WEST PALM BEACH, Fla., July 11, 2025 (GLOBE NEWSWIRE) -- Solitron Devices, Inc. (OTC Pink: SODI) (“Solitron” or the “Company”) is pleased to announce fiscal 2026 first quarter results. FISCAL 2026 FIRST QUARTER HIGHLIGHTS – Net sales decreased 32% to approximately $2.70 million in the fiscal 2026 first quarter versus $3.97 million in the fiscal 2025 first quarter.Net bookings increased 37% to $2.80 million in the fiscal 2026 first quarter versus $2.04 million in the prior year first quarter.Backlog increas ...

允中 发自 凹非寺 量子位 | 公众号 QbitAI 强化学习改变了大语言模型的后训练范式，可以说，已成为AI迈向AGI进程中的关键技术节点。 然而，其中 奖励模型 的设计与训练，始终是制约后训练效果、模型能力进一步提升的瓶颈所在。 当前，大模型在 Next Token Prediction 和 Test-time Scaling 两种扩展范式下，通过大规模的数据和模型扩展，实现了能力的持续跃 升。但相比之下，奖励模型缺乏系统性的预训练和扩展方法，导致其能力难以随计算量增长而持续提升，成为阻碍强化学习链路进一步扩展的 短板。 如何解决？ 现在，来自上海人工智能实验室的研究团队提出了一种新的思路： 回答三： 彩虹是阳光通过空气中的水滴折射和反射后形成的。生活中我们经常在雨后或喷泉、水幕附近看到彩虹，有时候还会看到双彩虹甚至三重彩 虹呢！不过很可惜，彩虹本身只是光学现象，没法真正走近摸到。 对于这种开放问题，不同的用户可能会偏向不同风格的回复。此时，传统奖励模型的"绝对偏好"无法灵活应对不同的定制化场景。 而POLAR 只需要根据不同的参考回复，即可为三种回答给出不同的偏序关系，无需重新训练奖励模型 。 他们找到 ...

Latest achievements highlight over 60 years of space heritage and commitment to high-reliability, radiation-tolerant technology for next-generation satellite and avionics systems.CHANDLER, Ariz., July 10, 2025 (GLOBE NEWSWIRE) -- Continuing to support the evolving needs of space system developers, Microchip Technology (Nasdaq: MCHP) has announced two new milestones for its Radiation-Tolerant (RT) PolarFire® technology: MIL-STD-883 Class B and QML Class Q qualification of the RT PolarFire RTPF500ZT FPGA and ...

回顾大模型（LLM）的成功之路，是利用 Next Token Prediction 的形式统一了所有任务，解决了任务形式不同导致无法泛化的难题。而奖励模型（RM）的设计仍 然在重蹈传统方案的老路，即为特定场景标注偏好数据，训特定场景的 RM。因此，是否可以仿照 LLM 的成功之路，重新设计 RM 的训练范式呢？消除 RM 的 "打分标准"，就像消除 LLM 的 "任务形式" 一样，找到一个 脱离于 "打分标准" 之外的更本质的优化目标函数 来进行预训练 ，从而达到真正的通用性。 机器之心发布 机器之心编辑部 在大语言模型后训练阶段，强化学习已成为提升模型能力、对齐人类偏好，并有望迈向 AGI 的核心方法。然而，奖励模型的设计与训练始终是制约后训练效果的 关键瓶颈。 目前，主流的奖励建模方法包括 " 基于偏好的奖励建模 "（Preference-based Reward Modeling）和 " 基于规则的验证 "（Rule-based Verifier）两种方法。 其中，"基于偏好的奖励建模" 一般利用标注的偏好对数据来训练奖励模型，这种方法存在着诸多局限。首先，高质量偏好数据的 获取成本极高，难以大规模扩展 ...

Core Insights - SoundHound AI (SOUN) has experienced a significant stock price increase of 55% over the past three months, outperforming major market indices and sectors [1][9] - Despite this growth, the stock is still 49.1% below its 52-week high of $24.98, but 222.6% above its 52-week low of $3.94, raising questions about further upside potential [2] Revenue Growth and AI Innovation - In Q1 2025, SoundHound reported revenues of $29.1 million, reflecting a year-over-year growth of 151%, driven by successful acquisitions [6] - The proprietary Polaris foundation model supports multi-modal, multi-lingual voice AI, enhancing the performance of the Amelia 7.0 platform [7] - SoundHound is expanding its voice commerce capabilities, with over 13,000 restaurant locations using its AI and pilots with major automotive OEMs [8] Diversification and Enterprise Traction - SoundHound has diversified its customer base beyond automotive AI, achieving strong renewals and expansions in sectors like healthcare, telecom, and financial services [10] - The Amelia 7.0 platform allows businesses to deploy AI agents for complex workflows, positioning SoundHound as a key player in the Agentic AI market [11] Operational Metrics - Active cloud users increased by over 50% in Q1, with a tenfold rise in restaurant activations year over year [12] - SoundHound's AI platforms handle over 2 billion voice queries per quarter, indicating strong operational momentum [12] Valuation Concerns - SoundHound's forward Price-to-Sales (P/S) ratio stands at 27.19X, significantly above the industry average of 19.56X, highlighting valuation concerns [15] - The company remains unprofitable on a GAAP basis, with a non-GAAP net loss of $22.3 million in Q1 [18] Competitive Landscape - SoundHound faces intense competition from tech giants like Alphabet, Amazon, and Apple, which dominate the voice AI market [22] - The company must focus on niche, embeddable voice solutions to differentiate itself from larger competitors [23] Future Outlook - Management projects full-year 2025 revenues between $157 million and $177 million, with a commitment to achieving adjusted EBITDA profitability by the end of 2025 [21] - The stock's recent rally reflects confidence in SoundHound's scalable AI platform and vertical expansion, but elevated valuation and operating losses present risks [24][25]

Core Insights - Richardson Electronics has entered a global technology partner agreement with Pakal Technologies to supply advanced 650V and 1200V silicon-based power switches, enhancing efficiency for customers worldwide [1][3] - Pakal Technologies has introduced the IGTO(t) - a new high-voltage silicon power semiconductor, which offers up to 30% lower conduction losses compared to modern IGBTs, making it suitable for various applications including electric vehicles and renewable energy systems [2][8] - The partnership aims to leverage Richardson Electronics' global reach and experience in launching high-performance power products to bring Pakal's innovative technology to market [3] Company Overview - Richardson Electronics, Ltd. is a global manufacturer specializing in engineered solutions, green energy products, and power conversion components, with over 50% of its products manufactured in the U.S. and Germany [4] - The company has a strong focus on providing specialized technical expertise and engineered solutions across various markets, including alternative energy, healthcare, and industrial sectors [4][6] - Pakal Technologies is a pioneering power semiconductor company based in San Francisco, known for its proprietary IGTO(t) architecture that supports next-generation electrification in multiple sectors [8]

Polaris团队 投稿 量子位 | 公众号 QbitAI 4B模型的数学推理能力和顶尖商业大模型差在哪里？ 香港大学NLP团队联合字节跳动Seed、复旦大学发布名为 Polaris 的强化学习训练配方： 通过Scaling RL，Polaris让4B模型的数学推理能力（AIME25上取得79.4，AIME24上取得81.2） 超越了一众商业大模型 ，如Seed-1.5- thinking、Claude-4-Opus和o3-mini-high(25/01/31)。 并且，Polaris-4B的轻量化允许在消费级显卡上部署。 详细的blog、训练数据、模型和代码都已 全部开源 ，链接可见文末。 围绕待训练模型进行参数配置 之前的RL训练配方，如DeepScaleR，已经展示了Scaling RL在较弱基模型上强大的效果。 但对于目前最前沿的开源模型（如Qwen3），Scaling RL是否也能复现如此显著的提升呢？ Polaris的研究团队给出了明确回答：可以！ 具体来说，Polaris通过仅仅700步的RL训练，成功地 让Qwen3-4B在数学推理任务上接近了其235B版本的表现 。 只要方法得当，RL还存在 ...

Key Takeaways MSFT offers a full-stack data platform spanning storage, analytics and AI tools. Enterprise adoption of Fabric and OneLake highlights MSFT's growing platform traction. SNOW is deepening AI integration and expanding its unified data experience for enterprises.Snowflake (SNOW) and Microsoft (MSFT) are major players in the expanding cloud data platform market. Snowflake provides a scalable data warehouse with integrated AI and machine learning. Microsoft delivers a suite of data services on Azu ...

Covenant Relief Enhances Financial Flexibility and Supports Long-Term GrowthMINNEAPOLIS, July 2, 2025 /PRNewswire/ -- Polaris Inc. (NYSE: PII), the global leader in powersports, today announced that on June 27, 2025 it executed an amendment to its existing credit facilities, originally dated July 2, 2018. In addition, the company also fully repaid all $350 million of outstanding senior notes due 2028. The debt prepayment was funded through revolving loans under the company's senior credit facilities.The ame ...

招商通知 第三届全球手术机器人大会 第二届全球医疗科技大会 心未来 2025年7月1日，心脏电生理创新领军者 心航路医学 （ EnChannel Medical ）宣布，已完成对美国公司 Acutus Medical 的 AcQMap™高分辨率成像与标测平台 核心资产 收购。 在持续性房颤治疗路径尚不明朗的全球背景下，这一交易或将重塑复杂心律失常标测技术的竞争格局，并 为电生理介入提供更精准的可视化工具。 心航路医学首席执行官冯君 表示 ： "这项战略性收购将加速我们实现房颤高效、精准可视化治疗的使 命。借助已有成果并结合我们的自主技术，我们正在重塑持续性房颤诊断与治疗的未来。" # 交易概况：资产收购而非企业合并的 "轻型整合" 本次收购并非传统意义上的企业合并，而是对特定技术资产的 "点对点"收购。 根据心航路方面披露， 交易涉及内容包括 AcQMap ™平台相关的关键知识产权、质量与法规文档、部分 系统设备，以及用于研发用途的导管产品 ，但不包括 Acutus Medical 整体业务或员工团队的并入。 这类 "轻型整合"交易日渐成为医疗器械行业创新企业优化资产结构、提升技术占有率的常用方式。与资 本密 ...