作为一家量化交易公司，Jump Trading创建预测模型，利用市场数据、新闻和事件以及一系列另类数据源，对资产价格做出尽可能准确的预测。由于市场复杂、充满噪声且随时间不断变化，几乎不可能准确预判将会发生什么。但据Jump LLM研发负责人Lucas Baker所说，只要在规模上比抛硬币略胜一筹，就足以形成成功的策略。
Baker领导智能体研发，专注于构建智能体、运行框架和基础设施，让量化研究人员能够更广、更深地探索自己的想法。加入GPT‑6 Astra后，可以交给智能体处理的工作流在规模和复杂度上都大幅扩展，从日常编码到用于验证新假设的高级量化研究，皆在其列。
“借助GPT-6系列，尤其是GPT-6 Astra，OpenAI为长周期任务解锁了新的自主性层级，这类任务需要灵活的智能体协调，以及在复杂工作流上的极度持久性。过去我们经常需要人工指导和干预，现在我们完全可以专注于定义一个安全且受到良好监控、目标明确的环境，然后让智能体自行探索路径。”
——Lucas Baker，Jump Trading LLM研发负责人
从短小代码片段到全面分析
在过去一年里，AI已从一种有用的工具——适合编写一次性代码片段或查找小错误——转变为一种能力强大、用途广泛的系统，可以自行开发完整的代码库和服务。如今，Baker和他的团队发现，当AI更像一位同事那样被对待时，效果最好。研究人员可以定义关键问题、工作环境，以及评估结果质量和重要性的方法，然后实时引导一个或多个智能体，告诉它们应将分析重点放在哪里，或下一步运行哪个任务。
Baker表示，借助GPT‑6 Astra，智能体现在不仅能找到有意义且实用的改动，还能在递归改进的过程中将这些成果合并并叠加起来。在一个长时间运行的任务中，系统可以分析自己的发现，依据最初提案中商定的标准进行判断，并主动调整方向，而不需要人来分析每一轮改动。正如他所说：“你可以定义一件需要运行数天的事情——它需要从许多数据源拉取数据，需要对什么是重要的、什么是不重要的做出那些微妙的判断，还需要把一切相互关联起来——从而形成一份真正可行的全面分析，”他说。
让人类判断始终处于核心位置
Jump Trading的业务覆盖所有时间周期和资产类别。流程中的每一步都很复杂，而在金融这样受到严格监管的行业中，错误可能同时带来财务和合规后果。Baker表示，必须意识到将工作交给AI所带来的风险，但同样重要的是，智能体智能也可以用于提升质量、安全性和监控，而不仅仅是增加功能。强大的系统设计和边界、清晰的约束、有助于可引导性和可观测性的基础设施，以及对改动的人工审查，都能帮助Jump Trading团队确保AI赋能的工作流已准备好在受监管环境中规模化运行。
“如果你有一个安全的环境，让智能体或系统可以自由产出它需要的任何输出，但最后又有一个由人类验收、进行关键验证的人工审查流程，这就是让我们有信心的原因，”他说。例如，如果某个智能体生成一个交易信号，它就会像任何输出一样被界定范围并接受审查：作为一个信号，通常具有信息价值，但也可能是错的，并且会在严格审查和受控的执行环境中与所有其他信号整合。
下一步是什么
Baker说，他认为我们正走向一个“自动研究”无处不在的世界，以至于它会被视为量化研究人员日常工作流的一部分；所谓自动研究，就是由智能体研究人员对可度量系统进行递归改进。如今，即便是GPT‑6 Astra运行时间最长的工作，仍然需要与定义任务的人定期沟通——不仅包括拉取哪些数据、运行多久、什么算作重要，还包括中间结果是否合理。高级自动研究仍将从人类定义的系统开始，包括输入、环境定义、评估指标、权衡取舍和总体优先级，但会把其余流程交给一个结构较为松散的智能体“集群”，而这些智能体本身又由其他智能体协调。在一个结构良好的研究流水线中，这些智能体将能够就如何探索想法、分配计算时间和整合有前景的发现做出智能决策，而这一切都始于一个开放性问题。
“当你能够把这一切端到端地串起来，提出一个连你自己都不知道答案的通用问题，并让一个有用的结果返回时，那会是什么样子？”他问道。“如果你能解决一个千禧年大奖难题，你大概也能弄清楚一些关于量化金融的相当有趣的事实。”
每一年带来的不仅是基准测量上的稳步提升，还有模型所能执行的工作类型上的阶跃式变化。这些发展往往很难提前几个月预测：2024年，早期智能体能够不出错地写出单个文件就已令人印象深刻；到了2025年，从零创建完整代码库已成为可能；而到了2026年，许多智能体已经能够动态协作、并肩推进开放研究问题。“明年我们会拥有什么？”他说。“这是一个相当令人难以置信的前景。”
As a quantitative trading firm, Jump Trading creates predictive models that use market data, news and events, and a range of alternative data sources to make the best possible predictions about asset prices. Because markets are complex, noisy, and changing over time, it is rarely possible to anticipate exactly what will happen. But according to Lucas Baker, Head of LLM R&D at Jump, predicting even slightly better than a coin flip at scale is enough to result in a successful strategy.
Baker leads agentic research and development, and he’s focused on building the agents, harnesses, and infrastructure that let quantitative researchers explore their ideas in greater breadth and depth. Adding GPT‑6 Astra has dramatically expanded the scale and complexity of workflows that can be handed off to agents, from day-to-day coding to advanced quantitative studies to validate new hypotheses.
“With the GPT-6 series, especially GPT-6 Astra, OpenAI has unlocked a new tier of autonomy for long-horizon tasks that require flexible agent coordination and extreme persistence on complex workflows. Where we used to require frequent human guidance and intervention, we can now focus fully on defining a secure and well-monitored environment with clear goals and letting the agents find their own way.”
—Lucas Baker, Head of LLM R&D, Jump Trading
From short snippets of code to comprehensive analysis
Over the past year, AI has transformed from a helpful tool, useful for writing one-off code snippets or finding small bugs into a capable, versatile system that can develop entire codebases and services by itself. Now, Baker and his team find that AI works best when treated more like a colleague. Researchers can define a key problem, a work environment, and a way of evaluating the quality and significance of results, then steer one or many agents in real time about where to focus the analysis or which job to run next.
Baker says that with GPT‑6 Astra, agents are now capable of not only finding meaningful and practical changes, but merging and stacking those wins together in a process of recursive improvement. Over the course of a single long-running task, the system can analyze its findings, judge them against the agreed-upon criteria from an initial proposal, and actively redirect its efforts rather than needing a person to analyze each round of changes. As he puts it, “You can define something that needs to run for days—it needs to pull from many data sources, it needs to make those subtle calls about what is important and what is not, and it needs to interrelate everything—to create a comprehensive analysis that actually works now,” he says.
Keeping human judgment at the center
Jump Trading works across every time horizon and asset class. Every step of the process is complex, and in a heavily regulated industry such as finance, mistakes can have both financial and compliance consequences. Baker says that it is critical to be aware of the risks that arise from entrusting work to AI, but also that agentic intelligence can also be applied to improving quality, security, and monitoring, not just adding features. Strong system design and boundaries, clear constraints, infrastructure that promotes steerability and observability, and human review of changes help Jump Trading’s team ensure that AI-enabled workflows are ready to scale in a regulated environment.
“If you have a safe environment where the agent or system is free to produce any output that it needs, but there is also a human review process at the end of it where critical validation takes place with human acceptance, that’s what gives us confidence,” he says. For example, if an agent produces a trading signal, it’s scoped and reviewed the same way any output would be: as a signal, usually informative but potentially wrong, and integrated with every other signal in a stringently reviewed and controlled execution environment.
What’s next
Baker says he thinks we’re headed towards a world where “autoresearch,” or the recursive improvement of measurable systems by agent researchers, will become so ubiquitous it is considered simply part of a quantitative researcher’s ordinary workflow. Today, even GPT‑6 Astra’s longest-running work still involves regular check-ins with the person defining the task—not only what data to pull, how long to run, and what counts as important, but whether the intermediate results make sense. Advanced autoresearch would still begin with a human-defined system, including inputs, environment definition, evaluation metrics, trade-offs, and overall priorities, but would entrust the rest of the process to a loosely structured “fleet” of agents themselves coordinated by other agents. Within a well-structured research pipeline, these agents would be able to make intelligent decisions about how to explore ideas, allocate compute time, and integrate promising findings, all starting with little more than an open question.
“What does it look like when you can put all of this end to end, ask a general question that even you don’t know the answer to, and have a useful result come back?” he asks. “If you can solve a Millennium problem, you can probably also figure out some pretty interesting facts about quant finance.”
Every year has brought not only steady advancements in benchmark measurements, but also step changes in the type of work models were capable of performing. Often these developments were difficult to predict even months in advance: while it was considered impressive in 2024 for early agents to write a single file without mistakes, it became possible in 2025 to create entire codebases from scratch, and in 2026 to make progress on open research questions with many agents collaborating dynamically side by side. “What are we going to have next year?” he says. “It’s a pretty incredible prospect.”
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文