AI progress and recommendations

当大众惊讶地发现 Turing test 好像被“越过”之后，很多人觉得奇怪：日常生活居然几乎照常。这个里程碑被谈论了几十年，起初看似遥不可及，后来又突然近在眼前，再后来我们就站在了另一边。我们确实得到了很多出色的新产品，但世界并没有因此发生翻天覆地的变化——尽管计算机现在可以对话、并能处理复杂问题。

大多数人仍把 AI 等同于聊天机器人和更好的搜索，但今天已经出现能在最艰难的智力竞赛中胜过最聪明人类的系统。尽管这些系统仍然表现不均、存在严重弱点，但在能解决如此艰难问题这一点上，它们离成为真正的 AI 研究者看起来更像是“近八成”，而不是“不到两成”。公众如何使用 AI 与 AI 当前能做到的事之间存在巨大落差。

能独立或通过提高人类效率来发现新知识的 AI 系统，很可能对世界产生重大影响。

在短短几年内， AI 在软件工程等领域的能力从只能做那些人几秒钟能完成的任务，进步到能替人做需一小时以上的工作。我们预计不久就会出现能完成需数日或数周人力的系统；但对于能替人完成需数百年工作的系统，我们目前还无从设想。

与此同时，达到某一智力水平的单位成本急剧下降；过去几年每年大约下降 40 倍是一个合理估计。

我们预计到 2026 年， AI 将能做出极小但真实的发现。到 2028 年及以后，我们相当有信心会出现能够做出更重大发现的系统（当然我们也可能判断失误，但这是基于当前研究进展的估计）。

我们长期认为 AI 的进展往往出人意料，社会会与技术共同演化。尽管未来几年 AI 能力预计会迅速且显著提升，但日常生活仍会在很多方面保持出乎意料的稳定——即便工具大幅改进，生活方式也有很强的惯性。

尤其是，我们预计未来会提供新的、更可能更好的实现充实生活的方式，且比今天有更多人能体验到这样的生活。诚然，工作会发生变化，经济转型在某些方面可能非常艰难，甚至可能需要重新界定基本的社会经济契约。但在一个广泛分配的丰裕世界里，人们的生活有可能比今天大为改善。

AI 系统将帮助人们理解健康状况，加速材料科学、药物开发和气候建模等领域的进展，并扩大全球个性化教育的可及性。展示这些切实利益，有助于构建一种共识： AI 不只是提高效率，更能让生活更美好。

OpenAI 对安全深度承诺，我们把安全视为在减轻负面影响的同时促成 AI 积极作用的实践。尽管潜在好处巨大，我们认为超级智能系统的风险可能具有灾难性，因此相信通过实证地研究安全和对齐问题，可以为全球性决策提供依据，例如在接近具备递归自我改进能力的系统时是否应放缓整个领域的发展以便更谨慎地研究这些系统。显然，在不能可靠地对齐和控制超级智能系统之前，任何人都不应部署它们——这需要更多技术性工作。

为实现与 AI 共存的积极未来，我们认为有几件事可行：

• 来自 前沿实验室 的共享标准与洞见：我们认为 前沿实验室 应就安全原则达成一致、共享安全研究与对新风险的认知、研究减少竞赛性行为的机制等。比如前沿实验室就 AI 控制评估达成某些标准可能大有裨益。社会过去通过制定建筑规范和消防标准来保护公众安全，这挽救了无数生命。
• 与能力相称的公共监督与问责框架，既促进 AI 的积极影响，也缓解负面影响：关于 AI 有两种观点。一种认为 AI 属于“普通技术”，会像印刷术、互联网等历史性技术革命一样推进，社会有时间适应，常规公共政策工具足以应对，我们应优先鼓励创新、保护与 AI 的对话隐私，并通过与联邦政府合作防范强大系统被滥用。我们认为当前能力水平的 AI 大体属于这一类，应该广泛扩散，这意味着大多数开发者、开源模型以及现有技术的绝大多数部署不应承受超出既有监管的重大新增负担，尤其不应面对各州之间的零敲碎打式监管。

另一种观点则设想超级智能以人类前所未见的方式和速度出现并扩散。在这种情形下，尽管上述很多做法仍然必要，我们还需更多创新手段；既有监管可能无力应对。因此我们可能需要与行政机构（ executive branch ）以及多个国家的相关机构（例如各类安全研究所）紧密合作，尤其在减少 AI 被用于生物恐怖主义的风险（以及用 AI 来检测和阻止生物恐怖主义）和处理自我改进型 AI 的影响等领域要协调一致。总体方向应是让公共机构承担问责，但实现路径可能需有别于以往。

• 构建 AI 韧性生态系统：无论哪种情境，构建类似网络安全那样的生态系统至关重要。互联网出现时，我们没有靠一条政策或一家企业来保护它，而是发展出一整套网络安全产业：软件、加密协议、标准、监测系统、应急响应团队等。该生态并未消除风险，但将风险降低到社会可以接受的水平，使人们信任数字基础设施并以其构建生活与经济。 AI 也需要类似的配套体系，国家在推动相关产业政策方面可以发挥重要作用。
• 来自 前沿实验室 和政府的持续报告与测量：了解 AI 如何在现实中影响世界，有助于把这项技术引导向积极方向。预测很难：例如 AI 对就业的影响难以预判，部分原因在于今天的 AI 在长处与短处上与人类有很大不同。对实际发生的情况进行衡量将非常有价值。
• 以个体赋能为出发点：我们认为成年人应能在社会划定的宽泛界限内按自己的意愿使用 AI 。我们预计，未来几年获取高级 AI 将成为像电力、洁净水或食物一样的基础性公共服务。最终，社会应支持让这些工具广泛可及，而北极星目标应是帮助人们实现自己的目标并获得更大自主权。

When the popular conception of the Turing test went whooshing by, many of us thought it was a little strange how much daily life just kept going. This was a milestone people had talked about for decades. It felt impossibly out of reach, then all of a sudden it felt close, then all of a sudden we were on the other side. We got some great new products and not much about the world changed, even though computers can now converse and think about hard problems.

Most of the world still thinks about AI as chatbots and better search, but today, we have systems that can outperform the smartest humans at some of our most challenging intellectual competitions. Although AI systems are still spikey and face serious weaknesses, systems that can solve such hard problems seem more like 80% of the way to an AI researcher than 20% of the way. The gap between how most people are using AI and what AI is presently capable of is immense.

AI systems that can discover new knowledge—either autonomously, or by making people more effective—are likely to have a significant impact on the world.

In just a few years, AI has gone from only being able to do tasks (in the realm of software engineering specifically) that a person can do in a few seconds to tasks that take a person more than an hour. We expect to have systems that can do tasks that take a person days or weeks soon; we do not know how to think about systems that can do tasks that would take a person centuries.

At the same time, the cost per unit of a given level of intelligence has fallen steeply; 40x per year is a reasonable estimate over the last few years! 

In 2026, we expect AI to be capable of making very small discoveries. In 2028 and beyond, we are pretty confident we will have systems that can make more significant discoveries (though we could of course be wrong, this is what our research progress appears to indicate).

We’ve long felt that AI progress plays out in surprising ways, and that society finds ways to co-evolve with the technology. Although we expect rapid and significant progress in AI capabilities in the next few years, we expect that day-to-day life will still feel surprisingly constant; the way we live has a lot of inertia even with much better tools.

In particular, we expect the future to provide new and hopefully better ways to live a fulfilling life, and for more people to experience such a life than do today. It is true that work will be different, the economic transition may be very difficult in some ways, and it is even possible that the fundamental socioeconomic contract will have to change. But in a world of widely-distributed abundance, people’s lives can be much better than they are today.

AI systems will help people understand their health, accelerate progress in fields like materials science, drug development, and climate modeling, and expand access to personalized education for students around the world. Demonstrating these kinds of tangible benefits helps build a shared vision of a world where AI can make life better, not just more efficient.

OpenAI is deeply committed to safety⁠, which we think of as the practice of enabling AI’s positive impacts by mitigating the negative ones. Although the potential upsides are enormous, we treat the risks of superintelligent systems as potentially catastrophic and believe that empirically⁠ studying⁠ safety⁠ and alignment⁠ can help global decisions, like whether the whole field should slow development to more carefully study these systems as we get closer to systems capable of recursive self-improvement. Obviously, no one should deploy superintelligent systems without being able to robustly align and control them, and this requires more technical work. 

Here are several things we think could help with achieving a positive future with AI:

Shared standards and insights from the frontier labs. 

We think that frontier labs should agree on shared safety principles and to share safety research⁠, learnings about new risks, mechanisms to reduce race dynamics, and more. We can imagine ideas like frontier labs agreeing to certain standards around AI control evaluations being quite helpful.

Society went through a similar process to establish building codes and fire standards, which have saved countless lives.

An approach to public oversight and accountability commensurate with capabilities, and that promotes positive impacts from AI and mitigates the negative ones. 

There are two schools of thought about AI. One is that AI is like “normal technology,” in that it will progress like other technological revolutions in the past, from the printing press to the internet. Things will play out in ways that give people and society a chance to adapt, and conventional tools of public policy should work. We will need to prioritize ideas like promoting innovation, protecting the privacy of conversations with AI and defending against misuse of powerful systems by bad actors by partnering with the federal government.

We believe AI at around today’s capability levels is roughly here, and should diffuse everywhere, which means most developers and open-source models, and almost all deployments of today’s technology, should have minimal additional regulatory burdens relative to what already exists. It certainly should not have to face a 50-state patchwork

The other one is where superintelligence develops and diffuses in ways and at a speed humanity has not seen before. Here, we should do most of the things above, but we also will need to be more innovative. If the premise is that something like this will be difficult for society to adapt to in the “normal way,” we should also not expect typical regulation to be able to do much either. In this case, we will probably need to work closely with the executive branch and related agencies of multiple countries (such as the various safety institutes) to coordinate well, particularly around areas such as mitigating AI applications to bioterrorism (and using AI to detect and prevent bioterrorism) and the implications of self-improving AI.

The high-order bit should be accountability to public institutions, but how we get there might have to differ from the past.

Building an AI resilience ecosystem. 

In either scenario, building out an AI resilience ecosystem will be essential. When the internet emerged, we didn’t protect it with a single policy or company—we built an entire field of cybersecurity: software, encryption protocols, standards, monitoring systems, emergency response teams etc. That ecosystem didn’t eliminate risk, but it reduced it to a level society could live with, enabling people to trust digital infrastructure enough to build their lives and economies on it. We will need something analogous for AI, and there is a powerful role for national governments to play in promoting industrial policy to encourage this.

Ongoing reporting and measurement from the frontier labs and governments on the impacts of AI. 

Understanding how AI is concretely impacting the world makes it easier to steer this technology towards positive impact. Prediction is hard: for example, the impact of AI on jobs has been hard to anticipate, in part because today’s AIs strengths and weaknesses are very different from those of humans. Measuring what’s happening in practice is likely to be very informative.

Building for individual empowerment. 

We believe that adults should be able to use AI on their own terms, within broad bounds defined by society. We expect access to advanced AI to be a foundational utility in the coming years—on par with electricity, clean water, or food. Ultimately, we think society should support making these tools widely available and that the north star should be helping empower people to achieve their goals.

Generated by RSStT. The copyright belongs to the original author.

Source