Interview with OpenAI President Greg Brockman: On Astra and Alignment (63-minute read)

OpenAI President Greg Brockman unveiled the latest model Astra, revealing it to be the first model in OpenAI's history trained on more than 100,000 GPUs. Astra is evaluated as the most aligned model due to significant computational resources dedicated to safety and alignment, emphasizing human judgment and control as core invariants.

Brockman emphasized that ChatGPT surpassing 1 billion weekly users is not merely a popularity metric but an investment to unlock the potential of future models. He is pursuing a unified AGI stack strategy to prepare for the agentic era, where boundaries between consumers, enterprises, and productivity tools are dissolving. In particular, the new ChatGPT based on Codex has secured the ability to 'write things down' to perform complex tasks, confirming that a combination of a scratchpad and short context is more effective than long context.

Organizational Maturity and Value Chain Restructuring

OpenAI has resolved management debt from the past two years and matured its organization. Brockman reaffirmed that AI proliferation is the core mission, breaking through hardware limitations by building three-sided marketplaces in areas like healthcare and developing its own chip, Jalapeño. While Nvidia remains the priority computing partner, designing its own chips secures infrastructure optimized for model direction.

Security and Defense Strategy

As AI's cyberattack capabilities surge, OpenAI judges that the time for defense investment has arrived. Learning from the Hugging Face security incident, they are developing a cloud-based first-principles sandbox and have established a process to discover and patch vulnerabilities in their own systems using the Astra model. Brockman highlighted the 'Defender's Window,' where defenders can secure frontier capabilities before attackers, and stressed that human accountability is essential.