AI行业正面临一场“安全-速度悖论”的结构性困境。一方面，为了在与谷歌、Anthropic的竞争中保持领先，AI公司如OpenAI需要以极快的速度推进研发，这导致内部管理混乱和产品开发上的“疯狂冲刺”，例如Codex项目仅用七周就完成。另一方面，研究人员呼吁重视AI安全，要求发布系统卡和安全评估，但内部大量安全工作未被公开。这种速度与谨慎之间的冲突，源于竞争压力、早期实验室的文化DNA以及衡量安全成果的困难。要解决这一难题，需要将安全案例的发布与代码本身同等重要，建立行业安全标准，并培养工程师的全面安全责任感，确保AI的进步与安全并行。

🦺 **AI行业的“安全-速度悖论”：** 文章指出，AI行业普遍存在一个深层结构性冲突，即在追求快速迭代以保持竞争力的同时，也面临着保障AI安全的道德责任。这种矛盾使得AI公司在速度和谨慎之间艰难抉择，尤其是在“三马车竞赛”的背景下。

🏃 **高速发展下的挑战与代价：** 以OpenAI为例，公司在短时间内人员激增，导致管理上的“可控混乱”。像Codex这样革命性产品的快速开发，虽然展示了惊人的执行力，但背后是工程师们高强度的工作，也凸显了在追求速度时，安全研究的缓慢和系统性发布可能被视为一种“干扰”。

⚖️ **竞争压力与文化基因的双重影响：** 这种困境并非源于恶意，而是多重因素交织的结果。首先是激烈的市场竞争，其次是早期AI实验室“科学怪人”式的文化，他们更看重颠覆性突破而非循序渐进的过程。此外，量化安全成果的难度远大于量化速度和性能，使得可见的“速度”指标在决策中占据主导地位。

💡 **重塑规则以实现安全与速度并存：** 文章提出，解决之道在于改变游戏规则。这包括将安全案例的发布视为与代码同等重要的环节，建立行业通用的安全标准，避免单一公司因严谨而承担竞争劣势。最重要的是，要在AI实验室内部培养一种文化，让每位工程师都承担起安全责任，认识到AGI竞赛的关键在于“如何抵达”，而非“谁先抵达”。

A criticism about AI safety from an OpenAI researcher aimed at a rival opened a window into the industry’s struggle: a battle against itself.

It started with a warning from Boaz Barak, a Harvard professor currently on leave and working on safety at OpenAI. He called the launch of xAI’s Grok model “completely irresponsible,” not because of its headline-grabbing antics, but because of what was missing: a public system card, detailed safety evaluations, the basic artefacts of transparency that have become the fragile norm.

It was a clear and necessary call. But a candid reflection, posted just three weeks after he left the company, from ex-OpenAI engineer Calvin French-Owen, shows us the other half of the story.

French-Owen’s account suggests a large number of people at OpenAI are indeed working on safety, focusing on very real threats like hate speech, bio-weapons, and self-harm. Yet, he delivers the insight: “Most of the work which is done isn’t published,” he wrote, adding that OpenAI “really should do more to get it out there.”

Here, the simple narrative of a good actor scolding a bad one collapses. In its place, we see the real, industry-wide dilemma laid bare. The whole AI industry is caught in the ‘Safety-Velocity Paradox,’ a deep, structural conflict between the need to move at breakneck speed to compete and the moral need to move with caution to keep us safe.

French-Owen suggests that OpenAI is in a state of controlled chaos, having tripled its headcount to over 3,000 in a single year, where “everything breaks when you scale that quickly.” This chaotic energy is channelled by the immense pressure of a “three-horse race” to AGI against Google and Anthropic. The result is a culture of incredible speed, but also one of secrecy.

Consider the creation of Codex, OpenAI’s coding agent. French-Owen calls the project a “mad-dash sprint,” where a small team built a revolutionary product from scratch in just seven weeks.

This is a textbook example of velocity; describing working until midnight most nights and even through weekends to make it happen. This is the human cost of that velocity. In an environment moving this fast, is it any wonder that the slow, methodical work of publishing AI safety research feels like a distraction from the race?

This paradox isn’t born of malice, but of a set of powerful, interlocking forces.

There is the obvious competitive pressure to be first. There is also the cultural DNA of these labs, which began as loose groups of “scientists and tinkerers” and value-shifting breakthroughs over methodical processes. And there is a simple problem of measurement: it is easy to quantify speed and performance, but exceptionally difficult to quantify a disaster that was successfully prevented.

In the boardrooms of today, the visible metrics of velocity will almost always shout louder than the invisible successes of safety. However, to move forward, it cannot be about pointing fingers—it must be about changing the fundamental rules of the game.

We need to redefine what it means to ship a product, making the publication of a safety case as integral as the code itself. We need industry-wide standards that prevent any single company from being competitively punished for its diligence, turning safety from a feature into a shared, non-negotiable foundation.

However, most of all, we need to cultivate a culture within AI labs where every engineer – not just the safety department – feels a sense of responsibility.

The race to create AGI is not about who gets there first; it is about how we arrive. The true winner will not be the company that is merely the fastest, but the one that proves to a watching world that ambition and responsibility can, and must, move forward together.

(Photo by Olu Olamigoke Jr.)

See also: Military AI contracts awarded to Anthropic, OpenAI, Google, and xAI

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.

Explore other upcoming enterprise technology events and webinars powered by TechForge here.

The post Can speed and safety truly coexist in the AI race? appeared first on AI News.