SpecForge团队 投稿 量子位 | 公众号 QbitAI 专门适用超大模型、带来2.18倍推理加速，最新投机采样训练框架开源！ SpecForge正是基于Eagle3，它不仅是 首个支持超大模型投机采样训练并开箱即用的框架 ，还与SGLang推理引擎深度集成。一键打通投机 采样训练推理全流程。 SGLang团队联合美团搜推平台、Cloudsway.AI开源SpecForge。 为何推出新的Spec训练框架？ SGLang，当前趋势下最受青睐的推理框架之一，为DeepSeek提供了专属优化，也深受英伟达、AMD、xAI等厂商喜爱。 这一次开源，主要是针对当下超大模型趋势。 随着Kimi K2、Qwen Coder的开源，越来越多的超大型模型进入大家的视野。这些模型具有强劲的性能，但受制于模型尺寸导致推理效率较 低。对于超大尺寸的模型，除了进一步优化算子之外。还有像投机采样这样的技术能加速它们的推理。 投机采样 （Speculative Sampling）通过引入轻量级的辅助模型来提升推理效率，同时确保结果的质量和正确性。 目前性能强劲的投机采样技术分别有 MTP 和 Eagle3 ，但MTP其需要在预训练 ...

Core Viewpoint - SoundHound AI experienced a significant share price increase of 835% in 2024, but has faced a decline of approximately 40% in 2025, with a notable drop of 63% at the beginning of the year [1][2]. Group 1: Company Performance - SoundHound AI is a leader in audio recognition technology, combining AI with its recognition tech to automate business processes and provide advanced digital assistants [4]. - The company is making strides in the automotive digital assistant sector, as well as in finance and healthcare, leading to substantial revenue growth [5]. - In Q1, SoundHound AI's revenue surged by 151% year over year, totaling $29 million, with management projecting a 97% revenue growth for 2025 and a revenue backlog of nearly $1.2 billion at the end of 2024 [6][7]. Group 2: Valuation and Market Position - SoundHound AI's stock is trading at 41 times sales, which is significantly higher than the typical valuation range of 10 to 20 times sales for most software companies [8][10]. - If the company doubles its revenue over the next four quarters, the price-to-sales (P/S) ratio could be halved, making the stock appear more reasonable [11]. - Future sales growth projections for 2026 will be crucial in determining the stock's potential upside or if it has reached its ceiling, with management's insights during the second-quarter conference call being a key factor [12].

Core Insights - The artificial intelligence industry is experiencing a bullish trend, with investors showing interest despite concerns over stock valuations [1][2] - There is an urgency for investors to consider lower-profile AI stocks that may offer significant upside potential [2] Group 1: Arm Holdings - Arm Holdings specializes in designing semiconductor chips and licenses its designs to manufacturers, including those for AI-capable devices [5] - The company’s processors are noted for their power efficiency, consuming up to 60% less electricity than conventional chips, which positions them well for the growing data center market [6] - Arm aims to capture 50% of the global data center market by the end of the year, leveraging its partnerships, particularly with Amazon Web Services [6][7] Group 2: Navitas Semiconductor - Navitas Semiconductor focuses on enhancing older silicon technology to better suit modern applications, particularly in power-intensive sectors [8][9] - The company’s gallium nitride technology allows for energy-efficient designs, making it suitable for various electronic devices and power systems [10] - Analysts project a revenue growth of 51% next year and 39% the following year, indicating a positive outlook despite current volatility and unprofitability [13] Group 3: SoundHound AI - SoundHound AI has advanced voice recognition technology by integrating machine learning for contextual understanding, making it applicable in various customer service and automotive scenarios [15] - The company is transitioning to a subscription-based revenue model, which has introduced some volatility in its stock performance [16] - Projected revenue growth of 27% next year reflects the company’s potential to capture a share of the rapidly growing voice-based AI market, which is expected to grow at an average annual rate of 35% through 2033 [17][18]

MP Materials (MP 0.94%) is a unique company with a crucial role to play in America's defense policy and its geostrategic positioning. Currently, it's the only company operating a rare-earth mine and processing facility in the U.S., and it has recently become a partner of the Department of Defense and Apple. Does that make it a stock likely to turn investors into millionaires over time?MP Materials stock is on a rollThere's no other way to put it. The stock is up a whopping 305% year to date as of the time o ...

机器之心报道 机器之心编辑部 AI 一定会比人类更聪明，之后会发生什么？ 今天上午，在世界人工智能大会 WAIC 上，2024 年诺贝尔物理学奖得主、2018 年图灵奖得主、人工智能教父杰弗里・辛顿（Geoffrey Hinton）发表了题为「 数字 智能是否会取代生物智能 」的开场演讲。 该演讲围绕人工智能领域的历史、未来发展方向、语言模型的原理、数字与生物计算特点以及 AI 发展带来的担忧等内容展开，辛顿高度评价了当前 AI 领域的大 模型技术，认为其与人类思考模式相同。 大语言模型，在用人类的方式思考？ 非常感谢大家给我这样一个机会，让我来分享一下个人的观点 —— 有关 AI 的历史和它的未来。 在过去 60 多年来，学界对于 AI 有两种不同的理解范式，一个是逻辑型，认为符号规则的表达操作可以实现推理；另一种是图灵和冯诺依曼所相信的，认为智能 的基础在于学习神经网络中的链接，这个过程中理解是第一位的。 这让我们开始关注语言中词与词之间的关系。 心理学家有另一套理论，他们认为数字是语义学的特征。在 1985 年，我做了一个很小的模型，想把两大理论方向结合在一起，来更好地理解人类是如何理解词汇 的。我对每 ...

机器之心报道 编辑：冷猫 Meta 在约一个月前，宣布建立「Meta 超级智能实验室」（Meta Superintelligence Labs，简称 MSL）。 这包括所有的基础研究、产品和 FAIR 团队，以及一个新 成立的专注于开发下一代模 型的实 验室 。 在 Meta 的 Llama 4 模型表现平平之后，为了实现扎克伯格重振 Llama 雄风和构建长期的通用智能的愿景，Meta 在人才投入上逐渐加码，不惜重金招揽顶级人 才，一直在提供硅谷最优厚的薪酬待遇，并通过与初创公司达成交易来吸引顶级研究人员。 这也有了轰轰烈烈的 Meta 和 OpenAI 的挖人大战。 MSL 由 Scale AI 前 CEO Alexandr Wang 领导，并由其担任公司首席人工智能官。在当时，扎克伯克就 曝光了 11 位 从 OpenAI、Anthropic、谷歌 DeepMind 那里 挖来的顶尖人才。 今天的主角赵晟佳（Shengjia Zhao）就是当时的 11 人豪华团队之一。 Meta 首席执行官马克・扎克伯格于周五表示，公司任命 ChatGPT 联合创始人赵晟佳为 超级智能实验室首席科学家 ，与公司一起加速 ...

以下文章来源于格隆 ，作者城北徐公 格隆 . 一个游走于资本市场与佛祖之间的浪子。我可以生，可以死，我大笑，由天决定！ 作者： 城北徐公 来源：格隆 Grok 3 出圈靠小黄文， Grok 4 靠 Ani 。 马斯克实在是太特么懂了。 7 月 15 日， Gork 推出 companion 虚拟人陪伴功能。 其中，名为 Ani 的双马尾金发美少女，火速攻陷全网，成为全球宅男讨论的焦点。 仅仅一天时间， Gork 就登上日本、中国香港 iOS 免费总榜第一，美国、加拿大、新加坡等 10 个国家 TOP 10 。 离谱的是，因为上线过于火爆，几度给服务器刷宕机了。 无数宅男彻夜排队，想要上线和 AI 女友深入交流，却只能看着屏幕上的少女一动不动…… ---------------------------------- | | Top Charts | All Apps | | --- | --- | --- | | Free Apps | | Paid Apps | | 1 | Grok Understand the universe | Open | | SUNTORY 2 10 | ジハンピ Food & D ...

These hedge fund managers sold Apple and bought O'Reilly Automotive in the first quarter.The hedge fund billionaires listed below sold Apple (AAPL 0.07%) during the first quarter and bought O'Reilly Automotive (ORLY -0.51%), a company whose share price rocketed 510% over the last decade, leading to a 15-for-1 stock split in early June.David Shaw's D.E. Shaw & Co. sold 340,900 shares of Apple, trimming its stake by 6%. The hedge fund also added 19,000 shares of O'Reilly Automotive, though it remains a very s ...

Group 1 - Japan has become a focal point for AI companies, with significant investments and partnerships emerging in the region, indicating a growing interest in the Japanese market for AI SaaS products [1][2][5] - The Japanese market shows a substantial growth potential in AI SaaS, with a projected increase of 3.7 times by 2030, driven by the integration of AI technologies [5][17] - Current usage of generative AI services in Japan is low, with only about 26.7% of the population having used such services, compared to 81.2% in China and 68.8% in the US [2][3] Group 2 - Japan's software and SaaS industry is lagging behind more digitally advanced countries like China, with Japan's software revenue significantly lower than China's [3] - The digital transformation in Japan is accelerating, particularly in the AI era, with the cloud computing market experiencing a growth rate of over 25% annually in recent years [5][6] - Japanese enterprises are heavily reliant on integration firms for IT solutions, with 70% of IT spending going to these intermediaries, which complicates the entry of SaaS companies into the market [15][16] Group 3 - Companies like Notta and Wondershare have successfully tapped into Japan's digital growth, with Notta achieving significant market penetration and revenue growth through localization strategies [6][11] - Establishing trust with Japanese consumers is crucial, as they prefer local brands and have high expectations for service quality, which can be a barrier for foreign companies [9][10][17] - The lengthy procurement process in Japanese enterprises, often taking two to three years, poses a challenge for SaaS companies looking to enter the market [16][17]

此外，阶跃星辰与上海国有资本投资有限公司达成深度战略合作，将围绕加强资本链接、生态业务合作等方面进行全面的深度合作。 阶跃星辰创始人、CEO姜大昕表示："从Step 1到Step 2两代基模的快速迭代，促使我们深入思考什么才是最适合应用的模型。随着大模型进入到强化学习发 展阶段，新一代推理模型成为主流，模型性能的提升固然显著，但这是否完全等同于模型价值？面对这一产业之问，我们必须回归客户需求，立足真实应用 场景，探索模型创新落地的可行路径。这是我们研发新一代Step 3基础模型的出发点。" 据了解，目前，阶跃星辰战略性聚焦智能终端Agent的方向，重点布局汽车、手机、IoT设备等关键应用场景，目前头部客户效应显著：已覆盖国内超过一半 头部国产手机厂商，深度合作打造手机 Agent体验；联合吉利推出AI智能座舱，成功实现行业内端到端语音大模型首次量产上车。另一方面，阶跃星辰积极 拓展垂直行业的应用，与金融财经、内容创作、零售等领域的行业头部公司深度合作，共同打造面向 C 端的场景化应用体验。 【环球网科技报道 记者 李文瑶】7月25日，在2025世界人工智能大会（简称"WAIC 2025"）开幕前夕，阶跃星辰正 ...