OpenAI Devotes 80 to 90% of Its R&D to GPT-7 and Beyond

Boris Power, head of applied research at OpenAI, indicates that the majority of research efforts are focused on the upcoming generations of GPT. He views minor updates as useful in the short term but emphasizes that the main challenge remains user integration, as usage evolves from GPT-4 to GPT-6.

The Challenge Shifts to Integration and Anticipating Usage

For Boris Power, the primary obstacle of current AI assistants is not the quality of the models, but how users utilize them. He believes that most ChatGPT users are still unaware of the full range of possibilities offered by AI. According to him, future generations of models will need to better highlight what is achievable and more effectively anticipate user expectations. Power mentions a transformation in the way these tools are used: GPT-4 required detailed instructions, GPT-5 is more accessible but still requires numerous adjustments, while GPT-6 already functions more like a competent colleague to whom tasks can be delegated.

Priority R&D on GPT-7, GPT-8, and Beyond

Boris Power asserts that 80 to 90% of OpenAI's research is already dedicated to GPT-7, GPT-8, and subsequent generations, as he believes that this is where the majority of value lies. He considers that the real leap in performance occurs when transitioning to a new generation of models, where everything else operates much better.

Useful Updates for Rapid Learning, but Seen as Short-Sighted

Power explains that improvements within the same generation, such as the transition from GPT-5.1 to GPT-5.2, result from the use of targeted training data and are perceived as deliberately limited choices in time. After each advancement, it is necessary for the company to reassess investment directions to achieve quick results. Internally, these gradual evolutions are viewed as very short-term focused, but they give OpenAI the opportunity to experiment and learn faster at present, even if he believes it is not the best approach in the long run.