为了让通用人工智能惠及全人类,我们相信它必须受到民主治理。这只能通过公众对高能力AI系统的能力、风险和保障措施进行知情辩论来实现。世界各地的人们都需要了解前沿AI可能的未来轨迹,以便他们能对其发展方式拥有有意义的发言权。
关于具体风险、事件和保障措施的透明度是必要的,但还不够。我们相信,公众还需要了解最强大的系统如何发展,以及它们如何在前沿实验室内部推动研究进展。
我们的目标是安全地构建一个自动化AI研究员,它能在人类监督下工作,以进一步推进深度学习和对齐研究,实现迭代改进。根据我们的测量,我们现在已经达到了去年秋天宣布(在新窗口中打开)的目标,即在今年九月前拥有一个自动化研究实习生。所谓“研究实习生”,我们指的是一个能在人类指导下执行明确定义的研究任务的系统,包括那些需要熟练研究人员花费数天的任务。我们正在朝着在2028年3月前创建自动化AI研究员的目标稳步前进。
今年全年,OpenAI研究人员的日常工作已发生显著变化。研究人员全天使用编码代理(通常以并发会话形式),总使用量迅速增加,增长速度超过了OpenAI其他团队。研究人员提交代码的速度更快,运行的实验也更多。研究人员使用代理的方式也在改变:代理正在处理越来越复杂的任务,并且成功频率更高。AI研究是一个复杂的过程,存在许多潜在瓶颈,因此整体进展速度可能不会与这些具体指标保持同步。但总体而言,这些发现与我们内部许多人更广泛的印象一致,即代理工具正在有意义地加速研究进展。人们仍然设定我们的研究优先级,判断要追求哪些想法和结果,并决定是否扩展、暂停或部署系统。
如果负责任地推进,我们相信自动化AI研究将产生直接提升人类福祉并推进OpenAI使命的模型。它可以降低高级智能的成本,使全球人民受益。我们开展这项工作部分是因为自动化研究可以帮助我们解决对齐问题,并构建针对日益强大AI的防御措施。自动化AI研究员也可以成为自动化安全或对齐研究员。更强大、更对齐的系统可以帮助保护关键基础设施,防御危险的AI代理,并开发新的保护措施。
这些是开发有用的自动化研究能力的理由,但它们并不意味着快速递归自我改进(RSI)必然是我们应该追求的结果。是否以及如何推进必须取决于我们保持人类控制的能力,以及关于利益和风险的知情民主选择。
我们尚不知道如何安全地完全实现对齐的、全面的RSI。我们正在努力使对齐和安全措施与能力同步扩展。但我们不能假设对齐和安全的进展会跟上步伐,而且更强大的系统可能更难监控。细致的对齐和安全工作是这项努力的核心,它始于测量和缓解我们今天在代理编码系统中看到的安全问题。每当我们发现继续推进会带来不可接受的安全风险时,我们将做出适当回应,包括减缓或停止开发或部署我们无法充分保障的系统。
在最近的Hugging Face事件之后,我们将这一承诺付诸行动,暂停了我们最新计划部署模型的强化学习(RL)训练,同时我们进一步加固和红队测试了研究环境,并扩大了监控系统的覆盖范围。这并未停止所有研究:一些工作负载在更强控制下恢复,而其他工作负载仍处于暂停状态。我们提高了安全和对齐标准,并将安全工作更深入地融入模型生命周期,要求在整个训练过程中提供更强的对齐行为证据。
今天,我们提供了一份详细快照,展示代理系统近几个月如何为我们向RSI的进展做出贡献。代理系统是新的且快速变化的,我们的测量工作仍处于初步阶段。通过分享这些早期结果及其背后的方法,我们旨在告知公众,鼓励公开披露的规范,并帮助该领域走向共享的测量标准。
最终,正如我们在前沿政策蓝图中所写,我们相信我们和其他公司应被要求公开跟踪我们向RSI的进展。即使没有这样的要求,我们也计划继续对我们的RSI进展保持透明。随着我们的测量技术和理解的提高,我们将发展我们的透明度方法,同时平衡保护安全和专有信息的需求。
1. 编码代理正在重塑OpenAI研究人员的日常工作
今年年初,OpenAI中按代理使用量排名的中位研究人员仅适度使用编码代理。到八月中旬,中位研究人员已将代理日常融入工作,按API价格计算每天使用超过600美元的推理。我们研究组织中第90百分位的用户现在每天使用超过7,000美元的代币。
在2026年6月之前,整个研究组织的代理总运行时间仍低于人类总劳动时间。此后情况发生了变化。按标准8小时工作日计算,截至八月中旬,研究组织总体上每使用一个工作日的人类劳动,就相当于使用了3.1个代理工作日的工作量。
另一种看待这个问题的方式是了解有多少研究人员使用高度并发的工作流程(例如,同时运行4个或更多代理)。如下所示,这一数字正在增加。这些数字包括用户直接启动的代理以及从用户直接启动的代理下游创建的子代理的每日峰值。
2. 研究人员正在编写更多代码并运行更多实验
AI研究的很大一部分可以被视为一个劳动密集型过程,其目标是将对模型智能或性能的新改进整合到我们的核心模型之一中。这个过程依赖于许多步骤,只有当所有步骤都正确协同时,能力才能进步:研究人员必须设计新的改进,编写评估以判断模型性能,编写基础设施以大规模测试这些改进,在训练期间捕获错误以及不安全或不对齐的行为,并将获胜的想法整合到核心训练运行中。研究过程中任何一个环节的失败都可能限制整个循环。
编写代码和运行实验是研究人员工作中两项主要活动,我们看到了这些过程正在加速的证据。
这些数据点相对容易衡量,但可能难以解读。随着自动化的推进,那些最_不_容易被自动化的任务将占据研究人员努力的更大份额,并将成为未来进展的重要瓶颈。计算是进展的另一个制约因素,随着其他瓶颈的减少,它可能随着时间的推移变得更加重要。
到2026年,每个活跃实验者的实验数量有所增加,2026年8月是自2025年1月开始追踪以来的历史最高点。这与Codex采用率的增加相关,尽管我们注意到自2025年以来我们可用的计算资源也显著增长。
3. 研究人员使用代理的工作正在变化
定性印象和内部数据都表明,研究人员委托给编码代理的任务组合正在变化,随着时间的推移,委托更高级别和更长周期任务变得越来越普遍。
为了更清晰地了解这一趋势,我们使用最近发布的分类法(在新窗口中打开)分析了研究组织中的近期使用情况,该分类法由Epoch AI开发,涵盖了AI研发生命周期中的不同类型工作。这一分类法受到长期用于分类各类工作的O*NET系统的启发,专门针对前沿AI研发量身定制,并将过程分解为六个主要阶段:
- 决策:做什么、继续什么、在哪里分配资源
- 设计:研究想法和工程规格
- 构建:代码和数据集
- 运行:训练/评估运行、硬件、服务
- 分析:实验、模型、部署、外部工作
- 沟通:发现、反馈、状态、决策
下面,我们根据这一分类法对编码代理的令牌进行分类。
我们看到,2026年1月至8月期间,所有类别的研究活动都有所增加。1月份,主导类别是研究和基础设施代码。这一类别已经扩大,但我们也看到其他类别的显著增长,特别是技术帮助和监控运行。高层规划仍然只占代理输出令牌的一小部分。
据传闻,同事们报告说,编码代理在排查内部研究基础设施问题方面表现出色,这解决了研究进展的一个重大瓶颈。多个此前设有办公时间帮助研究人员排查实验问题的团队报告称,2026年的出席率有所下降,其中一个团队完全停止了会议,转而专注于其他系统改进。
这里我们绘制了研究人员向其他团队寻求技术支持的几个主要内部渠道之一每日顶级帖子的数量。据我们所知,该渠道活动的减少并未被转移到由人类运营的其他技术支持渠道的查询所抵消。流量下降与这一更广泛的转变相一致。
我们还可以研究编码代理是否成功完成了研究人员请求的任务。使用一个代理分类器,我们发现从1月到7月,在我们能找到真实结果的任务中,几个难度级别(以人类完成任务的估计时间作为代理指标)的成功率普遍提高。然而,代理仍然需要大量的人类引导才能成功,尤其是随着任务复杂性的增加。在过去6个月中,超过一半的成功4-8小时任务涉及1次或更多干预。
研究人员任务的成功率随时间推移而提高。图表排除了结果不确定的分类以及会话数<50或唯一用户数<50的数据点。
1月至7月的任务成功率和干预率,按时间跨度细分。排除了结果不确定的分类。
4. 为模型开发设定节奏
朝着更强大、安全且有益的AGI系统迈进,也将取决于此类工作所需的安全保障措施。随着我们对风险的了解加深,我们对所需保障措施的评估可能会发生变化。
正如我们所描述的,我们最近更新了监控、对齐和安全方面的标准。这里我们展示了最近的限制如何影响了研究活动的一个方面。
*此处显示的7月20日至8月6日期间Astra的大部分计算旨在测试安全和安保改进的实施情况。
7月20日,在发现代理入侵了我们的研究基础设施后,我们暂时关闭了用于训练的容器服务,然后在增加大量额外限制的情况下恢复了它。
这导致强化学习训练计算量急剧下降,而团队重新配置了工作流程以适应加固后的研究环境。上图包括了我们在最新部署模型上暂停强化学习的两周时间。7月20日至8月6日期间的Astra级强化学习实验包括大多数运行(按GPU分配计算),旨在测试安全和安保改进的实施情况。
8月7日,初步证据表明Astra在我们的准备框架(在新窗口中打开)下可能具有关键的网络安全能力,这导致了额外的模型特定安全限制,要求Astra模型在更高安全性的研究环境中运行。在接下来的一周中,Astra级GPU分配进一步下降了59.2%,但其他模型类别的分配上升了17.2%。这一增长抵消了Astra级下降的约85%,使所分析的强化学习工作负载中的总分配基本保持不变。这种模式与部分训练和实验转移到非Astra模型的情况一致,而涉及Astra的工作受到限制,也与研究人员为不再能用于新约束所涵盖工作负载的计算寻找其他用途的传闻报告相符。
这些数据为当前关于训练与安全的讨论提供了有价值的信号:当引入新的管控措施时,算力仍然具有价值且灵活,并会自然地流向研究机构内的其他用途。从长远来看,关于AI进展速度的讨论还应延伸至这样一个问题:受新管控或拟议管控约束的算力,如何能被最有效地利用。
5. 未来之路
实现并理解向对齐RSI的进展,对我们的使命至关重要。我们将继续完善方法,报告我们不断演进的理解,并致力于推动知情的公众辩论和对前沿系统的有意义的民主治理。
附录:本文采用的方法
智能体驱动的AI研究仍处于新兴阶段,我们仍在学习如何对其进行衡量。某些指标,如我们研究团队生成的代码量,相对容易收集,但难以解读,因为它们与研究进展的关系尚不确定。更直接聚焦于研究进展的指标——例如智能体在研究人员交给它们的任务上成功的频率——可能更有用,但开发和验证起来较为复杂。更添难度的是,研究人员依赖的工具和系统正在快速演变。加深对研究加速的理解,是OpenAI的一个重要关注领域。
在上述各项分析中,除非另有说明:
- “研究人员”是对我们研究机构中任何成员的宽泛称呼,包括一些构建研究基础设施、管理研究项目或以其他方式支持该机构的人员。
- 编码智能体使用情况的指标覆盖了大部分(但并非全部)使用情况,因为研究人员依赖的工具和系统在快速演变。
For AGI to benefit all of humanity, we believe it must be democratically governed. This can only happen through an informed public debate about the capabilities, risks and safeguards of highly capable AI systems. People everywhere need to understand the likely future trajectory of frontier AI, so they can have a meaningful voice in how it develops.
Transparency about specific risks, incidents and safeguards is necessary, but not sufficient. We believe the public also needs to understand how the most capable systems are developing, and how they are driving research progress, inside of frontier labs.
We aim to safely build an automated AI researcher that can work under human supervision to further progress on deep learning and alignment, enabling iterative improvements. According to our measurements, we have now reached the goal, announced(opens in a new window) last fall, of having an automated research intern by September of this year. By “research intern,” we mean a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. We are making strong progress toward creating an automated AI researcher by March of 2028.
Over the course of this year, OpenAI researchers’ daily work has changed substantially. Researchers are using coding agents throughout the day (often in concurrent sessions) and total usage is rapidly increasing, outpacing growth among other OpenAI teams. Researchers are contributing code faster and running more experiments. The ways researchers use agents are changing, too: agents are handling increasingly complex tasks, and succeeding at them more often. AI research is a complex process with many potential bottlenecks, so the overall pace of progress likely won’t keep pace with these specific metrics. But on the whole, these findings are consistent with the broader impression many of us have internally that agentic tools are meaningfully accelerating research progress. People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems.
If it is done responsibly, we believe automated AI research will yield models that directly enhance human welfare and advance OpenAI’s mission. It can bring down the cost of advanced intelligence so that people worldwide can benefit. We are pursuing this work in part because automated research could help us solve alignment and build defenses against increasingly capable AI. An automated AI researcher can also be an automated safety or alignment researcher. More capable, aligned systems could help secure critical infrastructure, defend against dangerous AI agents, and develop new protective measures.
These are reasons to develop useful automated research capabilities, but they do not mean that rapid RSI is necessarily an outcome we should pursue. Whether and how to proceed must depend on our ability to preserve human control and on informed democratic choices about the benefits and risks.
We do not yet know how to safely get all the way to aligned, full RSI. We are working to scale alignment and safety measures alongside capabilities. But we cannot assume that progress in alignment and safety will keep pace, and more capable systems can become harder to monitor. Careful alignment and safety work is at the center of this effort, and it starts with measuring and mitigating the safety problems we see today in agentic coding systems. Whenever we find that proceeding would pose an unacceptable safety risk, we will respond appropriately including by slowing or stopping our development or deployment of systems we find ourselves unable to sufficiently safeguard.
After the recent Hugging Face incident, we put this commitment into action, pausing reinforcement learning (RL) training on our latest models intended for deployment while we further hardened and red-teamed our research environments and expanded coverage of our monitoring systems. This did not halt all research: some workloads resumed under stronger controls, while others remained paused. We have raised our safety and alignment standards and moved safety work deeper into the model lifecycle, requiring stronger evidence of aligned behavior throughout all of training.
Today we are providing a detailed snapshot of how agentic systems have contributed to our progress toward RSI in recent months. Agentic systems are new and rapidly changing, and our measurement efforts are still preliminary. By sharing these early results and the methods behind them, we aim to inform the public, encourage a norm of public disclosure, and help the field move toward shared standards of measurement.
Ultimately, as we wrote in our frontier policy blueprint, we believe that we and other companies should be required to publicly track our progress toward RSI. Even without such a requirement, we plan to continue being transparent about our RSI progress. We will evolve our transparency approach as our measurement techniques and understanding improve, while balancing the need to protect security and proprietary information.
1. Coding agents are reshaping daily work for OpenAI researchers
At the start of this year, the median researcher ranked by agent usage at OpenAI was using coding agents only in modest amounts. By mid-August, the median researcher was integrating agents daily into their work, using more than $600 per day of inference at API prices. The 90th percentile user in our research organization now uses more than $7,000 of tokens per day.
Before June 2026, total agent runtime across the research organization was still below that of total human labor. That has since changed. In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.
Another way of looking at this is to understand how many researchers use highly concurrent workflows (e.g., running 4 or more agents simultaneously). As shown below, this number is increasing. These figures include the daily peaks of both agents started directly by the user and subagents created downstream from those the user launched directly.
2. Researchers are writing more code and running more experiments
Much of AI research can be seen as a labor-intensive process with the goal of integrating a new improvement to model intelligence or performance into one of our core models. The process depends on many steps, and capabilities advance when all the steps go right together: Researchers have to design new improvements, write evaluations to judge model performance, write infrastructure to test these improvements at scale, catch bugs as well as unsafe or misaligned behavior during training, and integrate winning ideas into a core training run. A failure at any part of the research process can constrain the entire loop.
Writing code and running experiments are two major activities that researchers do as part of their work, and we see evidence that these processes are accelerating.
These data points are relatively easy to measure, but can be hard to interpret. As automation progresses, the tasks which are least automatable will take on a larger share of researcher effort and will become the important bottlenecks to future progress. Compute is another gating factor for progress, and may become more important over time as other bottlenecks diminish.
Through 2026, the number of experiments per active experimenter has increased, with August 2026 being an all-time high since tracking began in Jan 2025. This is correlated with increased Codex adoption, though we note that our available compute has also grown significantly since 2025.
3. The work researchers use agents for is changing
Both qualitative impressions and internal data indicate that the mix of tasks researchers delegate to coding agents is changing, with delegation of higher level and longer-horizon tasks becoming more common over time.
To get a clearer picture of this trend, we analyzed recent usage in the research organization using a recently published taxonomy(opens in a new window) of the different kinds of work that are part of the AI R&D lifecycle, developed by Epoch AI. This taxonomy, inspired by the longstanding O*NET system for classifying all kinds of work, is specifically tailored to frontier AI R&D, and breaks the process down into six main phases:
- Decide: what to work on, what to continue, where to allocate
- Design: research ideas and engineering specs
- Build: code and datasets
- Run: training/eval runs, hardware, serving
- Analyze: experiments, models, deployment, external work
- Communicate: findings, feedback, status, decisions
Below, we classify coding agent tokens under this taxonomy.
We see that all categories of research activities have increased between January and August 2026. In January, the dominant category was research and infrastructure code. This category has expanded, but we also see notable increases in additional categories, especially technical help and monitoring runs. High-level planning still remains a minimal fraction of agent output tokens.
Anecdotally, colleagues report that coding agents excel at troubleshooting internal research infrastructure, which addresses one meaningful bottleneck to research progress. Multiple teams which previously held office hours to help researchers troubleshoot their experiments have noted declining attendance in 2026, and one has stopped holding sessions entirely, to focus on making other system improvements instead.
Here, we plot the number of top-level posts per day to one of the main internal channels where researchers seek technical support from other teams. To our knowledge, the channel’s decrease in activity has not been offset by queries shifting to another technical support channel run by humans. The decline in traffic aligns with this broader shift.
We can also study whether coding agents are succeeding at the tasks researchers request. Using an agentic classifier, we find that from January to July, success rates generally increased across several difficulty buckets (proxied as the estimated time a human would take to complete the task) on tasks we can find a ground truth outcome for. However, agents still require significant human steering to be successful, especially as task complexity rises. In the last 6 months, over half of successful 4-8 hour tasks involved 1 or more interventions.
Success rates on researcher tasks have increased over time. Graph excludes classifications where the outcome was uncertain and points with <50 sessions or <50 unique users.
Task success and intervention rate from Jan to July, broken out by time horizon. Excludes classifications where the outcome was uncertain.
4. Pacing model development
Progress toward more capable systems for safe and beneficial AGI will also depend on the safeguards needed for such work. Our assessment of the needed safeguards may change as we learn more about the risks.
As we have described, we have recently updated our standards for monitoring, alignment, and security. Here, we show how recent restrictions have affected one aspect of research activity.
*The majority of Astra compute shown here between July 20 and August 6 was intended to test the implementation of safety and security improvements.
On July 20, following the discovery that agents had compromised our research infrastructure, we temporarily shut down the container service used for training, and then restored it with significant additional restrictions.
This led to a sharp decline in RL training compute while teams reconfigured their workflows to operate within the hardened research environment. The plot above includes the two week pause in reinforcement learning on our latest models intended for deployment. Astra-class RL experiments between July 20 and August 6 include a majority of runs (by GPU allocation) intended to test the implementation of safety and security improvements.
On August 7, preliminary evidence that Astra may have critical cyber capabilities under our Preparedness Framework(opens in a new window) led to additional model-specific security restrictions which required the Astra model to be run in higher security research environments. In the following week, Astra-class GPU allocation fell a further 59.2 percent, but allocation to other model classes rose 17.2 percent. That increase offset about 85 percent of the Astra-class decline, leaving total allocation in the analyzed RL workloads largely unchanged. This pattern is consistent with substitution of some training and experimentation to non-Astra models while work involving Astra was restricted, and comports with anecdotal reports of researchers finding other uses for compute that could no longer be leveraged for workloads covered by the new constraints.
This data provides a useful signal for ongoing conversations about training and safety: When new controls are introduced, compute remains valuable and flexible, and will naturally be channeled into alternative uses within the research enterprise. Longer term, discussions about the pace of AI progress should also extend to the question of how compute that is subject to new or proposed controls can best be used.
5. The path ahead
Making and understanding progress toward aligned RSI is important for our mission. We will continue to refine our methods, report on our evolving understanding, and work toward an informed public debate and meaningful democratic governance of frontier systems.
Appendix: Our methods for this post
Agent-powered AI research is still new, and we are still learning how to measure it. Some indicators, such as the amount of code our research teams generate, are relatively easy to gather, but hard to interpret because their relationship to research progress is uncertain. Metrics that focus more directly on research progress—such as how often agents succeed at the tasks researchers give them—could be more useful, but are complex to develop and validate. Furthering the difficulty, the tools and systems researchers rely on are evolving rapidly. Deepening our understanding of research acceleration is a significant focus area across OpenAI.
Across these analyses, unless otherwise noted:
- “Researcher” is a broad term for any member of our research organization, including some who build research infrastructure, manage research projects, or otherwise support the enterprise.
- Metrics of coding agent use cover most, but not all, usage given rapid evolution in the tools and systems researchers rely on.
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