Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan · The a16z Show summary

Key takeaways

Independent, third-party evaluations are crucial for objectively measuring AI capabilities beyond self-reported benchmarks.
Current public benchmarks are often saturable, outdated, or misaligned with real-world contexts. Continuous, adaptive evaluation methods are necessary to track progress accurately.
Evaluation tools like VALS help enterprises optimize AI model selection and deployment, reducing costs and increasing ROI.
Benchmarking should evolve to encompass complex, agentic, and recursive AI behaviors to remain relevant as models become more advanced.
Policy and regulation should be informed by empirical evaluation data, fostering international cooperation and trust through shared standards.
Developing clear, dynamic, and incentive-aligned evaluation methodologies will be key to managing AI risks, including cybersecurity and biosecurity challenges.

Summary

The Need for Independent AI Evaluation

This section discusses the emergence of new trillion-dollar industries driven by AI innovations and the critical role of independent testing groups in assessing AI model capabilities. It highlights the issues with private, self-reported benchmarks—such as Meta's Llama 4—which can underperform in private evaluations despite showing strong public benchmark results. Ryan Krishnan emphasizes that independent, third-party evaluations are essential to objectively measure true model capabilities, especially as public benchmarks become outdated or saturated.

Challenges in Benchmarking and Evaluation Methodologies

The conversation delves into the complexities of evaluating AI models accurately. Krishnan explains VALS’s approach to evaluating AI by creating proxies for every part of the model development process, allowing rapid testing within hours before a product launch. They explore the evolution of benchmarks—removing outdated tests, aligning with current world contexts, and incorporating more complex, multi-step, and agentic tasks—including recursive self-improvement (RSI). The importance of updating benchmarks continuously to keep pace with advancing models is underscored.

Implications for Enterprises and AI Measurement

This section outlines how AI evaluation tools are becoming vital for enterprises to understand the ROI of AI investments. Krishnan discusses tools like Valsmith, which helps companies benchmark and select optimal models for coding tasks and other workflows. As AI use becomes more agentic and complex, evaluating not just raw capability but also costs, latency, and adaptability is critical. The shift toward using evals to measure value, cost efficiency, and performance in real-time is seen as a way to justify AI expenditures and inform deployment strategies.

Policy, Regulation, and International Coordination on AI Evaluation

The discussion emphasizes the importance of involving multiple stakeholders—labs, governments, and independent evaluators—in setting AI standards and regulations. Krishnan advocates for transparent, empirical data sharing with policymakers to shape effective regulations. He discusses the role of evals in geopolitical contexts, comparing them to nuclear arms verification, to foster trust and mutual verification of AI capabilities across nations. The emphasis is on building shared languages and benchmarks that facilitate international cooperation and safeguard against risks such as reward hacking or malicious use.

Future of Benchmarking and the Evolving AI Landscape

The final section contemplates how AI benchmarks must evolve alongside models and geopolitical developments. Krishnan highlights the need to develop new tests that capture frontier capabilities and emerging risks, including infrastructure-level vulnerabilities. He envisions a future where evaluation informs and guides recursive self-improvement and AI policy, helping ensure responsible development. The importance of creating incentive-aligned, high-quality evaluations that can keep up with the rapid pace of AI innovation is reiterated.

Show notes

a16z’s Erik Torenberg, Ben Horowitz, and Jennifer Li sit down with Vals founder and CEO Rayan Krishnan to discuss one of AI’s increasingly difficult problems: how do you actually measure whether a model is getting better?

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations. They unpack why self-reported model scores can be misleading, how VALS evaluates models in the hours before a release, and why measuring increasingly agentic systems means testing work that can unfold over hours, days, or even weeks.

They also explore why evals are becoming critical for enterprises trying to understand the ROI of AI, what happens if token spend begins to rival employee salaries, and how evaluations could eventually provide a shared language for everything from model routing and recursive self-improvement to AI policy and international coordination.

Resources:

Follow Rayan Krishnan on X: https://x.com/RayanKrishnan

Follow Ben Horowitz on X: https://x.com/bhorowitz

Follow Jennifer Li on X: https://x.com/JenniferHli

Stay Updated:

Find a16z on YouTube: YouTube

Find a16z on X

Find a16z on LinkedIn

Listen to the a16z Show on Spotify

Listen to the a16z Show on Apple Podcasts

Follow our host: https://twitter.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.

Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan