Inside the Race to Measure Frontier Intelligence

2024年初，AI模型评估领域面临一个关键的空白：公共基准不足以准确衡量进展，这导致了对新方法论的需求。这一认识促成了 Vals 的创立，这是一个独立的测试小组，致力于建立高质量的评估和基准。联合创始人 Ryan 拥有基准研究背景，他观察到“建立新的生成系统所需与实际新的评估机制之间存在紧密关系。”他指出，尽管许多有趣的模型正从各个实验室涌现，但越来越难以确定它们的真实能力。 Vals 识别出的一个关键问题是模型自报能力的问题。Ryan 引用了 Meta 的 Llama 4 发布，称其为一场“灾难”，该模型在 Vals 的“保留私有基准”上表现不佳，但在开源公共基准上却展现了“令人难以置信的能力。”这种“巨大脱节”强调了对中立第三方实体的需求，以提供客观评估，帮助实验室证明其数十亿美元投资的合理性，并培育一个“理性购买市场。”Demis (Hassabis) 等行业人士也呼吁建立这样一个第三方评估者生态系统。 从历史上看，每一个新兴的“万亿美元产业”都需要独立的测试，类似于评级机构或审计公司。Vals 旨在同时服务于模型实验室（通过验证新的模型能力）和企业（通过指导采用策略以实现最大投资回报率）。尽管实验室内部基准有利于推动进展，但它们往往缺乏外部验证所需的中立性。 在运营上，Vals 优先考虑速度和效率，以避免延迟模型发布。最初，这意味着通宵工作，但他们此后在用于“大规模分布式”评估的基础设施上投入巨资，并开发了一个名为“Steve，经济型 Vals 员工”的内部 AI 系统，以自动化人工任务。 评估 AI 模型提出了一个“AI完全问题”，类似于评估人类智能所面临的挑战（例如，智商测试与公认框架）。Vals 通过将“模糊或分布式形式的评估明确化”来解决这个问题，本质上是为可能缺乏明确人类先例的真实世界工作流程定义了性能标准。他们也承认需要避免其他行业中出现过的陷阱，例如安然丑闻中审计师也提供咨询，从而产生利益冲突。Vals 做出了一个早期决定，即“绝不向实验室出售训练数据”，以此区别于那些“创建噱头式基准作为销售数据机制”的实体。 Vals 已经开发了一个广泛的基准目录，专注于“具有经济意义的应用。”其中受欢迎的包括他们的金融代理基准和 VibeCode，后者衡量模型从自然语言提示构建全栈网络应用的能力。他们还开发实验性基准，例如递归自我改进（RSI）指数，它评估模型训练其下一版本的能力潜力。Vals 秉持“总有更高峰”的信念，弃用饱和的基准，并创建反映“世界当前状况”的新基准，推动模型走向新前沿。 模型向“代理工作”的演进已经改变了评估方式，从简单的多选问题转向可以运行“数小时、数天，有时数周”的复杂多步骤任务。这需要强大的基础设施，并转向“更少的样本量，但更大的标准或预期集合。”随着模型处理更复杂的工作流程，这种趋势将继续。 对于企业而言，“对智能的错误估值”正在成为一个“关乎存亡”的问题。Ryan 分享了一个关于 Vals 内部代币支出的轶事，该支出在一个月内达到了“150万美元的代币”，是其员工薪资支出的10倍，突显了当前代币使用限制的任意性。为解决这个问题，Vals 推出了“VALS myth”，这是一个允许公司从其 GitHub 代码库构建内部编码基准的产品。这有助于他们确定“性能最佳”和“投资回报率最高”的编码代理，因为内部测试通常会揭示前沿模型和更具成本效益的替代方案之间非直观的性能差异。 从政策角度来看，Vals 旨在提供经验数据，以促进关于 AI 风险和监管的复杂对话。尽管政府制定规则，但像 Vals 这样的独立评估者可以“测试并执行”。这种分工使政府能够专注于制定政策（例如，针对生物黑客、网络黑客），而第三方则评估模型是否能执行此类行动以及它们是否能被提示执行此类行动。 在全球范围内，“主权 AI”倡议的兴起带来了挑战。Vals 倡导一种“共享的评估语言”以促进国际合作，借鉴核不扩散的“信任但核实”原则。这一共同框架对于解决生物安全、网络安全以及“递归自我改进的可能性”等共同关切至关重要，后者可能导致一个实体迅速超越其他实体。Vals 仍致力于构建能够捕捉这些“前沿”能力和风险的基准，确保评估是动态的，并能反映不断发展的 AI 格局。

In early 2024, the landscape of AI model evaluation faced a critical gap: public benchmarks proved insufficient to accurately measure progress, leading to a need for new methodologies. This realization prompted the inception of Vals, an independent testing group dedicated to building high-quality evaluations and benchmarks. Ryan, a co-founder with a background in benchmark research, observed a "tight relationship between what it takes to build new systems for generation and actually new mechanisms for evaluation." He noted that while many interesting models were emerging from various labs, it was increasingly difficult to ascertain their true capabilities. A key issue Vals identified was the problem of self-reported model capabilities. Ryan cited the release of Meta's Llama 4 as a "disaster," where the model underperformed on Vals' "held-out private benchmarks" but showed "incredible capability" on open-source public benchmarks. This "huge disconnect" underscored the need for a neutral, third-party entity to provide objective assessments, helping labs justify their billion-dollar investments and fostering a "rational buying market." Industry figures like Demis (Hassabis) have also called for such an ecosystem of third-party evaluators. Historically, every new "trillion-dollar industry" necessitates independent testing, akin to rating agencies or audit firms. Vals aims to serve both model labs, by validating new model capabilities, and enterprises, by guiding adoption strategies for maximum ROI. While internal benchmarks within labs are good for driving progress, they often lack the neutrality required for external validation. Operationally, Vals prioritizes speed and efficiency to avoid delaying model releases. Initially, this meant all-nighters, but they have since invested heavily in infrastructure for "massively distributed" evaluations and developed an internal AI system called "Steve, the Economic Vals employee," to automate human tasks. Evaluating AI models presents an "AI-complete problem," similar to the challenges in evaluating human intelligence (e.g., IQ tests vs. agreed-upon frameworks). Vals addresses this by making "fuzzy or distributed forms of evals explicit," essentially defining the criteria for performance in real-world workflows that may lack clear human precedents. They also acknowledge the need to avoid pitfalls seen in other industries, such as the Enron scandal where auditors also consulted, creating conflicts of interest. Vals made an early decision to "never sell training data to labs," distinguishing itself from entities that create "gimmick-style benchmarks as a mechanism to sell their data." Vals has developed an extensive catalog of benchmarks, focusing on "economically interesting applications." Popular ones include their finance agent benchmark and VibeCode, which measures a model's ability to build full-stack web applications from natural language prompts. They also develop experimental benchmarks like the recursive self-improvement (RSI) index, which evaluates a model's potential to train its next version. Vals believes in "always a higher peak," deprecating saturated benchmarks and creating new ones that reflect the "current state of the world" and push models towards new frontiers. The evolution of models towards "agentic work" has transformed evaluation, moving from simple multi-answer questions to complex, multi-step tasks that can run for "hours, days, sometimes weeks." This necessitates robust infrastructure and a shift to "fewer sample size, but a larger set of criteria or expectations." This trend will continue as models handle more complex workflows. For enterprises, the "misvaluing of intelligence" is becoming an "existential" concern. Ryan shared an anecdote about Vals' own internal token spend, which reached "$1.5 million worth of tokens" in one month, 10 times their employee salary spend, highlighting the arbitrary nature of current token usage limits. To address this, Vals launched "VALS myth," a product allowing companies to build internal coding benchmarks from their GitHub codebases. This helps them determine the "most performant" and "highest ROI" coding agents, as internal testing often reveals non-intuitive performance differences between frontier models and more cost-effective alternatives. From a policy perspective, Vals aims to provide empirical data to inform sophisticated conversations about AI risks and regulation. While the government sets rules, independent evaluators like Vals can "test and enforce that." This division of labor allows the government to focus on setting policies (e.g., for biohacking, cyberhacking) while third parties evaluate whether models can perform such actions and if they can be prompted to do so. Globally, the rise of "sovereign AI" initiatives presents challenges. Vals advocates for a "shared language of evals" to facilitate international cooperation, drawing parallels to nuclear non-proliferation's "trust, but verify" principle. This common framework is crucial for addressing shared concerns like biosecurity, cybersecurity, and the "recursive self-improvement possibility," which could lead to one entity rapidly outstripping others. Vals remains focused on building benchmarks that capture these "frontier" capabilities and risks, ensuring evaluations are dynamic and reflective of the evolving AI landscape.