Ideology – not security – gave birth to open source.

In September 1983 Richard Stallman, a then-software programmer at MIT, posted an online message announcing he was going to build GNU, a free, Unix-compatible operating system. Richard’s motivation was simple, albeit unpopular at a time when proprietary software was the norm: he could not violate his own “golden rule [that] requires that if I like a program, I must share it with other people who like it.”

What Stallman called “free software” at the time eventually became “open source”, but not until 15 years later when Netscape decided to release its browser source code to the public. Apparently “free software” didn’t properly connote value; thus, was born “open source.”

Nor could Stallman have predicted in 1983 that his free GNU software would ultimately become Linux. In 1992 Linus Torvald’s Linux kernel was combined with Stallman’s GNU tools to create what we know now as the Linux operating system.

Most interestingly, Linux is the dominant computer operating system still today – and it remains open source. As much as 90% of the public cloud infrastructure runs on Linux and two-thirds of servers and mobile devices run on Linux. Open source won.

So, why are we talking about a 43-year-old mailing list post?

Because open source matters and is misunderstood in the current Super Intelligence (SI) safety and security debate. The argument generally goes like this: the safest SI is SI kept in a few vaults, controlled by a few large, proprietary model companies. “Closed source” in fewer hands = less risk. It sounds intuitive.

Not so fast – think about the technology that actually runs much of what you as a consumer (and businesses) run on today. How much of it is closed? This is not a trick question – although it probably isn’t obvious to most consumers.

The internet runs on open standards – TCP/IP, DNS, HTTP, and HTML – that anyone can read, implement, and improve. Web encryption runs on open implementations such as OpenSSL. Google Chrome and Microsoft Edge, despite the two companies being bitter competitors, share the same underlying open engine. Data centers run on Linux, Git, Docker, and Kubernetes (by the way, which Google invented, gave away and now runs on its rivals’ clouds too!). And nearly every college student learning to build with SI runs VS Code and Homebrew, free and open.

Even the vaults are built on open source. In 2022 Meta handed PyTorch to the Linux Foundation. And nearly all of PyTorch is written in Python, a free, open language supported by a nonprofit for the past twenty-five years. In effect, “closed” companies are built on open source code.

Why has open source essentially won?

Security – it turns out that the wisdom of the crowd to identify and fix bugs works better than simply relying on the badged employees alone. Open source software advocate Eric Scott Raymond wrote about this in his seminal 1997 essay “The Cathedral and the Bazaar”. The Bazaar wins because "given enough eyeballs, all bugs are shallow".

Cost – open source injects price competition into the market, inuring to the benefit of all consumers, without sacrificing quality. In fact, it typically takes just a matter of months for frontier open source models to reach performance parity with closed models, yet often at as much as 70% lower overall cost to users.

Customizability – open source allows consumers to adapt frontier models to the specific needs of their personal or organizational use cases, at lower cost and greater specificity. Brian Armstrong, CEO of Coinbase, wrote about this recently in his company’s adaptation of an open source model to create a better, lower cost, faster processing fraud prevention tool. As does Brian, I suspect we will see many more “smaller” model adaptations of open source to use cases that are more customized to the specific needs of users.

The U.S. government has long recognized the importance of a thriving open source ecosystem. For example, all recipients of federal research grants are required to make available their code for open inspection. As part of President Trump’s “Restoring Gold Standard Science” executive order last year, the President expanded this requirement to medical research by requiring data openness for grant recipients.

This same logic holds for SI.

First, you can’t secure what you can’t inspect. Interpretability research, red-teaming, bias audits, and backdoor detection all require looking inside the model. Open models let thousands of independent researchers do that work, rather than hoping one lab’s internal team catches everything.

Second, most organizations don’t need the biggest model for all things. Managing customer support inquiries, summarizing contracts, and extracting data from forms are narrow jobs. A small open model that you run on your own infrastructure, tune on your own data, and keep entirely behind your own walls is often more secure than shipping sensitive information to someone else’s API. Fewer moving parts, smaller attack surface, and you decide where the data goes.

Third, resilience. As my colleague Adam Starr wrote recently, attackers are using SI to shrink the time between a known flaw and a working exploit from months to hours. Defenders need SI, too, and they need to be able to run it where their most sensitive systems live. A world where critical SI capabilities depend solely on three or four providers is a world with single points of failure, whether technical, commercial, or political. Open models are the hedge.

Finally, setting security aside, open models are good for SI consumers. They introduce price competition that bends the cost curve for all – whether users of open source or proprietary models.

To be clear, however, the development of SI models is not an “either/or” but rather an “and”. Linux and proprietary operating systems have coexisted for decades. And that is a good thing. A healthy SI ecosystem will have frontier closed models and thriving open ones. That is best for consumers – from both a security and cost-efficiency perspective. Everyone is better off with a plethora of choices. 

Richard Stallman may have been an ideologue in his defense of free software, but he was ultimately right about something more foundational – free is the best protection.