George Cameron

George Cameron is the co-founder and CPO of Artificial Analysis, based in San Francisco. With co-founder Micah Hill-Smith, he has built a benchmarking platform that compares AI models and providers across intelligence, speed, and cost. His public priorities include broadening evaluation beyond raw intelligence to measure hallucination and agent performance.[3][1]

Cameron puts independence at the center of the business: “you can’t pay us for better results,” he said in the Latent Space interview published January 8, 2026. He described paid enterprise subscriptions providing standardized benchmarking reports and insights, alongside custom private benchmarking for enterprises and AI companies. A deployment report, for example, helped companies compare serverless inference, managed deployment, and leasing chips to run inference themselves.[1]

He also explained a distinction between the platform’s published evaluation cost and its actual expenditure. The reported cost of running the Intelligence Index assumed one repeat; the team spent more because it repeated evaluations. Hill-Smith explained that those repeats helped the team establish confidence intervals rather than rely on a single run.[1]

Cameron described the team’s GDPVal AA work as adapting OpenAI’s GDPVal dataset into an evaluation runnable across models. The team built a reference agent harness to perform the tasks and an evaluator using Gemini 3 Pro Preview to compare the resulting deliverables. Cameron said the evaluator had been tested extensively for alignment with human preferences. Documents, videos, and audio outputs often lacked a simple correct-or-incorrect ground truth, so the evaluation used relative Elo comparisons between model outputs.[1]

That reference harness became Stirrup, which Cameron said the team released on GitHub. He described a minimalist generalist agent with context management, web search and browsing, and code execution; Hill-Smith added that the GDPVal setup included an image-viewing tool. Cameron favored letting increasingly capable models control the workflow with a small set of tools, rather than having a more elaborate framework dictate the sequence of actions.[1]

Cameron argued for evaluating more than raw intelligence. He reported that the team’s measurements showed no strong correlation between general intelligence and a model’s willingness to acknowledge that it did not know an answer. He also qualified the goal: factual knowledge tasks benefit from avoiding invented answers, while coding and generating new ideas can benefit from a model attempting uncertain possibilities.[1]

At the 2025 AI Engineer World’s Fair, Cameron presented “Trends Across the AI Frontier.” The organizer’s recap highlighted his slide contrasting savings from distillation, inference improvements, and hardware with increased consumption from larger models, reasoning, and agents.[2]

In the Latent Space interview, Cameron predicted that demand for smarter AI would be “insatiable.” He said Artificial Analysis would continue benchmarking raw intelligence while exploring model behavior and personalities to support more nuanced choices. These were proposed directions alongside its work on hallucination.[1]

The platform began with a narrower practical need: helping its founders choose models while building AI applications. That earlier work provides context for its comparisons of capability, speed, and cost.[1]

Before Artificial Analysis, Cameron worked in strategy consulting focused on data centers and technology companies.[3]

The founders began Artificial Analysis as a side project in 2023 and launched it publicly in January 2024. Hill-Smith’s work on a legal AI research assistant had exposed repeated model-selection problems involving accuracy, performance, and cost. Cameron recalled that they built the tool for their own needs, connected a domain to their Vercel deployment, and tweeted about it. He credited swyx with an early retweet and spotlight, and said the tool became more useful as model releases accelerated.[1]

The founders subsequently joined AI Grant. Cameron described Nat Friedman and Daniel Gross as mentors who helped them work methodically through difficult startup decisions. Contact with other companies in the program also helped the founders decide how Artificial Analysis could remain useful to people building with AI.[1]