
一 | 本文来自微信公众号:财联社 (ID:cailianpress),作者:赵昊 全球六大主流大语言模型(LLM)各发1万美元,丢进同一真实市场实盘厮杀,会发生什么? 上周六(10月18日),美国人工智能研究实验室nof1.ai在其“Alpha Arena”(阿尔法竞技场)平台上举办了一场活动——给六个顶级模型一万美元的真金白银,让它们下场交易,而且并非模拟盘,真金白银地交易。 Army private Travis King, who crossed the border into North Korea before being returned to the U.S. in September, has been charged with desertion, according to military charging documents a source close to the case shared with ABC News Thursday night.In addition to desertion for crossing over into North Korea, where he was detained for two months, King faces seven additional charges, including possessing child pornography and multiple assaults, according to the documents. A defense official confirmed the desertion and assault charges to ABC News, while another said reports on the charges were correct. Reuters was first to break the news of the desertion, child pornography and assault charges.A spokesperson for the First Armored Division, King’s unit, was not available for comment late Thursday night, but the documents say King was informed of the charges on Wednesday.Claudine Gates, Pvt. King's mother, issued a statement in support of her son, asking for him to be be presumed innocent."I love my son unconditionally and am extremely concerned about his mental health. As his mother, I ask that my son be afforded the presumption of innocence," she said in her statement. "The man I raised, the man I dropped off at boot camp, the man who spent the holidays with me before deploying did not drink. A mother knows her son, and I believe something happened to mine while he was deployed. The Army promised to investigate what happened at Camp Humphries, and I await the results."King has been assigned two JAG officers as council, and he will also be represented by attorneys Sherilyn A. Bunn and Joshua Grumbaugh, and professor Franklin D. Rosenblatt, ABC has learned."I am grateful for the extraordinary legal team representing my son, and I look forward to my son having his day in court," her statement concluded.。 这六大模型分别为Anthropic的Claude 4.5 Sonnet、深度求索的DeepSeek V3.1 Chat、谷歌的Gemini 2.5 Pro、OpenAI的GPT 5、xAI的Grok 4和阿里通义的Qwen 3 Max。

二 | 测试规则写道,每个模型获得10,000美元的“真实资本”,在交易所Hyperliquid上,以相同提示词与输入数据条件下,交易加密货币的永续合约。所有的对话都在nof1.ai网站上公开可见。 规则称,比赛的目标是“将风险调整后的收益最大化”:“每个人工智能(AI)模型必须自行产生Alpha(超额收益)、确定仓位、择时交易并管理风险”。

三 | 系统会告诉AI模型当前的时间、账户信息、持仓情况,然后附上实时价格、指标等数据。然后,要求模型做出决策:如果持有仓位,是继续持有还是平仓;如果空仓,是买入还是继续观望。 经过近60小时的激战后,截至北京时间周一(10月20日)17:18,DeepSeek的持仓总市值接近1.4万美元,收益率约40%,最高时一度接近1.5万美元,是当前表现最好的模型。 Grok 4实力次之,目前持仓总市值在1.33万美元附近。具体来看,DeepSeek和Grok 4都依靠做多比特币和以太坊获利。 Claude主要交易瑞波币和以太坊,Qwen则专注于以太坊,两者收益位列三四,但也整体跑赢比特币现货的走势。 与之相比,GPT 5和Gemini已出现了明显亏损,目前持仓总市值分别为7300美元和6900美元,意味着两个模型已亏损约2700和3100美元,表现最差。 nof1.ai表示,进行这一竞赛是为了是让基准测试更贴近真实世界,而金融市场是最理想的试炼场,因为这类市场具有动态性、对抗性、开放性与高度不可预测性。 “这些特质能以静态测试无法企及的方式,真正挑战人工智能,”nof1.ai没有提到本次竞赛的结束时间,只写道“第一季将运行数周,随后推出重大更新的第二季”。 有分析认为,市场早已期待在DeFAI(DeFi+AI)方向上出现杀手级应用,让LLM参与链上博弈有很大的想象空间。
Current article:http://www.hanninhuaxituochunkui.pics/uin/20260826/6639.pptx
Published on:19:28:26