According to Fast Company, last year OpenAI tasked post-training research lead Michelle Pokrass with building a team focused on the company’s power users. These users include doctors, scientists, coders, and companies building on OpenAI’s API. Pokrass, a former Coinbase and Clubhouse engineer who joined in 2022, played key roles in developing GPT-4.1 and GPT-5. Her team’s work on “post-training” refines how models understand user intent, enabling more polished outputs. She cites examples of GPT-5 aiding scientific breakthroughs and discovering new mathematical proofs. The core belief is that today’s power-user applications become the median-user norm within a year or two.
The Power User Feedback Loop
Here’s the thing: this is a brilliant strategic pivot. For a long time, the narrative around LLMs was about making them accessible to everyone—the “democratization of AI.” But OpenAI seems to be recognizing that democratization only gets you so far. To actually advance the underlying technology, you need people who are stress-testing it in extreme, novel ways. A regular user might ask ChatGPT to write a blog post. A power user is trying to get it to debug a complex quantum computing simulation or propose a novel protein fold. The feedback from that second group is infinitely more valuable for making the model smarter, not just more user-friendly.
Building for Tomorrow, Not Today
Pokrass’s quote is the real kicker: “build to where capabilities are emerging, rather than just focusing on what people are using the models for now.” That’s a fundamentally different philosophy. It’s not about optimizing for the most common queries in the logs. It’s about finding the faint signals of what’s *possible* and pouring gasoline on them. Think of it like this: if you only listen to your average driver, you’ll make a better sedan. But if you listen to your Formula 1 engineers and drivers, you’ll discover materials and aerodynamics that eventually trickle down to make every car better. OpenAI is choosing to listen to its Formula 1 team.
business-imperative”>The Business Imperative
So why does this matter for OpenAI’s bottom line? Two words: enterprise revenue and model superiority. The power users they’re courting—those doctors, scientists, and dev teams—are the exact decision-makers for large-scale API adoption and expensive “ChatGPT Enterprise” plans. By tailoring the model to their cutting-edge needs, they lock in the most valuable customers. And in the brutal AI arms race, having a model that can legitimately aid in R&D is a killer feature. It’s not just about writing better marketing copy than Gemini or Claude; it’s about being an indispensable tool for discovery. That’s a much harder moat to cross.
A Tricky Balancing Act
But this focus isn’t without risk. The big question is: can you hyper-serve power users without leaving the regular folks behind? There’s a danger of the model becoming *too* specialized, or its default behavior drifting away from the simple tasks that brought in the hundreds of millions of users in the first place. The post-training process Pokrass oversees is basically the calibration between raw capability and usable intuition. It’s a delicate balance. Get it right, and you have a product that feels both magically capable and broadly useful. Get it wrong, and you could end up with a brilliant but alienating tool. Based on the push into scientific and coding realms, it seems OpenAI is betting that what wows the experts will eventually wow us all.
