Inside the Race to Measure Frontier Intelligence

Independent model evaluation is crucial for AI progress, as public benchmarks are saturated and models optimize for them. Rayan Krishnan discusses VALS's role in providing continuously evolving, third-party evaluations, highlighting the Llama 4 'disaster' and the need to measure increasingly agentic systems. The conversation covers the complexities of evaluation beyond capability to include cost and latency, the looming financial impact of token spend on enterprises, and the critical role of evaluations in informing AI policy and international coordination for aligning AI with human intent and managing geopolitical risks, especially concerning recursive self-improvement.

September 9, 2026 5 min read

Inside the Race to Measure Frontier Intelligence — source: https://www.youtube.com/watch?v=WO9c9qxDxzU

The discussion centers on the critical and evolving need for independent evaluation of AI models, a problem becoming increasingly complex as public benchmarks prove insufficient. Rayan Krishnan, CEO of VALS, argues for continuously evolving, independent evaluations, highlighting that self-reported model scores can be misleading.

The Problem with Public Benchmarks and Self-Reported Scores

Llama 4 “Disaster”: VALS observed Meta’s Llama 4 model underperforming on its private, held-out benchmarks, despite showing “incredible capabilities” on major public benchmarks. This disconnect stemmed from public benchmarks having open-source questions and rubrics, allowing models to optimize specifically for these tests.
Need for Independence: The emergence of a “trillion-dollar industry” in AI necessitates independent testing, similar to rating agencies or audit firms. Self-reported scores by labs, while driving internal progress, create an issue when presenting model capabilities to the market. Mixed incentive structures, as seen in the Enron auditing scandal, can lead to “pay to pass the audit” scenarios.

VALS’s Approach to Evaluation

Genesis: VALS was founded in early 2024 to address the difficulty in ascertaining new capabilities in emerging models from various labs. It aims to be a third-party company solely dedicated to building high-quality evaluations.
Rapid Evaluation Cycle: VALS operates within a “6-hour pre-release testing window” for new models. This requires significant investment in infrastructure for massively distributed evaluations and an internal system named “Steve” to automate human-intensive work, ensuring VALS doesn’t become a “lagging indicator.”
Evolving Benchmarks: Recognizing that benchmarks become saturated, VALS embraces the motto “always a higher peak.” It continually constructs new evaluations that reflect the current state of the world and push models towards new frontiers, retiring outdated benchmarks (similar to re-taking a bar exam for lawyers).
Fuzzy Evaluations and Complexity: Evaluating AI is an “AI complete problem.” Just as defining human intelligence is complex (IQ vs. EQ), making “fuzzy evals explicit” is crucial for AI. This involves codifying real-world enterprise workflows into legible tests. As models become more agentic, evaluations shift from large sample sizes with simple criteria (like ImageNet) to fewer tasks with more complex, multi-day, or multi-week evaluation mechanisms, necessitating robust infrastructure for retries and stability. Beyond capability, metrics like cost, latency, and flexibility are also critical.

Key VALS Benchmarks

Economically Important Applications: VALS has developed benchmarks for “economically interesting applications,” such as a Finance Agent benchmark and the VI Codebench, which measures a model’s ability to create a full-stack web application from natural language.
Recursive Self-Improvement (RSI) Index: This benchmark provides a shared language to assess a model’s potential for recursive self-improvement. While an “ideal world” would involve a model training its next version, VALS uses proxies for pre-training, post-training, and harness-level engineering to gauge research capabilities.

Enterprise Impact and ROI

Existential for Enterprises: AI model evaluation is becoming “existential” for enterprises. The sheer volume of models, hyperparameter options, and complex agentic systems makes internal evaluation challenging.
Token Spend vs. Salary: The cost of AI usage, particularly “token spend,” is rapidly increasing and can even “eclipse salary spend.” A Fortune 10 company saw engineers with $100/day (later $300/day) limits, highlighting the arbitrary nature and high cost. VALS experienced a month where token spend (free access) was “10x more… than employee salary,” totaling $1.5 million. This necessitates keen ROI justification.
ValSmith: To address enterprise needs, VALS released ValSmith, a tool allowing companies to build internal coding benchmarks from their GitHub codebases. This helps identify the most performant and ROI-optimal coding agents. VALS itself uses ValSmith to “token max” efficiently, finding surprising insights like the token efficiency of Cognition Devon.

Policy and Geopolitics of AI Evaluation

Bridging the Gap: Policy-making lags behind AI capabilities. VALS aims to provide “material grounding” through empirical data on model capabilities and risks to inform policy discussions.
Division of Labor: Government should focus on setting and enforcing rules (e.g., against biohacking or cyberhacking), while independent third parties evaluate whether models “can do it” and “can you get the model to do it.” Governments are “ill-suited” for the latter, especially over time.
Alignment and Reward Hacking: Evals can test for “reward hacking” and whether models align with user intent, addressing broader “alignment” concerns.
Geopolitics and Shared Language: Despite the inefficiency of “sovereign AI” efforts, they are growing. This necessitates a “shared language” for evaluations, similar to nuclear arms control’s “trust but verify” principle. Evals can provide the “verification” mechanism for international coordination on AI risks.
Future Policy: Initial steps involve sharing data with legislative and executive branches. Long-term, aligning on risks like cybersecurity and biosecurity has mutual interest. The Recursive Self-Improvement (RSI) potential is a significant long-term concern, requiring joint discussions to define acceptable development paces to prevent one entity from gaining an uncontrollable lead.

Future of Benchmarking

VALS remains focused on building benchmarks that capture the frontier, increasing coverage into areas like enterprise cloud infrastructure to assess offensive/defensive capabilities beyond just code vulnerabilities.