"Two years of research成果": OpenAI co-founder discusses Spud model for the first time, responds to Dario's "life-or-death" criticism

According to 1M AI News monitoring, OpenAI co-founder Greg Brockman, in an interview, first discussed from a technical perspective a new-generation pretraining foundation model inside OpenAI codenamed Spud. He described Spud as “the culmination of about two years of research,” and as the foundation for all subsequent improvements in OpenAI’s capabilities. He explained the model production process: first produce the foundation model through pretraining, then use reinforcement learning so the model can practice solving problems across a variety of scenarios, and finally carry out additional post-training fine-tuning focused on behavior and usability. In the past 18 months, his main efforts have been focused on expanding GPU infrastructure and the training framework.

On Spud’s capabilities, Brockman offered qualitative rather than quantitative expectations: the model will be able to solve harder problems, understand instructions more precisely, and grasp context more deeply—reducing that “big model smell.” He gave an example from an engineer with whom he has worked closely: between GPT-5.2 and GPT-5.3, this engineer went from “completely unable to use AI for underlying systems engineering” to “give it a design document and it can implement the functionality, add monitoring metrics and observability, run a performance analyzer and optimize again—producing exactly what he wants.”

The interviewer cited earlier remarks from Anthropic CEO Dario Amodei: “Some players are gambling with their lives (yoloing), pushing the risks too high—I’m very concerned,” and also mentioned Amodei’s claim that if forecasts were even slightly off, the company could go bankrupt. Brockman responded directly: “I disagree. We’ve been thinking very carefully the entire time, and we’re very clear about the trends that are coming.” He believes OpenAI was the earliest company to recognize the compute bottleneck, while other companies “probably only realized this at the end of last year—starting to rush and buy up compute, but it was already gone.” Regarding the bankruptcy risk, he thinks there are “actually more exits on the ramp,” and he boils the bet down to a judgment about the entire industry rather than a single company: “Do you believe this technology can produce and deliver the kind of enormous value we’re seeing?”