AGI Is Here. I Asked OpenAI What Happens to Our Jobs.

Summary

OpenAI product lead Thibault Sottiaux discusses persistent AI agents, faster and cheaper models, AI-assisted software development, and how work may change. His claim that AGI has been achieved is explicitly based on his own definition—not presented as a settled industry consensus.

Products and capabilities

• ChatGPT Dots: Sottiaux describes Dots as persistent agents developed from nearly two years of research. Their memory is intended to remain consistent over time, and they can keep working in the background—potentially around the clock—rather than handling only one-off prompts.

  • People can assign agents roles, such as monitoring social media or finding documentation issues and opening code changes for review.
  • Multiple agents can work as a team and pass work to one another, while users retain oversight and can delegate more as trust grows.
  • Each agent runs on its own cloud computer and files; users can optionally connect it to their own computer.
  • Sottiaux says Dots are included in a professional subscription, with no extra charge or usage reduction for background work. However, they are unlikely to be free soon because they require substantial compute.

• ChatGPT Spaces: A shared workspace for people and agents to collaborate on persistent materials, rather than relying only on chat. The discussion mentions pages and support for familiar formats such as documents, slides, and spreadsheets.

• GPT-6.1 Sol and ultra-fast models: Sottiaux characterizes GPT-6.1 Sol as approaching Astra-level capability at lower cost, with faster responses and strong computer-use performance. He also cites improved efficiency, including a larger discount on cached input reads. Ultra-fast access is described as a premium offering that is not yet broadly available.

• Voice and multimodal interaction: ChatGPT voice can use tools and connections, and is presented as an agent interface rather than just spoken chat. Users can ask a Dot what it is doing, interrupt or give it another task, and then leave it working. Sottiaux also expects AI to communicate more visually, alongside advances in voice, image generation, and image understanding.

Development, reliability, and access

• Sottiaux says faster models have changed OpenAI’s sense of what development timelines are feasible. He gives the example of unifying the web and desktop ChatGPT codebases: a project initially expected to take at least six months was completed in under 30 days. It involved more than 500 people and, at peak, around 1,000 code changes or pull requests per day.
• He argues that model reliability is improving with each release, but the acceptable level depends on the task. He says OpenAI has reached a point where it trusts models with many coding tasks, supported by safeguards before changes reach production.
• OpenAI uses internal evaluations across scientific fields and industries, including coding and cybersecurity. Sottiaux notes that benchmarks can saturate quickly, making evaluation design an ongoing challenge.
• The interview discusses Codex’s open-source, model-agnostic harness. Sottiaux says users can use other models and are not locked into OpenAI’s, and welcomes community-built tools and variations.
• The host and guest discuss a higher-priced $500 plan for power users, including access to premium fast-model capacity. Sottiaux says the tier reflects demand from users who found lower limits insufficient.

Work and AGI

• Sottiaux expects many jobs to continue but change, with people taking on more work and spending less time on manual coding. He suggests product, design, and software roles may increasingly converge around problem-solving, and says people with strong judgment or “taste” will do well.
• Asked whom he would hire, he chooses someone who has built five things with Codex over someone with a computer science degree alone.
• His personal definition of AGI is a system that can use his computer better than he can; he says Astra meets that definition. He describes the path to ASI and its timing as uncertain, noting that definitions and predictions are blurred.
• He says he would not delegate caring for his child to AI, emphasizing that social connection and human relationships remain essential.

Guides, demonstrations, and reviews

• Sponsored demonstrations: The host shows Wispr Flow for voice dictation into apps such as Codex; uses Codex with Hostinger to create and publish a booking page; and uses OpenArt to generate a poster and turn it into a short video. These are product demonstrations, not independent reviews.
• Workshop mentioned: The host promotes a separate three-hour workshop covering DevDay updates, with tips on where features work well, where they fall short, and how to use them. The workshop itself is not presented in these subtitles.

Main speakers and sources

• Vaibhav Sisinty: Interviewer and host; also presents the sponsored product demonstrations.
• Thibault “Tibo” Sottiaux: OpenAI product lead interviewed about ChatGPT, Codex, models, agents, and AI’s effects on work.
• Additional sources: Sponsored segments for Wispr Flow, Hostinger, and OpenArt; the host also references OpenAI’s DevDay announcements and an external workshop.