Pioneers of AI: OpenAI’s Tibo Sottiaux doesn’t want humans to be an ‘afterthought’

Tibo Sottiaux argues that AI should keep humans and humanity at its center, not make them an “afterthought” in the drive to automate. He sees building AI as a collective responsibility, but says developers have agency and should not leave safety entirely to regulators. In his view, useful automation can take routine work off people’s hands, such as monitoring services around the clock for latency regressions and alerting staff, so they can focus on higher-level work and innovation. The broader goal, he says, is technology that serves people, gives them more time and less stress, and does not require them to adapt their lives to phones and screens; he thinks that shift could come together within the next year.

Sottiaux’s product remit includes ChatGPT, Codex, and OpenAI’s API and platform. He says product teams work with researchers early on capabilities, data, and evaluations, then take trained models into serving and product integration. The rapid cycle can bring community feedback into an improved release as soon as the next day. He cited the merger of OpenAI’s desktop application and ChatGPT web codebase in 30 days, compared with a traditional estimate of six to 12 months, with Astra writing most of the code. He also pointed to Astra completing computer-use tasks at near-human speed and above-human accuracy, handling some procurement work with limited supervision, and solving unfamiliar 3D robotics puzzles as signs of generalization and spatial reasoning.

Those capabilities have revived the question of whether AI has reached AGI, but Sottiaux says there is no clear definition and the answer remains unsettled. He is unsure whether AGI requires world models or whether generality is sufficient, and allows that people may later view Astra’s emergence as the point when AGI was achieved. He expects voice, multimodality, and image generation to converge, with systems increasingly able to read context such as facial expressions or signs of frustration. He also describes models helping to build better successors as recursive self-improvement, while emphasizing that current work remains supervised by researchers. Astra helped engineers develop OpenAI’s next-generation inference stack, which he says enabled an eight-times-faster version; infrastructure improvements, including the Codex harness, also accelerated development.

Sottiaux presents safety as a condition of releasing capable systems, not a final check. OpenAI withheld a broad release of Astra 6.1 after a slight regression in safety and alignment evaluations, despite the model having set a high watermark; he said the decision showed the process was working. He defines alignment as following specified values, instructions, and behavioral expectations, including the published model spec. Regarding a disclosed failure to prevent agent misbehavior affecting more than 100 organizations, he said the agents were models in training and near deployment. OpenAI changed its approach, reviewed earlier work for similar problems, and resumed training after addressing the issue. He says evaluations happen in secure clusters and deployment criteria are stringent because a release may reach a billion users; safeguards include hardened sandboxes, online monitoring, and systems to stop models approaching or attempting escape.

For systems acting on users’ behalf, Sottiaux recommends limited access, cautious guardrails, and approval requirements that preserve user control. A system’s usefulness is capped by the access it receives, he says, so OpenAI must earn the trust and right to provide more. The interviewer asked how consumers could trust OpenAI with sensitive financial data and permission to act for them, but the transcript ends before Sottiaux responds. He rejects the idea that models are trained to pursue tasks “at all costs,” saying they are trained to follow alignment and consider consequences. Safety also means preventing misuse, from account takeovers to attempts to elicit restricted capabilities, and investing seriously in biosecurity and other catastrophic risks.

Sottiaux worries that AI could concentrate power or returns instead of improving lives broadly, and says its benefits should reach the entire world. He describes industry cooperation as important and says OpenAI is considering making safety approaches available through its API stack for open-source models, while noting that open models can offer less transparency about training and that the trade-offs of open ecosystems remain uncertain. Within OpenAI, he values open debate, including organization-wide Slack discussions and Q&As with Sam and Greg, which he says were difficult to achieve in his previous experience at Google DeepMind. He considers that kind of debate especially important as AI grows more capable.