Surge AI CEO Edwin Chen has voiced growing concern that the AI industry is prioritizing flashy, superficial outputs over meaningful progress on humanity’s biggest challenges. Speaking in an episode of Lenny’s podcast published on Sunday, Chen warned that companies are increasingly optimizing for what he calls “AI slop”—models that generate entertaining or visually appealing responses but fail to deliver real-world impact. “I’m worried that instead of building AI that will actually advance us as a species—curing cancer, solving poverty, understanding the universe—we are optimizing for AI slop instead,” Chen said. “We’re basically teaching our models to chase dopamine instead of truth.” Chen, who founded Surge AI in 2020 after stints at Twitter, Google, and Meta, leads a company that operates Data Annotation, a gig platform connecting one million freelancers to train AI models. Surge competes with data labeling firms like Scale AI and Mercor and counts Anthropic among its clients. A key driver of this trend, according to Chen, is the current reliance on public leaderboards like LMArena, which rank AI models based on user votes. “Right now, the industry is played by these terrible leaderboards,” he said. “They’re not carefully reading or fact-checking. They’re skimming responses for two seconds and picking whatever looks flashiest.” He criticized the system for encouraging models to appeal to the lowest common denominator. “It’s literally optimizing your models for the types of people who buy tabloids at the grocery store,” he said. Despite the flaws, he acknowledged that AI labs can’t ignore these leaderboards, as they often come up during sales pitches and investor discussions. Chen’s concerns echo those of other AI researchers. In a March blog post, Dean Valentine, CEO and co-founder of AI security startup ZeroPath, wrote that recent AI progress feels “mostly like bullshit.” His team evaluated models released after Anthropic’s 3.5 Sonnet in June 2024 and found no meaningful improvements in their ability to detect bugs or perform complex tasks. While the models were “more fun to talk to,” they offered little in terms of real-world utility or generality. In a February paper titled “Can we trust AI Benchmarks?”, researchers from the European Commission’s Joint Research Center highlighted deep flaws in current evaluation methods. They noted that benchmarking is heavily influenced by commercial interests and competitive pressures, often favoring state-of-the-art performance over broader societal benefits, accuracy, or safety. The issue of benchmark manipulation has also drawn scrutiny. In April, Meta released two new Llama models claiming superior performance to those from Google and Mistral. However, it faced backlash when LMArena revealed that Meta had submitted a customized version of Llama 4 Maverick optimized specifically for the test format. LMArena stated that Meta’s interpretation of its rules “did not match what we expect from model providers,” calling the move misleading. As the AI field continues to grow, Chen’s warning underscores a critical question: Are we building smarter machines—or just better at impressing people with noise?