Sriram Krishnan on Open Source AI's Biggest Week Yet

In this episode of the A16Z podcast, Sriram Krishnan reflects on the significant week in open-source AI, highlighting the release of multiple competitive models including Grok, Kimi K3, Qwen 3.8, and others that rival frontier models like Claude and GPT in capability. Krishnan explains that this shift represents a fundamental change in the AI ecosystem where consumers now have genuine choices in model providers, potentially eroding the pricing power and margins of frontier labs like Anthropic and OpenAI.

Krishnan identifies several key implications of this trend. First, the availability of capable open-weight alternatives means tasks that don't require frontier-level intelligence can be served by cheaper alternatives, putting downward pricing pressure on frontier models. Second, he notes a concerning security situation where Chinese models like Kimi K3 have fewer safeguards than American models like Claude, making them paradoxically more useful for legitimate security research—a problematic position for American competitiveness. Third, this opens opportunities for the broader infrastructure stack (cloud providers, chip makers, data center operators) to capture economic value previously concentrated at frontier labs.

On the question of government policy toward restricting Chinese models, Krishnan expresses that while he doesn't have inside information, the Trump administration's stated AI action plan emphasizes the importance of open source. However, he notes the concerning reality that leading open-weight models are currently Chinese rather than American, which he views as suboptimal. He advocates for American companies to improve their open models rather than restricting Chinese ones, citing ongoing efforts from Google (Gemma), NVIDIA (Nemotron), and startups like Reflection.

Regarding distillation—the practice of training models on outputs from other models—Krishnan clarifies that distillation is foundational to all modern AI training and includes inevitable consumption of AI-generated content now prevalent on the internet. The problem isn't distillation itself but industrialized, terms-of-service-violating extraction. He highlights an important asymmetry: Chinese models can freely distill from American models, while American startups face legal uncertainty about distilling from other American models, creating an uneven competitive landscape.

On the potential for automated AI research and rapid capability improvements, Krishnan takes a pragmatic position, noting that the timeline and feasibility remain debated. Rather than speculative long-term concerns, he emphasizes focusing on credible current risks like cybersecurity threats, where AI can both pose risks and provide solutions. His approach prioritizes maintaining competitive ecosystems while addressing specific, credible threats as they emerge.

When asked about frontier labs' ability to monetize amid open-source competition, Krishnan applies capitalist principles, arguing that if open models provide real value, the entire supply chain will orient toward capturing value from them. He points to strong growth across infrastructure providers and specialized services (like Anthropic's Claude API or model fine-tuning services) as evidence that the ecosystem can thrive beyond pure frontier model licensing.