Key facts
- Perplexity AI co-founder Andy Konwinski argues AI safety is a pretext for concentrating power.
- Konwinski's essay followed a meeting of about 100 AI researchers convened by him.
- He criticized Anthropic's temporary measure to silently degrade responses for suspected AI trainers.
- Jennifer Chayes stated Berkeley researchers rely on Chinese models due to a lack of open Western frontier models.
- Yann LeCun echoed concerns about AI power concentration, comparing it to historical censorship.
Perplexity AI co-founder Andy Konwinski has argued that the discourse around AI safety is being weaponized to consolidate power within a few private laboratories, rather than genuinely addressing risks. In an essay published this week, Konwinski criticized companies like Anthropic for assuming the authority to make decisions about who can train competing AI models, citing Anthropic's temporary policy to silently degrade responses for suspected AI trainers.
Konwinski's essay, titled "Concentration of power in AI is a risk, not a solution," was released following the Open Frontier meeting he organized on June 30 in San Francisco, which gathered approximately 100 researchers. He posits that AI is becoming foundational infrastructure, similar to the internet or electricity, and that centralizing access creates significant risks. Konwinski advocates for a research commons that provides frontier-scale compute access to top researchers without requiring permission from private entities.
Jennifer Chayes, dean of UC Berkeley's College of Computing, Data Science, and Society, noted that Berkeley researchers are building upon Chinese models due to a perceived lack of open Western frontier models. She characterized the safety messaging from OpenAI and Anthropic, particularly ahead of their IPOs, as an "effective fear campaign."
Yann LeCun, a Turing Award winner and former chief scientist at Meta, responded to Konwinski's essay on X, stating that the concentration of power in AI is the most significant danger. He drew a historical parallel to the Ottoman Empire banning the printing press, suggesting that foundational models will inevitably become commoditized infrastructure, with future profits lying in the application layer. LeCun himself has launched AMI Labs, which plans to open-source its research.