OpenAI ships new capability and risk every Tuesday, its former safety-report lead says

When Robinson started at OpenAI in May 2023, a day after Sam Altman's first Senate testimony, shipping a frontier model generally meant training one from scratch ^[1]. "We were going to bake a fresh cake with a new pretraining run, do the whole thing from scratch," Robinson told Ezra Klein on "The Ezra Klein Show" ^[2]. That process took months and held major releases to a few a year ^[3]. Post-training and safety testing came only after each new base was finished ^[3].

His description of the current stack separates the layers by cost. "You've got the pretraining; that's the baking of the underlying model," Robinson said ^[8]. He said "But then, in addition to post-training, you have reasoning training" and that "those steps are easier to do quickly, so you can redo them" ^[9]. This is sound engineering: if the bottom layer takes months, you design the upper layers so they can be replaced without touching it. A better reasoning recipe can be layered onto the base model OpenAI already has ^[6].

The second layer is tooling. "It's not just a chat anymore," Robinson said of the tools and affordances now wired into models ^[11]. Those can make a system more capable and change how it behaves with no new training run beneath it ^[11]. "All of those things are changing what the model can do and what the risks are, and we're shipping new capability and risk every Tuesday," he said ^[12].

"Every Tuesday" is Robinson's characterization of the pace. OpenAI once made a point of the "month or months" of safety work that went into GPT-4 after the model was finished ^[7]. That window is longer than the whole gap between GPT-6 Sol and GPT-6.1 Sol ^[18]. It does not prove GPT-6.1 Sol got less testing, because testing for one release can run while the previous one ships.

For a team building on these models, the base model is now only one of the layers that can change between major releases ^[6]. The interview does not say which of those changes reach a model a developer calls through the API. In my context, where an application's behavior is the product, I would treat each OpenAI release as a dependency upgrade and run our own evaluations against it before users see it.

The third accelerator is OpenAI's own research speed. Its September report said that by mid-August the organization was running 3.1 eight-hour agent workdays for every human workday ^[15]. That is about 24.8 agent-hours for each human workday ^[19]. The least glamorous evidence is the easiest to believe. Robinson said the research infrastructure can be "pretty janky" ^[16]. Researchers who once took a broken experiment to an internal Slack channel now ask Codex, and OpenAI's report shows traffic to that channel falling as agent use climbed ^[16].

Competition adds pressure to use that speed. Klein pointed to a more crowded field than OpenAI faced a few years ago, with Anthropic, xAI, Chinese labs and open-weight models all competing for enterprise customers ^[17]. "All of these things push toward speed," Klein said ^[17].