Workflow
Reasoning
icon
Search documents
‘The Nvidia Way’ author Tae Kim: Jensen Huang always positions Nvidia ahead of the next big trend
CNBC Television· 2025-07-11 18:13
So, how did Jensen Wong and Nvidia get to where they are now. And what will it take to keep the momentum going. Let's bring in Tay Kim, senior technology writer at Barren and author of The Nvidia Way, a history of the company's unexpected rise to the top of the tech world.And Tay, it's great to have you back. Let's just set the scene here. A lot of people now asking like, come on with this guy, this company, you know, also because Bitcoin's at an all-time high this week, they feel like they're related.And N ...
阿里多模态推理模型开源!精准捕捉视频隐藏信息,三大杀手锏让AI更懂“人情世故”
Sou Hu Cai Jing· 2025-07-09 00:28
智东西 编译 | 程茜 编辑 | 心缘 AI能听懂你的"话外音"了? 智东西7月8日消息,近日,阿里通义实验室开源多模态推理模型HumanOmniV2。 HumanOmniV2通过引入强制上下文总结机制、大模型驱动的多维度奖励体系,以及基于GRPO的优化训练方法,实现了对多模态信息的全面理解,使得模 型不会错过图像、视频、音频中的隐藏信息,一定程度上规避其在全局上下文理解不足和推理路径简单上的问题。 如在生成最终答案前,模型会输出一个标签内的上下文概括,系统性分析多模态输入内容中的视觉、听觉、语音信号,为后面的推理过程提供依据。如下图 提问"女人为什么翻白眼",HumanOmniV2基于视频、音频等信息给出正确答案"她的翻白眼更像是对一个潜在敏感话题的夸张、俏皮的反应,非对其他人表 示不满"。 现阶段HumanOmniV2已开源。阿里通义团队还推出包含633个视频和2689个相关问题的评测基准IntentBench,在此之上,HumanOmniV2准确率达到 69.33%。 Hugging Face:https://huggingface.co/PhilipC/HumanOmniV2 IntentBench评 ...
DeepSeek推理最高提速6倍!开源研究:加装「思维进度条」,计算量减少30%
量子位· 2025-07-07 06:13
不圆 发自 凹非寺 量子位 | 公众号 QbitAI DeepSeek推理要详细还是要迅速,现在可以自己选了? 来自特拉维夫大学的研究团队开发出了一种新方法,可以 监控和控制LLM中的思考路径长度 。 超频能够减少不必要的推理步骤,使模型更快地得出结论,同时避免因过度推理导致的性能下降。 该模型已在gitHub上开源。 给LLM的推理任务装上进度条,还能控制推理的深度、调整推理速度。 加速后的模型和原模型相比, 使用的token数减少了近6倍,且都得出了正确答案 。 LLMs在显示结构化推理时,会隐式跟踪其在思考阶段的相对位置,并通过隐藏状态编码这一信息。 而论文提出了一种"思维进度向量"(Thinking Progress Vector, TPV ),可用于实时预测模型在推理阶段的相对位置,并通过可视化进度条 展示模型的推理动态。 通过干预TPV,可以加速或减速模型的推理过程,实现"超频"(overclocking)和"降频"(downclocking)。 方法:实时监控并控制推理深度 在有效推理学习过程中,模型必须 隐式地学习跟踪其思考阶段进度 ,并保持对例如距离最终答案有多近的估计。 由于进度跟踪依赖于 ...
重温《英伟达GTC 2025》:挖掘AI算力需求预期差?
2025-07-07 00:51
重温《英伟达 GTC 2025》:挖掘 AI 算力需求预期差? 20250706 摘要 美股算力领域表现突出,主要由推理和训练需求共振驱动,需关注大模 型和应用,而非仅依赖产业链数据。GTC 大会参会人数增加,AI 产业人 士占比提升,表明其对 AI 产业的重要性增强,蕴含大量信息差和预期差。 算力需求空间与 TOKEN 量密切相关,计算需求不断增加推动了这一趋 势的发展。海外算力公司的涨幅显著,仅靠传统业绩思路无法解释,需 深入分析 TOKEN 量如何影响计算需求,以及这些因素如何驱动未来趋 势。 Agentic AI 是推理模型衍生出的 AI 范式,强调任务分布执行和规划, 以完成某个任务为终极目标,通过拆解、分布、规划和执行来实现,能 够处理连续、多步骤的复杂或简单任务。 黄仁勋指出,现在不仅有预训练阶段,还有后训练(post training)和 测试时间(test time),这三个阶段都存在算力需求通胀,因此现在有 三条 skin law 曲线。 Q&A 全球 AI 算力跟踪的现状如何?与以往相比有哪些变化? 当前全球 AI 算力的跟踪方式与以往有显著不同。过去主要通过产业链数据进行 跟踪,但这种 ...
喝点VC|红杉美国对谈OpenAI前研究主管:预训练已经进入边际效益递减阶段,其真正杠杆在于架构的改进
Z Potentials· 2025-07-04 03:56
图片来源: Sequoia Capital Z Highlights Bob McGrew , OpenAI 前首席研究官,主导推动 GPT ‑ 3 、 GPT ‑ 4 以及内部称为 o1/o3 模型的研发,提出预训练( pre-training )、后训练( post-training )和推理( reasoning )的 " 三位一体 " 模型。现为多家 AI 初创企业的顾问或投资人,持续推动 AGI 的落地。本次访谈视频由 Sequoia Capital 在 2025 年 6 月 17 日发布,和 Bob 共同探讨了 从模型训练重点、 Agent 和机器人的未来发展、 AI 时代的教育心得与管理经验等主题,洞察人工智能的发展轨迹,并 指出初创企业依然可以挖掘并构建可持续竞争优势的领域。 预训练、后训练和推理,未来如何发展? Stephanie Zhan : 欢迎来到 Training Data 。今天我们非常高兴邀请到 Bob McGrew——OpenAI 前首席研究官,带我们深入探讨 frontier AI 的幕后发 展。 Bob 分享了预训练( pre-training )、后训练( post-tr ...
X @Decrypt
Decrypt· 2025-06-28 22:40
We break down China’s new open-source reasoning model, MiniMax-M1: real benchmarks, hidden tradeoffs, and how it stacks up against competitors. https://t.co/sjgqJ13pgk ...
From Quora to Poe: Adam D'Angelo on Building Platforms for LLMs and Agents | LangChain Interrupt
LangChain· 2025-06-27 16:44
Adam D'Angelo, co-founder and CEO of Quora, shares insights from building Poe—the AI platform that gives users access to multiple language models and agents under one subscription. He reveals surprising patterns in consumer AI usage, explains how Poe's bot creators are earning millions annually, and discusses why reasoning models are driving growth. D'Angelo also explores the unique challenges of building in AI's rapidly changing landscape, where planning cycles have shrunk from years to just two months. Wa ...
The AI-boom's multibillion-dollar blind spot: Reasoning models hitting a wall
CNBC Television· 2025-06-27 12:49
AI reasoning models were supposed to be the industry’s next leap, promising smarter systems able to tackle more complex problems and a path to superintelligence. The latest releases from the major players in artificial intelligence, including OpenAI, Anthropic, Alphabet and DeepSeek, have been models with reasoning capabilities. Those reasoning models can execute on tougher tasks by “thinking,” or breaking problems into logical steps and showing their work. Now, a string of recent research is calling that i ...
AI's reasoning blind spot
CNBC Television· 2025-06-26 16:26
Tech stocks continuing to rally, powering this market to a record high on the S&P. The NASDAQ 100 hitting its own record high. Nvidia, Microsoft, Broadcom at or near all-time highs on their own.But could the market be overlooking a major risk popping up in the next leg of the AI trade. Dear Drabosa digging into that in today's tech check, what are you worried about, Dearra. Well, this is what I'm worried about.AI's next big promise is reasoning. These are models that can think through problems, make plans, ...
X @TechCrunch
TechCrunch· 2025-06-26 16:18
Meta hires key OpenAI researcher to work on AI reasoning models | TechCrunch https://t.co/gwps46nZ9O ...