OpenAI capitulates on open-source models
OpenAI’s decision to go proprietary just as the world was pivoting open will have long-term consequences that will take serious time and careful work to fix
OpenAI is quietly atoning for recent sins, and the fallout will have critical implications for the future of global artificial intelligence.
Last week, the company launched 4o Image Generation, leading to such a deluge of Studio Ghibli memes across social media that CEO Sam Altman had to write, “our GPUs are melting” and temporarily turn off some functionality. “[T]he chatgpt launch 26 months ago was one of the craziest viral moments i'd ever seen, and we added one million users in five days,” Altman later noted. After the launch of 4o Image Generation, “we added one million users in the last hour.”
Tucked away amid the bread and circuses though was a massive strategic announcement: for the first time since the launch of its GPT-2 model in 2019, OpenAI said yesterday, the company will publish an open-weights AI model to the community. Altman wrote that “we’ve been thinking about this for a long time but other priorities took precedence. now it feels important to do.”
Six years in the artificial intelligence world is practically millennia. The torrential pace of research and product development means that even last week’s papers are outdated. This isn’t Greek history — the field changes in real time.
The story of OpenAI is one of the great reversals in startup history (or betrayals, if you are Elon Musk). Founded as a non-profit in 2015, OpenAI’s mission was to congregate the world’s best AI research scientists to develop open and safe AI models for the public. These technologies would be so powerful and influential, the thinking went, that no single company should have exclusive access to them. Instead, they should be offered liberally in the pursuit of human flourishing.
Indeed, that’s exactly what OpenAI did during its first few years, freely offering its frontier models to the community. That all changed in 2019 with the launch of GPT-2. Initially, the company refused to release the full source code for the model before ultimately releasing the full version later that year after persistent criticism. That was the canary in the coal mine: future models would not be open, and following the epic and public board room fight on OpenAI’s future at the end of 2023, the once non-profit entity announced its transition to a for-profit company at the end of last year. Given the intense drama of that transition, yesterday’s quiet announcement that the company will once again offer an open-weights model is head-spinning.
Why the sudden strategic reversal? The immediate trigger is DeepSeek, which launched its open-source V3 model on December 26, the day before OpenAI’s announcement that it was transitioning to a for-profit structure. As I wrote at the time, DeepSeek’s superior performance coupled with its fully open model meant that it was now the high-water mark of global AI.
Let’s pause for a second to relish the deliciousness of that timing. Just as OpenAI was firming up its fully-proprietary business strategy, the world had instantly shifted to open-source models as the future. The AI world truly changes in real time.
Yet, the competitive pressure from open-source isn’t just coming from a little-known Chinese hedge fund or from Alibaba’s Qwen, but also from America’s largest technology companies. Due to its strategic position, Meta has emphasized an open-source approach to AI, offering open-weight and public-source versions of its Llama models. Google DeepMind has now distilled its proprietary Gemini model into the open-source Gemma family, launching the third version last month to much anticipation.
In short, nearly every company has realized that open-source models will be crucial for capturing the future of the AI marketplace. It’s a lesson even DC policymakers are increasingly grokking. In a recent essay in Foreign Affairs, Jared Dunnmon wrote:
Clearly, the United States can no longer rely solely on closed AI systems from big companies to compete with China, and the U.S. government must do more to support open-source models even as it strives to limit Chinese access to cutting-edge chip technologies and training data. To continue its dominance, the United States should mount a comprehensive program to develop and deploy the best open-source LLMs while also ensuring that U.S. firms are still the ones building the most capable AI models—sometimes called “frontier systems”—that are still likely to reside within highly capitalized private companies.
While not calamitous, OpenAI’s decision to go proprietary just as the world was pivoting open will have long-term consequences that will take serious time and careful work to fix.
First, the company’s extremely public strategic reversal has broken the trust of many developers, customers and even the occasional policymaker. That six-year gap between GPT-2 and whatever it launches later this year cannot be erased. For those building on top of OpenAI’s products, how will they ensure that OpenAI doesn’t go fully closed-source once again? Altman said priorities other than continuing to publish open models “took precedence” these past few years. Will those priorities once again change in the future?
Not unlike similar crises at Twitter when it shut down its developer platform or Reddit when it restricted its APIs, it’s a long road to rebuild trust with developers and customers who once relied on a product and suddenly had it yanked from their hands. Of course, developers will slither back to a company if it becomes profitable to do so or if that company has a competitive product that is irreplaceable. Unfortunately for OpenAI, the extreme competition between AI foundation models gives developers many alternatives that will make its attractiveness hard to demonstrate.
Second, as one of the most crucial national AI champions for the United States, OpenAI is an integral leader in shaping the country’s policies toward AI systems. On open-source technology though, it got the strategy entirely wrong. If its proprietary approach had found greater and more immediate purchase among politicians, America might have disastrously pulled itself out of the AI race with China right in front. Thankfully, America’s political dysfunction protected itself this time, but it’s a reminder that there should be clear separation of the business goals of a handful of leading companies and the wider technology dominance the country needs.
Third, OpenAI confirmed yesterday that it is closing on a $40 billion round of capital at a $260 billion valuation. For the venture capitalists funding the company, it’s crucial to recognize that even as OpenAI clearly continues to build consumer market share at a torrential pace, it lost its grasp on the talent at the frontiers of AI technology. Behind the vanguard is a very dangerous place to find oneself in the whitewater rapids of AI development. Launching an open-weights model is a partial rectification, but the company has a lot more to do to recapture its once almost monopolistic hold on the technology.
OpenAI and Altman himself clearly made mistakes over the past few years regarding open-source. We shouldn’t criticize the recognition of those mistakes or new efforts to repair them, but neither should we ignore that the company zigged when the world zagged. Will it get the next critical decision right? The future of a now quarter trillion dollar company — and American economic security in the 21st century — rides on that answer.