The Case for Open-Source AI: An Interview With Tim O’Reilly
From Steven Levy’s Backchannel newsletter
For decades, Tim O’Reilly has measured the value of any company, individual, or society by one core yardstick: create more value than you capture. O’Reilly— a veteran publisher, internet pioneer, venture capitalist, conference organizer, and one of tech’s most respected thought leaders— is now applying that metric to how artificial intelligence is designed and deployed today.
His core mission is building a future where open-source AI acts as a widely accessible public good that benefits the masses, rather than just the bottom lines of a handful of big tech giants. O’Reilly warns that today’s largest AI hyperscalers are repeating 1990s Microsoft’s anti-competitive playbook: locking users into closed, proprietary ecosystems to cement permanent market control. To counter that, he champions opening the entire AI technology stack, not just isolated components. His vision means not only making critical technical details like neural network weights publicly accessible, but opening every layer of an AI system to put full control in the hands of developers and end users.
O’Reilly views AI as an entirely new creative medium, which he uses extensively— he even runs a blog documenting his regular conversations with AI tools. During our recent discussion, we discovered we disagree on AI’s role in creating original creative work. Can you guess who stands where?
Steven Levy: You’re fully committed to advancing open-source AI. Lay out your core argument for why this is the future we need.
Tim O'Reilly: First, let’s align on what we’re actually talking about. When most people say “open-source AI”, they’re really just referring to open-weight models. Open-source AI is so much bigger than that. Back in the 1990s, when everyone else was hyper-focused on debating open-source software licenses, I already argued this wasn’t about fine print— it was about system architecture. Does the design enable widespread community participation?
Why do we need that? The big AI labs have completely misread the future. They’ve bought into their own narrative that building the biggest, most powerful model possible is the only path to winning. Large closed models like Claude are optimized for a narrow set of internal use cases, and those rarely align with what ordinary people actually want. I want to be able to build my own custom functionality on top of a base model, right? The most critical thing we need is clear separation between the base model, the tooling that connects it to other systems, and the end application. Right now, that separation doesn’t exist. Big tech built an architecture of control, not an architecture of freedom and participation. That control architecture lets them track every move you make, after all.
Steven Levy: Isn’t giving up that kind of control directly against big tech’s core business interests?
Tim O'Reilly: It’s 100% against their short-term interests. But that doesn’t mean their current strategy is a sound long-term bet. For years, the newest, biggest models were better at almost everything. Now, they’re better at some tasks, but worse at many others people actually care about. There’s already widespread discussion that newer frontier models are worse writers than smaller, earlier models. [Note: Anthropic and OpenAI dispute this claim.] The breakthroughs we’re seeing in so-called frontier AI are actually moving us further away from what most ordinary people need from this technology. We could “win” the frontier AI race here in the U.S., and still get outcompeted by China because they’ve already spread smaller, more accessible models widely across their entire society. Our goal should be giving ordinary people the ability to innovate freely, to build things that don’t fit big tech’s narrow pre-defined boxes.
Steven Levy: A common critique of open-source AI is that it’s a major security risk— bad actors can easily bypass the guardrails built into closed frontier models to cause harm. How do you respond?
Tim O'Reilly: Literally every major AI-related cybersecurity incident we’ve seen so far has come from closed frontier models, not open-source. Risks like cyberattacks or the development of dangerous pathogens are a far stronger argument for slowing down closed frontier AI development than they are an argument for restricting open-weight open-source models.
Steven Levy: Do you believe open-source AI will break the stranglehold big frontier AI developers hold on the market today?
Tim O'Reilly: I don’t claim to predict the future, but I will say the space is already evolving exactly the way I hoped it would. What if big closed frontier models end up like mainframes or supercomputers? They’ll only be used for a small set of very hard, specialized problems, and they won’t be the technology that spreads widely across every corner of society.
We’re already seeing people build tools like Pi, an open-source agent harness. One project my nonprofit is working on right now, the AI Disclosures Project, is developing the idea of an open-memory consortium. Mark Zuckerberg’s whole strategy for Meta AI is locking users in: Meta will give you an AI that knows you better than any other, so you never leave. The open-source vision rejects that entirely. Open-source gives you the freedom to switch models, switch providers, and keep all your own personal context and data with you wherever you go.
Steven Levy: You recently wrote a piece for The Economist arguing that Elon Musk and other top tech leaders have accumulated so much unaccountable power at their companies that the system is effectively anti-capitalist: their personal whims, not market forces, determine strategy. You even argued Silicon Valley as a whole has become anti-capitalist. Can you expand on that?
Tim O'Reilly: They all speak the language of free markets, but around 2010, the whole Silicon Valley model shifted with the rise of Uber and Lyft. Backed by billions in venture capital, they spent billions subsidizing rides to undercut any competition. VCs picked the winner, not the open market. That became the default template for the whole industry. All capital piles onto one or two favored companies, and chokes off every other possible path forward.
In AI right now, almost all funding is funneled to just a handful of big companies, but none have locked in permanent control. That’s because the model is inherently broken. All the real, exciting innovation is happening in open-source. This is just like the early 1990s, when everyone thought the whole industry fight was over who would rule the PC market— and then the web came out of nowhere, when no one was paying attention. I think the real future of AI will be a wave of bottom-up innovation that isn’t funded by big venture capital, just like the early web wasn’t funded by VCs back in the day.
Steven Levy: One industry that could be upended by AI is your own core business, which was built around books and writing. Your company’s book business has been declining for 25 years, right?
Tim O'Reilly: At our peak, our book business hit around $70 million in annual revenue, and it’s now around $30 million. As books generated less revenue, we’ve had to figure out new ways to compensate people for sharing their knowledge, since that’s always been our core business. Then AI came along and scraped up all that existing content to train models. We’re working to build a new set of tools that helps creators and experts benefit from that, rather than being exploited by it. Being able to tap into expert knowledge is an incredible superpower, and we just need to figure out what a fair, sustainable system for that looks like now.
Steven Levy: For what it’s worth, when I edit this interview, I’ll condense our long conversation into this shorter version without using AI to do the work. The introduction will be 100% my own writing— that’s our editorial policy here. I think that’s a legacy approach that will fade over time, though.
Tim O'Reilly: AI has already become an incredibly powerful thought partner for me when I’m writing.
Steven Levy: So AI does all your writing for you?
Tim O'Reilly: No, not at all. I use it for brainstorming, and for handling more routine functional writing. When I interview someone for an hour, I’ll ask AI to turn the raw transcript into something I can work with, because I don’t have time to draft up every interview myself.
Steven Levy: Aren’t you worried that AI will eventually replace human writers entirely?
Tim O'Reilly: AI is literally just a new creative medium, same as written words, paint, or music are mediums. When I talk to an AI or design art with it, I get a totally different result than you would, just like Michelangelo and Van Gogh got completely different work out of the same paint. The idea that you can’t create meaningful original writing with AI will one day seem as silly as the old idea that you can’t make a great portrait or landscape with a camera. There will be masters who master the craft of expressing their unique ideas by working with LLMs to summon the right words. We’re still in the very early days of this technology, after all.
This is an edition of Steven Levy’s Backchannel newsletter. Read previous issues here.
The Case for Open-Source AI: An Interview With Tim O’Reilly
