OpenAI "We Are On The Wrong Side Of History" (of Open Source)

Sam Altman’s response to a question about releasing model weights and publishing research landed as a direct challenge to OpenAI’s current posture on open source. Altman said OpenAI is “on the wrong side of history” and should pursue a different open-source strategy, while also noting that not everyone at the company shares that view and that open sourcing isn’t the top priority right now. The remark matters because it reframes a long-running tension in AI: whether openness accelerates progress and builds trust—or whether it undermines a company’s ability to recoup massive development costs.

The discussion then pivots to the central risk question: does open sourcing parts of OpenAI’s models endanger the business? A common argument in favor of staying closed is cost and competitive leverage. Training and deploying frontier models is described as extraordinarily expensive, and giving away capabilities could let competitors build similar products without paying the same R&D bill—potentially weakening revenue at a time when ChatGPT’s burn rate is portrayed as high.

But the counterpoint emphasizes that openness can create value in ways that closed strategies often miss. One claim is that recent open releases—highlighted through the example of DeepSeek—demonstrate that “optimizations” can be “left on the table,” suggesting that open ecosystems can iterate faster and squeeze more performance out of existing model families. The transcript also argues that open source draws in talent that might not otherwise join the conversation, turning public artifacts into a proving ground for new ideas and implementations.

Another practical argument is that open sourcing doesn’t necessarily cannibalize the latest product. Even if weights are released, the newest and best results still come from using OpenAI’s hosted offerings. Meanwhile, running large models locally is framed as difficult for most people: it requires expensive GPUs, reliable uptime, and technical know-how. The transcript contrasts that reality with everyday users who just want tools—like email writing assistance—rather than self-hosting infrastructure. In that view, the average customer base is unlikely to disappear.

The only meaningful danger identified is that another organization could take released models and build something better, effectively leapfrogging OpenAI. Still, the transcript suggests a mitigation: don’t release the absolute latest model weights, and keep the newest capability tied to OpenAI’s platform.

Overall, the takeaway is a push for openness as a goodwill and ecosystem strategy—one that could increase adoption, attract contributors, and improve public perception without necessarily harming revenue. The transcript ends with a prediction that “o1 mini” will be open sourced within about a month, underscoring the belief that the company’s stance may shift soon.