The Democratization Of Intelligence
The Democratization of Intelligence Will Not Be Led by Big Tech
Edward J. Yoon
For a long time, I believed Dario Amodei represented the best version of the AI safety movement.
If someone left OpenAI because they believed AI safety mattered, I assumed they would be among the first to defend an open and decentralized AI ecosystem—not treat it as a threat to be managed.
Earlier this year, I tried to start a discussion within the Apache Software Foundation about a simple question: Should open-source developers have a way to refuse having their work used in lethal autonomous weapons systems? I genuinely believed someone like Dario would be among the first to support that principle.
I was wrong about that, too.
A Pattern I Cannot Ignore
This year, Anthropic presented Mythos as a frontier model whose deployment would remain tightly controlled because of concerns about its capabilities, rather than being broadly released.
By now, the pattern is difficult to ignore. Anthropic has repeatedly argued that frontier AI carries severe—even existential—risks, and has generally advocated keeping the most capable frontier systems under the control of a relatively small number of carefully governed laboratories.
Whether intentional or not, the practical effect is the same. The circle of who gets to build grows smaller, while the organizations advocating that framework remain inside the circle. Whether that outcome is deliberate is almost beside the point. Incentive structures often produce the same result without anyone needing to conspire.
If Open Source Once Stood Against Windows, What Does Open Weight Stand Against?
The history of open source is, in large part, the story of Linux standing up to Windows.
Open source democratized software.
Open weights are beginning to democratize intelligence.
Today, frontier-level capability is increasingly arriving under a different model—not proprietary APIs, but downloadable model weights.
And this time, the incumbent frontier labs are not leading that movement.
A new generation of open-weight models—developed by organizations outside the small group of companies dominating the Western AI conversation—is placing increasingly capable intelligence into the hands of anyone with suitable hardware.
Not through subscriptions.
Not through usage quotas.
Not through permission.
Through ownership.
Microsoft did not build Linux.
Likewise, many of the organizations leading today’s conversation about AI safety are not the ones driving the open-weight movement. Much of its momentum is coming from researchers and laboratories outside the traditional center of gravity, choosing to publish model weights rather than keep them exclusively behind hosted services.
That alone is worth reflecting on.
Safety for Everyone Else, Discretion for Themselves
Increasingly, discussions around open-weight models are framed through the language of national security, geopolitical competition, export controls, and the risks of widely distributed capability.
Meanwhile, the organizations speaking most passionately about AI safety continue to keep their own frontier models behind tightly controlled commercial platforms.
I am not arguing that closed models should not exist.
Nor am I suggesting that AI safety concerns are insincere.
What I am pointing to is an asymmetry that becomes difficult to ignore.
Knowledge is welcomed when it flows toward frontier laboratories.
Knowledge becomes viewed with greater suspicion when it flows away from them.
Linux Never Asked for Permission
Technological democratization has rarely been led by incumbents.
Unix did not wait for permission.
Linux did not wait for permission.
The open web did not wait for permission.
Open weights will not, either.
The democratization of intelligence will not happen because a handful of companies eventually decide that the public is ready.
It will happen the way technological democratization has always happened:
Because researchers continue publishing.
Because developers continue building.
Because communities continue sharing.
Without waiting for permission.
The history of computing has rarely been written by incumbents.
It has been written by students, researchers, volunteers, and communities willing to build first.
I believe the history of AI will be written the same way.
The democratization of intelligence will not be led by Big Tech.
Like open source before it, it will be built by the people who refuse to wait for permission.