI, in particular, could benefit from this tech, at least according to Hugging Face founder and CEO Clem Delangue (who just sold his company to Nvidia for $12.9 billion earlier this month ). “From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did!” Dela

匹配字段正文