OpenAI Teaches AGI to Find Childcare First, While Users Ask About Privacy and the Definition of AGI First

On November 2, OpenAI released LifeVal.

It is a benchmark, not a new model. OpenAI says that if AGI is to benefit everyone, it should help people handle their most important everyday tasks, so it built this test suite from feedback from thousands of ChatGPT users, covering scenarios such as finding childcare, accessing government services, and managing personal finances. Over the past year, OpenAI has used LifeVal to measure and improve models' ability to handle these tasks.

That sounds gentle, but it actually drags AGI's evaluation criteria toward life's mundane chores. Writing papers and solving math problems are certainly hard, but whether finding childcare means confirming a booking or just providing a shortlist, and whether accessing government services means filling out a form or getting through on the phone—these boundary-blurred tasks are harder for a benchmark to articulate clearly. One commenter wrote directly: if a model can get through to HMRC's phone line, that's AGI. Compared with solving problems, a phone queue system is more like a real-world Turing test.

The comments section isn't as tidy as a launch event.

Some users worry about privacy: finding childcare and using government services mean it sees everything about you, and they want that capability to run on their own machine—otherwise they'd rather not have it. Others pressed for where Terra ranks on LifeVal, and some asked why the page doesn't show 6.1 Sol and Terra. Another specific question: does passing the childcare item mean confirming a booking, or just producing a shortlist?

Someone posted a three-panel meme: a student uses ChatGPT to fudge homework, the person setting the assignment is also ChatGPT, and finally there's the teacher. In the image, the character sits in a car, expression shifting from surprised to gloomy. This meme fits well under LifeVal. What makes everyday tasks troublesome is acceptance: who gets to decide, and by what standard is it considered passed.

There was positive feedback too. One user said SOL helped them get a library card last month and even found free library audiobooks from their Amazon wishlist. They judged that sign-ups for various government programs might rise quickly. Another said LifeVal starts from real time-use surveys and includes international samples—a good starting point.

Criticism was equally direct. Some stopped reading as soon as they saw 'AGI,' believing OpenAI is far from AGI. Some asked why, when current models have already made progress in math, physics, engineering, and biology, there's such a rush for the next leap. Others saw LifeVal as 'if you can't beat them, build a new benchmark.' One more heated long reply criticized OpenAI for being closed-source, citing the shutdown of the 4-series model APIs and more than 24,000 people signing a petition to keep 4o. These voices don't necessarily represent the majority, but they show that AGI slogans amplify trust issues.

LifeVal's real value may lie in breaking model capabilities down into small, verifiable tasks. The three letters 'AGI' instead easily steer the discussion off track. Finding childcare, filling out government forms, handling finances—each can force out concrete questions: Was the task completed? Where did it fail? Can it be repeated? But the acceptance criteria must withstand follow-up questions. The privacy, model scores, and definition of passing that users ask about are all questions benchmark design cannot avoid.

As for whether it can get through to HMRC by phone, that can wait until someone actually tests it.