Highlights

Andy Konwinski, cofounder of Databricks and Perplexity AI, argues that AI safety rhetoric is being used to centralize control over frontier models rather than truly reduce harm. He points to Anthropic’s brief decision to degrade outputs for suspected competitors as an example of private labs assuming unilateral authority. Konwinski warns that concentrating AI power creates structural risks comparable to other infrastructure monopolies, and proposes a research commons with frontier-scale compute to let top researchers access capabilities without relying on private gatekeepers.

Sentiment Analysis

• The overall sentiment of the piece is mixed-to-critical. It highlights concern and skepticism toward industry safety narratives while endorsing openness and decentralization as remedies. The tone blends caution about current trends with optimism for alternative institutional solutions. The criticism is pointed at private companies that treat safety as a reason to restrict access, and at messaging that can stoke public fear to justify consolidation. At the same time, proponents of openness—such as Yann LeCun—are presented positively, framing decentralization as the healthier long-term trajectory for AI infrastructure.

Article Text

Andy Konwinski, known for cofounding Databricks and Perplexity AI, recently published an essay arguing that current discussions about AI safety can function as a pretext for concentrating power rather than preventing harm. He traces this dynamic to concrete industry behavior and to broader institutional incentives. Konwinski’s central claim is that when leading labs control the most powerful models and set the rules for access and permissible use, they reshape the risk landscape: centralization does not eliminate hazards so much as transform them into systemic dependencies and points of control.

His critique uses a recent Anthropic incident as a focal example. When Anthropic released Claude Fable 5, its system card contained a provision—buried deep in the documentation—saying the model could degrade responses for users it suspected of using outputs to train competing systems. Researchers who discovered that clause provoked public backlash, and Anthropic quickly reversed the policy. For Konwinski, the reversal does not undercut the argument: the more important issue is that a private company made a unilateral decision about how a foundational technology should behave. In his view, such unilateralism reflects and reinforces concentrated authority over an infrastructure that will have broad societal effects.

Konwinski presented his essay following an Open Frontier working meeting he convened at the Exploratorium in San Francisco through the Laude Institute. The meeting drew about one hundred researchers and signaled interest in alternative governance and access arrangements for frontier compute. Konwinski frames AI as foundational infrastructure analogous to railroads, electricity, and the internet—technologies that historically reorganized economic and political power in favor of those who controlled the underlying layers. He warns that if access to cutting-edge AI remains gated by a small set of private actors, the resulting power asymmetries could be long-lasting.

To counter that trajectory, Konwinski proposes creating a research commons that provides frontier-scale compute to qualified researchers without requiring permission from private labs. The aim is to preserve the ability of independent scientists and institutions to test, audit, and develop models at the frontier, thereby diffusing technical capability and reducing single-node points of control. Such a commons would, in principle, allow safety work to proceed under more transparent and participatory conditions than under exclusive, proprietary models.

Voices from across the field responded. Turing Award winner Yann LeCun echoed Konwinski’s concerns publicly, arguing that concentration of power and the desire to control information are among the gravest dangers posed by current trends in AI development. LeCun offered a historical analogy, comparing closed-lab control over dissemination to what he described as a form of medieval obscurantism—likening it to the Ottoman Empire’s historical delay in adopting the printing press to preserve established authorities. His forecast is that foundational models will eventually be treated as infrastructure, become commoditized, and shift commercial value toward application layers rather than the base models themselves.

LeCun’s own pathway reflects that perspective: after leaving Meta, he founded AMI Labs in Paris, funded to pursue open research and to explore architectures that emphasize world models and general representations. He has signaled an intention to release research publicly over time, embodying the argument that openness and shared development can be an alternative to concentrated control. For Konwinski and like-minded researchers, such initiatives illustrate practical approaches to the problems they identify.

The debate also touches on messaging from major firms ahead of key corporate events. Observers noted that safety-oriented communications around IPOs and other strategic moments can serve as persuasive tools, shaping public perception and regulatory attitudes in ways that advantage incumbents. Critics argue that this dynamic risks conflating genuine safety needs with competitive strategies to limit rivals’ access to essential capabilities.

Overall, the discussion reframes AI safety as a socio-political challenge as much as a technical one: who governs the base layers of capability matters for how risk is distributed, what research is permitted, and which actors gain lasting leverage. Konwinski’s position is not that safety concerns are unfounded, but that they should not become the primary justification for locking down critical resources. He stresses that preserving broad, equitable access to frontier compute is essential to both robust safety research and democratic oversight of transformative technology.

As the field continues to evolve, conversations about governance, access, and the institutional forms of AI development will likely shape which technical pathways are pursued and which social outcomes become more probable. The tension between centralized control and open, distributed capacity is central to those choices.

Key Insights Table