OpenAI just announced three major breakthroughs in mathematics, computer science, and physics. On the surface, this sounds fantastic. But beneath the headlines, there’s a growing frustration from researchers about how the company is choosing to share these findings, and I think there are important lessons here for how we build and deploy AI systems at scale.
The core issue is gatekeeping. OpenAI announced it would work with hand-picked expert “shepherds” to control how these results get released to the broader community. For the physics result especially, the company hasn’t released anything publicly yet, instead deferring to external experts about if and how the materials should be shared.
On its face, consulting experts sounds reasonable. But the research community isn’t buying it. The Advisory Group on Mathematics and Artificial Intelligence explicitly warned against this exact practice just a day before the announcement, calling it “unnecessary” and pointing out that it creates unequal access, risks biased selection, and can delay important scientific work.
The Real Problem with Selective Access
I understand why OpenAI might think controlled release makes sense. These are complex results. They need context. They need explanation. But what the company is doing is essentially saying: “We solved something important, but only a few people get to see it first.”
That’s the opposite of how science works best.
Andreas Karch, a physics professor at UT Austin, put it bluntly: “The fact that they handpick the group is a disaster. The large labs have enough power as it is. They should not be picking winners and losers.” He’s right. When you concentrate gatekeeping power in a few hands, you’re not enabling better understanding, you’re creating bottlenecks. You’re deciding whose voice matters in the conversation.
For developers and AI practitioners watching this play out, this should feel familiar if uncomfortable. We’ve seen how platform gatekeeping works in practice across tech. It doesn’t scale well. It breeds resentment. And it ultimately slows down progress.
Why Openness Matters for Adoption
There’s a practical angle here that I find particularly interesting. One researcher noted that having multiple groups around the world working on understanding these results actually supports faster adoption of AI tools. When you restrict access, you reduce the number of people who can contribute, experiment, and build on what you’ve created.
If OpenAI genuinely wants the AI research community to move faster and more broadly embrace new techniques, keeping results behind expert gatekeepers seems counterproductive. It’s like releasing a powerful API but only letting a few people use it.
I’ve been thinking about this in the context of our broader conversation around ai-research collaboration. The most impactful breakthroughs in computer science have typically come when knowledge flows freely, when junior researchers and established academics alike can access, critique, and build on new work. Restricting that flow doesn’t make the science more rigorous, it just makes it slower and potentially less democratic.
The Deeper Issue: Speed Without Scrutiny
But here’s what really troubles me about this situation: the gatekeeping might be a symptom of a bigger problem, not the root cause.
Xiaoyu He, a math professor at Georgia Tech, identified what I think is the core tension: “Fundamental results are being solved so quickly without human contribution, and the papers are difficult to read and verify.” That’s the real issue. OpenAI can’t solve this by consulting with shepherds or releasing in phases. The problem is that AI systems are solving important problems faster than the human community can understand and verify them.
That’s genuinely concerning for the field. And it’s concerning for anyone building with these systems. When breakthroughs happen so fast that even experts struggle to follow, we lose something important: collective confidence in the work.
I keep coming back to our work on open-science practices, because I think this is where real accountability starts. It’s not about removing guardrails or moving faster. It’s about creating structures where the community can actually participate in validation and understanding.
What This Means for Builders
As developers working with AI systems, we should be paying attention to how these institutions handle transparency. The practices they normalize become the industry standard. If OpenAI successfully establishes gatekeeping as acceptable for major breakthroughs, that sets a precedent that affects how we’ll access information, tools, and understanding as AI becomes more central to what we build.
The question isn’t really about three specific math problems. It’s about whether we’re building an AI ecosystem where knowledge concentrates or disperses, where progress accelerates or stalls, where everyone has a voice or just the people picked by the labs that built the systems.
What kind of AI future do we want to build?