Perplexity Co-Founder: AI Safety Is an Excuse to Lock Down Frontier

We have reached the stage of the AI cycle where the word safety is being weaponized. It happens in every transformative technology. A few big players get an early lead, realize how much money is on the line, and suddenly start worrying about the end of the world. They claim that because the tech is so powerful, only they should be allowed to touch it.

Perplexity co-founder Andy Konwinski recently put words to what many of us in the builder community have been feeling. He argued that AI safety has become a convenient excuse to lock down frontier models and prevent anyone else from catching up. It is a classic regulatory capture play, and we have seen this movie before in finance and energy. This time, the stakes are the very building blocks of the digital future.

The Anthropic Incident and the Transparency Gap

Konwinski pointed toward the recent fallout surrounding Anthropic’s Fable 5 as a prime example of why we should not trust private labs to be the sole arbiters of progress. When a proprietary model fails or behaves unexpectedly, the public is left in the dark. There is no open code to audit, no weights to inspect, and no community-driven fix. We are simply told to trust the creators.

As a founder, I find the trust me model fundamentally broken. In the crypto world, we learned that if you cannot verify it, it is not decentralized and it is rarely secure. AI is currently moving in the opposite direction. The companies lead with a mission of benefit for humanity, but their actions suggest a desire for a walled garden. When things go wrong behind those walls, the safety flag is raised not to protect the user, but to justify keeping the walls high.

Defining the Frontier Paradox

The term frontier model is being used to create a legal and psychological barrier. By defining certain levels of compute or capability as frontier, incumbents are essentially asking for a license to operate while demanding that startups face a mountain of red tape. They suggest that once a model reaches a certain level of intelligence, it becomes too dangerous for the open-source community to handle.

This is a paradox. If a technology is so critical and life-altering, it should be subject to more eyes, not fewer. Locking down the frontier ensures that the biases, blind spots, and commercial incentives of three or four companies dictate the direction of the entire industry. For builders, this is a nightmare scenario. It means you are not just building on a platform; you are building on a permissioned layer where the terms can change the moment you become a threat.

The Open Source Defense

Konwinski’s stance is a defense of the open-source ethos that built the internet. If we allow the safety narrative to win, we lose the ability to innovate at the edges. Small teams and independent researchers are the ones who usually find the creative breakthroughs or the critical security flaws. When you gatekeep the models, you starve the ecosystem of the raw materials it needs to grow.

The argument from the big labs is that an open-source frontier model could be used by bad actors. That is a real risk, but it is a manageable one. We do not ban the sale of hammers because someone might use one to break a window. We punish the person who breaks the window. In AI, the focus should be on the application and the actor, not the fundamental research. By criminalizing the research under the guise of safety, we are effectively choosing a monoculture over a diverse, resilient ecosystem.

The Commercial Reality Behind the Concern

Let’s call this what it is: a moat. Building these models costs hundreds of millions of dollars in compute alone. The companies that have spent that money want a return on investment. If a leaner, faster startup can achieve similar results using open-source weights and clever optimization, the billion-dollar moat disappears. Safety is a much more palatable marketing term than anti-competitive behavior.

When you hear a CEO of a major AI lab testifying before Congress about the risks of AI, you should be looking at their balance sheet. They are asking for regulations that they are already equipped to handle, knowing those same regulations will crush a ten-person dev shop. It is a strategic move to ensure that the next Perplexity or the next breakthrough doesn't happen in a garage, but stays within the confines of a corporate R&D department.

What This Means for Builders

For those of us in the trenches, this debate isn't academic. It determines what tools we will have access to in two years. If the gatekeepers win, your business model will always be at the mercy of an API provider. They can raise prices, censor your outputs, or shut you off entirely if your product competes with one of their internal projects.

The push for open-weights models and decentralized compute is more than just a technical preference; it is a necessity for survival. We need to support the founders and researchers who are pushing back against the lockdown. We need to demand transparency from the companies that claim to be building for the common good. If the technology is truly meant to serve humanity, then the keys to that technology cannot be held by a small committee of unelected executives.

Moving Beyond the Fear Narrative

The industry needs to move away from the binary choice of total lockdown or total chaos. There is a middle ground where safety is handled through robust testing, clear liability for misuse, and public transparency. We should be building better sandboxes, not bigger cages. Konwinski’s critique is a reminder that we are at a crossroads. We can either have an AI future that looks like the open internet—messy, competitive, and wildly innovative—or one that looks like cable television: curated, expensive, and controlled by a few incumbents.

I will take the messy version every time. The moment we stop questioning the safety excuse is the moment we stop being builders and start being tenants in someone else’s digital empire. Keep an eye on the lobbyists and the PR campaigns. When they talk about saving the world, they are usually just trying to save their share price.

Read the original at Decrypt →