RL

Search documents
X @Ashok Elluswamy
Ashok Elluswamy· 2025-07-07 19:50
RT TESLARATI (@Teslarati)Austin's Robotaxi service gets a review from a non-Tesla "influencer" 🤖🚗⚡️What's better than a ~$15 Uber ride? A $4.20 Robotaxi ride in a driverless Model Y. The rider even mentioned that it's smoother than typical Ubers he has ridden in the past.https://t.co/NeVLjml5ex ...
How Domestic and World Orders Change
Principles by Ray Dalio· 2025-07-07 14:36
Over the last few years, three big things that hadn't happened in my lifetime prompted me to do this study. First, countries didn't have enough money to pay their debts, even after lowering interest rates to zero. So, their central banks began printing lots of money to do so.Second, big internal conflicts emerged due to growing gaps in wealth and values. This showed up in political populism and polarization between the left who want to redistribute wealth and the right who want to defend those holding the w ...
VolitionRX (VNRX) Earnings Call Presentation
2025-07-07 11:45
Company Overview and Strategy - VolitionRx focuses on advancing low-cost, early detection, and treatment monitoring tests in cancer and sepsis[7] - The company's strategy involves licensing its IP to major players, following a low CapEx/low OpEx model similar to its Nu.Q® Vet business[11, 20] - Volition aims to be cash neutral on a full-year basis in 2025, balancing income with expenditure[87] Veterinary Commercial Progress - The Nu.Q® Vet Cancer test is available in over 20 countries[20] - Approximately 120,000 Nu.Q® Vet Cancer tests and test components were sold in 2024[20, 39] - The company has received $23 million in upfront and milestone payments to date, with an additional $5 million milestone payment anticipated in 2025[20] Expansion into Human Diagnostics - Volition is targeting licensing deals in the human sepsis and cancer space in 2025[10] - The company sees a $4 billion opportunity in lung cancer screening, prognostics, and MRD, and a $1 billion+ opportunity in sepsis testing and monitoring[21] - First revenue was recorded for CE-marked product in Q1 2025[85] Financial Performance - Volition recorded approximately $0.25 million in revenue in Q1 2025, a 44% increase over the first quarter of the prior year[87] - Net cash used in operating activities averaged $1.4 million a month, almost 50% lower than the first quarter of 2024[87]
大模型刷数学题竟有害?CMU评估20+模型指出训练陷阱
量子位· 2025-07-07 06:13
henry 发自 凹非寺 量子位 | 公众号 QbitAI 学好数理化,走遍天下都不怕! 这一点这在大语言模型身上也不例外。 大家普遍认同:具备更强数学能力的模型往往也更智能。 但,常识就是用来打破的。 最近,来自CMU的团队发现,一些数学好的模型并没有将它们的"天赋"带到其他更加通用的领域。 研究发现, 只有用强化学习(RL)训练的模型才能将数学推理技能广泛迁移到其他任务上。而用监督微调(SFT)训练的模型则表现出有限 的迁移甚至没有迁移。 网友直呼:又一个 苦涩的教训(bitter lesson) 。 这数学题,不做也罢? 很明显,人们训练大模型并不只是让它来做数学题的。 研究者之所以热衷于提高模型的数学表现,是因为希望它能够把数学那里学到的严密逻辑应用到其他更广泛的领域。 但在此之前,我们有必要知道,对于一个大模型,专门优化数学推理(math reasoning),它在其他任务(推理任务、非推理任务)上会变 得更好,还是更差? 换句话说: 做数学推理训练,会不会帮助或者损害模型在其他领域的能力? 为了解决这一疑问,研究评估了20多个模型在数学推理、其他推理任务(包含医学推理、医学推理、智能体规划)和非推 ...
X @The Economist
The Economist· 2025-07-06 22:40
While the Senate parliamentarian’s judgments are not enforceable, they almost always hold. We explain why https://t.co/Jusnh2aakA ...
X @The Economist
The Economist· 2025-07-06 00:40
While the Senate parliamentarian’s judgments are not enforceable, they almost always hold. We explain why https://t.co/wggiw923BD ...
X @Tesla Owners Silicon Valley
Tesla Owners Silicon Valley· 2025-07-05 17:58
Driverless Tesla Robotaxi https://t.co/AlPNghD8pe ...
X @The Economist
The Economist· 2025-07-05 00:20
Speaking out on world affairs is in vogue and much of it is doubtless heartfelt and sincere. But it can also have other motives—and unintended consequences https://t.co/5BBOscQO3K ...
图像目标导航的核心究竟是什么?
具身智能之心· 2025-07-04 12:07
点击下方 卡片 ,关注" 具身智能 之心 "公众号 作者丨 GianlucaMonaci 编辑丨具身智能之心 本文只做学术分享,如有侵权,联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货,欢迎加入国内首个具身智能全栈学习社区 : 具身智能之心知识星球 (戳我) , 这里包含所有你想要的。 研究背景与核心问题 图像目标导航(Image goal navigation)需要两种关键能力:一是核心导航技能,包括检测自由空间、障碍物 及基于内部表征做决策;二是通过比较视觉观察与目标图像计算方向信息。当前主流方法要么依赖专门的图 像匹配,要么预训练计算机视觉模块进行相对位姿估计。 研究聚焦于一个关键问题:该任务能否如近期研究所说,通过强化学习(RL)对完整智能体进行端到端训 练来高效解决?若答案为肯定,其影响将超出具身AI领域,有望仅通过导航奖励来训练相对位姿估计模型。 核心研究内容与方法 关键架构选择 研究探讨了多种架构设计对任务性能的影响,核心在于如何支持图像间的隐式对应计算,这对提取方向信息 至关重要。主要架构包括(figure 2): 实验设计 Late Fusion :分别编码观察图像和目标图像 ...