算法

Search documents
“复刻”幻方量化打造Deepseek 量化私募基金念空在大模型底层技术研发取得突破
经济观察报· 2025-06-03 11:17
随着AI大模型迭代升级,如今量化私募基金对AI大模型底层技 术的研发布局,日益侧重算法优化。在这个过程,产学研的结 合将是他们在大模型底层技术研发方面取得突破的"捷径"。 作者:陈植 封图:图虫创意 5月以来,全球大模型研发公司在大模型语义理解、多模态等方面的"较劲"悄然升级。 中国深度求索(DeepSeek)公司表示,DeepSeek R1模型已完成小版本升级,令大模型的思维深 度与推理能力显著提升。 国内量化私募基金念空科技与上海交通大学计算机学院开展合作,提出一种全新的大模型训练框架 (SASR),并发表论文投向全球顶级人工智能会议期刊NIPS。 念空科技创始人王啸在6月3日接受本报记者专访时表示,这项全新的大模型训练框架(SASR), 在GSM8K任务中,在仅使用1.5B模型的情况下,准确率就超过了80%,接近GPT-4o的表现;而在 KK逻辑推理任务中,其准确率比GPT-4o还高出约9个百分点。SASR让通用大模型变得更"聪明"。 他告诉记者,当前大模型技术的训练框架,主要围绕监督微调(SFT)和强化学习(RL),所谓监督微 调(SFT)就是不断给大模型输入资料和案例进行监督训练,相当于"刷题"; ...
“复刻”幻方量化打造Deepseek 量化私募基金念空在大模型底层技术研发取得突破
Jing Ji Guan Cha Wang· 2025-06-03 06:57
Core Insights - The competition among global large model development companies has intensified, particularly in semantic understanding and multimodal capabilities since May [2] - Domestic quantitative private equity funds are also entering the race, achieving breakthroughs in AI large model foundational technology [2][5] - A new training framework (SASR) proposed by NianKong Technology in collaboration with Shanghai Jiao Tong University has shown promising results, achieving over 80% accuracy on the GSM8K task with a 1.5B model, nearing GPT-4o's performance [2][4] Group 1: Training Framework and Algorithm Optimization - The current training frameworks for large models primarily focus on Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), with the challenge being to optimize the balance between these two methods [3][8] - The new training framework aims to dynamically adjust the relationship between SFT and RL, allowing the model to become "smarter" without increasing data volume [3][9] - The innovative training framework has been applied in quantitative investment strategy development, achieving approximately 80% market prediction accuracy compared to traditional models [4][13] Group 2: Industry Trends and Collaborations - Many quantitative private equity firms are establishing AI Labs to focus on foundational technology research for large models, emphasizing algorithm optimization [6][11] - The integration of academic research and private equity expertise is seen as a shortcut to breakthroughs in large model foundational technology [5][11] - The emergence of smarter large models with lower parameter counts but superior overall capabilities is attributed to innovations in training frameworks and algorithm optimization [10] Group 3: Future Directions and Challenges - The ability of large models to become "smarter" in various vertical fields depends on high-quality data and effective training modes [12] - NianKong Technology aims to empower large models to excel in more vertical fields, enhancing China's competitiveness in the global AI landscape [14]
晏庆华:推动算法向善向上、保障新就业群体权益需协同治理
Zhong Guo Xin Wen Wang· 2025-05-30 12:22
中新网北京5月30日电 (记者 夏宾)近年来,国家陆续出台了《关于维护新就业形态劳动者劳动保障权益 的指导意见》《互联网信息服务算法推荐管理规定》等政策文件,关注平台企业的新就业形态劳动者权 益保护。 中国物流与采购联合会网络事业部主任、物流信息服务平台分会秘书长晏庆华近日在北京出席互联网货 运行业算法研讨会时称,上述文件明确提出算法取中原则,要求平台企业优化算法规则,保障劳动者权 益,要推动算法向善向上,保障新就业群体权益问题,不能仅靠政府监管或企业自律,需要构建政府、 行业协会、平台、工会、劳动者等多方共同参与的体系,实现共同参与协同治理。 为深化多元共治,实现行业共融,晏庆华建议,强化政府统筹,完善政策法规,建议政府出台算法治理 专项法规,明确各方权责边界。发挥工会作用,构建协商平台,工会可探索建立与平台企业的常态化协 商机制,将算法优化、权益保障等纳入协商范围,并积极吸纳新就业形态劳动者入会,通过集体协商, 参与算法审计等方式维护权益。 同时,发挥协会作用,促进行业自律。平台应积极参与行业学会组织的活动,签署自律公约,承诺在算 法设计和应用中遵循公平、公正、透明的原则,行业协会可制定算法透明化和向善向 ...
豪掷22.87亿,大疆竞得深圳湾地块
Guo Ji Jin Rong Bao· 2025-05-30 11:54
深圳湾超级总部基地将再迎企业入驻。 据深圳公共资源交易中心最新公告,备受关注的T207-0066宗地使用权出让结果日前揭晓,全球无人机巨头深圳市大疆创新科技有限公司(以下简称"大 疆")以22.87亿元的成交价成功竞得该地块。 此次土地出让引发市场高度关注,外界普遍猜测这或与大疆未来总部建设规划相关。5月30日,《国际金融报》记者就地块具体用途及建设计划向大疆 方面求证,公司相关负责人回应称,目前除政府公开披露的信息外,暂无更多细节可以透露。 根据《项目产业发展监管协议》,该地块将建成"智能航空系统产业生态全球总部项目",整合研发、测试、展示、运营等全产业链环节,重点突破无人 机通信导航、人工智能算法、低空物流系统等核心技术。2024-2030年7年间,竞得方累计纳入深圳市统计核算的工业产值须不少于1960亿元,服务业营收不 低于840亿元,增加值不少于560亿元,形成地方财力不少于32.65亿元,进出口额不低于350亿元,研发投入不低于300亿元。 (大疆门店 蔡淑敏摄) 根据公开信息,大疆成立于2006年,总部位于深圳,由80后创始人汪滔带领团队从直升机飞行控制系统起步,逐步发展为全球无人机行业龙头,目前 ...
突发!600190停牌预退市!央国企市值管理放大招,这16只潜力股要火?
Zheng Quan Shi Bao Wang· 2025-05-29 23:53
Group 1 - *ST Jinguang (600190) will be subject to mandatory delisting due to major violations, with trading suspension starting from May 30, 2025 [2] - The company was found to have inflated profits through false trade activities and misreporting, with inflated profits of 36.10 million yuan in 2022, 68.09 million yuan in 2023, and 15.38 million yuan in Q1 2024 [2] - The stock will enter a delisting transition period for 15 trading days after the delisting decision, followed by a five-day period before being officially delisted [2] Group 2 - Central state-owned enterprises (SOEs) are accelerating market value management, with several companies releasing market value management systems [4] - Local and enterprise-level efforts are expected to enhance the pace of market value management, particularly for low-priced and low-market-cap companies, by the second half of 2025 [4] - 16 potential stocks have been identified based on criteria of stock price below 8 yuan and market capitalization below 8 billion yuan [5] Group 3 - Jishi Media, a local state-owned enterprise, is focusing on strategic emerging industries and digital transformation to enhance its asset scale and profitability [5][6] - The company reported a revenue of 499 million yuan in Q1, a year-on-year increase of 23.14%, marking the highest growth rate since its listing [7] - The company is actively collaborating on digital infrastructure projects to boost its capabilities in various sectors [6] Group 4 - Two potential stocks, Liyuan New Energy and ST Chenming, have seen continuous institutional buying over the past two years, with significant net purchases recorded [8] - Liyuan New Energy focuses on wind and solar power generation, with a notable increase in social security fund holdings [8]
小红书算法又变了....
Sou Hu Cai Jing· 2025-05-29 15:00
之前的"清朗·算法治理专项行动" ,直接给小红书这类平台上了紧箍咒。以前平台可以闭眼推流量,现 在必须公开算法逻辑、破除信息茧房、严控低俗内容。 "昨晚发的笔记突然限流","上周还能蹭的热点今天直接违规"…… 最近刷小红书,满屏都是创作者的哀嚎。 算法一变,有人一夜爆单,有人瞬间归零。这年头在小红书混口饭吃,简直比高考押题还难。 以前,小红书的新笔记哪怕数据一般,也会因为"新鲜度"获得一定曝光。但现在,算法开始考核"贡献 力"——点击、停留、点赞、收藏、评论、转发,每项指标都影响流量分发,而且权重不同,比如评论 就比单纯点击更"值钱"。 单纯靠搬运或低质量内容已经很难生存,必须真正懂用户、提供价值,才能获得推荐。听起来是好事, 但实际操作中,生手也难突围了。 一、政策高压下的算法变形记 但明面上,上线"内容偏好调节",让你自己选想看美妆还是数码。暗地里,把流量偷偷倾斜给能赚钱的 内容——带货笔记权重暴涨30%,纯分享类内容曝光量腰斩。 更狠的是"600字生死线",低于600字的笔记直接扔进流量垃圾场,逼着创作者写小作文。嗯,为了凑字 数,袜子颜色都得写200字心理活动! 二、曝光延迟缩短,商家利好 举个栗子: ...
货拉拉公布第二批算法公开举措:预计2025年将降低抽佣共计2.3亿元
news flash· 2025-05-29 12:16
继今年3月首次公开算法后,数字货运平台货拉拉今日公布第二批算法公开举措。记者现场获悉,此次 算法公开具体内容包括:平台动态评估算法与规则,进一步提升分播单的公平性;优化自动降抽佣算 法,持续推进司机降费减负举措等内容。货拉拉称,近期继续优化了"司机自主,就近优先"分单模式, 最大限度减少其他因素影响,以距离为标准优先中单的订单占比从90%提升到93%。货拉拉同时表示将 持续推出司机降佣产品,预计2025年平台将降低抽佣共2.3亿元。此外,平台将加大订单补贴力度,预 计每年将投入近7000万元,以提升司机收入。(每日经济新闻) ...
陈芋汐遭网暴,抖音回应
21世纪经济报道· 2025-05-28 15:36
5月28日,抖音黑板报微博账号回应有网友举报陈芋汐被网暴:经核查,平台已全面清理针对 相关运动员的攻击、侮辱、极端、不友善言论。平台持续针对体育饭圈现象展开专项治理, 严厉打击涉体育领域的网暴、诽谤、煽动对立和非理性追星行为。平台也呼吁广大用户理性 表达支持,抵制网络戾气。 值得注意的是,5月28日,中国体育报发布了《算法茧房正在编制体育的困局》一文,文章提 到:"在前不久的多哈世乒赛上,国乒小将林诗栋输球被网暴,国乒选手之间的比赛继巴黎奥 运会之后再度出现被喝倒彩的情景。 而陈芋汐在全国跳水冠军赛上虽然拿到了冠军,依然难 逃被网暴的结局。" 上述文章表示,虽然竞技体育中常常会有平局的存在,但终归还是要分出输赢、分出高下、 决出排名。当这种无论输赢都被"暴揍"的两难境地日益频繁地成为舆论环境中的常态,终将 会把体育推向不可输也不可赢的困局,妨碍广大青少年建立对体育精神的正确认知,制约体 据该网友反馈,2025年全国跳水冠军赛,陈芋汐与掌敏洁双人10米台夺冠后,抖音上出现大 量针对陈芋汐的攻击、侮辱言论。该网友表示,"刷了一下抖音,看到一群抖人疯狂网暴陈芋 汐,没人管一下吗?" 据此前报道,5月25日,陈芋汐在 ...
金盛贵金属投资被骗?权威数据揭开合规平台的「安全密码」
Cai Fu Zai Xian· 2025-05-27 07:09
近期网络上关于贵金属平台的争议频发,"金盛贵金属投资被骗" 的不实传闻引发投资者关注。诈骗团 伙克隆正规平台 APP、伪造交易数据、设置提现陷阱等手段,利用投资者对贵金属市场的关注实施精 准诈骗。 经查证,金盛贵金属是具备香港交易所合规资质、过人技术实力与全球化服务网络的合规合法平台,其 母公司金盛金融集团覆盖全球 113 个国家和地区,注册用户超 3000 万,日均现货黄金交易量突破 10 亿 美元。 一、合规基因:百年监管体系的「防火墙」 面对贵金属市场的剧烈波动,金盛贵金属的技术优势尤为突出。其 MT4/MT5 双平台支持止损、止盈自 动触发功能,即使在美联储利率决议等行情爆发时段,订单执行速度仍低于 0.03 秒,滑点率控制在行 业领先的 0.05% 以内。这种技术稳定性,在 2025 年 4 月国际金价突破 2400 美元 / 盎司的行情中,帮助 投资者精准捕捉交易机会。 平台创新性推出的 "智能行情预警系统",结合 AI 算法实时分析全球金融数据,当伦敦金 / 银价格波动 超过预设阈值时,自动通过 APP 推送预警信息。这种智能化工具与 2025 年黄金 ETF 规模突破 2500 吨 的市场趋势 ...
扑空、跌倒、卡住,机器人打架“意外连连”
Nan Fang Du Shi Bao· 2025-05-26 23:13
Core Viewpoint - The first human-shaped robot fighting competition, part of the CMG World Robot Competition series, took place in Hangzhou, Zhejiang, showcasing advanced robotic combat capabilities and technology [7][18]. Group 1: Competition Overview - The competition featured four UTree G1 robots controlled by operators from different fields, with one robot wearing black protective gear emerging as the champion [7]. - The event utilized a point system over three rounds, with effective strikes to the head and torso scoring points, and penalties for being knocked down [8]. Group 2: Technical Challenges - The competition tested robots' instantaneous explosive power, requiring strong power systems for quick action responses, and highlighted the importance of control algorithms for precise movement [9][10]. - Robots faced dynamic environments that challenged their stability and control, with many robots showing signs of wear and tear from the competition [11]. Group 3: Learning and Control - The G1 robot has eight basic fighting movements and various combinations, allowing it to perform punches and kicks [14]. - Operators used voice and remote control for real-time manipulation of the robots, emphasizing human-robot collaboration [15]. Group 4: Industry Implications - The extreme testing environment of the robot fighting competition promotes continuous algorithm optimization and performance enhancement, benefiting human production and daily life [17]. - The competition signifies a new phase in robotic demonstrations, pushing companies to improve algorithms and hardware, and increasing the demand for standardized testing protocols in the industry [18].