Boris Power says OpenAI research mostly targets later GPT generations

At Fellows Forum on September 23, Boris Power, OpenAI’s head of applied research, estimated that 80–90% of research at the company focuses on “GPT-7, GPT-8 and beyond.” He offered the figure while explaining how OpenAI balances work on current products with research into more capable models. It is his estimate in a public conversation; OpenAI has not published a research allocation breakdown to check it against.

The big change

• What changed: Power described two connected tracks: improvements to models people use now, and research aimed at later generations. He estimated that the latter accounts for most OpenAI research.
• Why it matters: A stronger large model can supply training examples for a smaller, cheaper model on a defined task. That gives developers a reason to follow frontier research even when they rely on smaller models in their products.
• What to watch: Future model releases and technical disclosures can show whether advances in large models reach smaller ones. The interview supplied no GPT-7 or GPT-8 release dates, specifications or results.

The context behind the estimate

Power was answering a question from Zachary Lipton about how customer requests feed back into research. He said OpenAI can improve a particular behavior, such as tool use, with specialized training data and applied research. In his account, incremental updates within a model generation often address such needs. A new generation of larger models, he argued, can change which specialized fixes are still necessary. He described some current investments as ways to learn and improve faster today, even when they are not the long-term strategy.

That distinction led to the 80–90% remark at 1:32:52 in the Fellows Forum recording. Power referred to research at OpenAI, not just research in his own team. Neither the recording nor OpenAI’s published material provides a definition of the research being counted, a time period, a budget, a staff count or a method for arriving at the estimate. The number should therefore be read as Power’s description of priorities, not an audited measure of company spending or staffing.

Power named GPT-7 and GPT-8 as examples of later generations. He did not announce either model, disclose a development milestone or give a release schedule. OpenAI’s current model catalog lists GPT-6 Astra, Sol and Luna; it provides public context for the names Power used, but no evidence about the state of the later generations.

Why he brought up distillation

Immediately after the estimate, Power described a large GPT-6 Astra model teaching a smaller GPT Luna model. His point was that a strong large model can provide examples for training a smaller one. He was explaining a research approach, not documenting that a particular released Luna model was trained from Astra.

OpenAI’s supervised fine-tuning guide describes the mechanism. A developer first checks that a larger model performs well on a defined task, keeps suitable outputs, uses them as training examples for a smaller model, and evaluates the smaller model on that task. OpenAI says this can bring the smaller model closer to the larger model’s performance on a specific task. It does not imply that the smaller model inherits every ability of the teacher. OpenAI’s model distillation announcement likewise describes evaluation, captured outputs and fine-tuning as an iterative process.

The idea has a longer history. In a 2015 research paper, Geoffrey Hinton, Oriol Vinyals and Jeff Dean studied transferring knowledge from larger systems into a model easier to deploy. That work explains the teacher–student concept; it says nothing about OpenAI’s current allocation of research.

This explains the link Power was drawing between frontier research and practical models. A more capable teacher can create useful training material for focused, lower-cost systems. Whether that works well depends on the task, the examples selected and the evaluation. Power supplied no distillation results or comparative measurements for the Astra–Luna example.

Power’s estimate is revealing because it states his view of where OpenAI expects the greatest research value. The public evidence establishes the remark and the general technique he invoked. It does not establish how OpenAI’s resources are actually divided or what a future GPT generation will deliver.