LifeVal

Most AI benchmarks test what a model knows. LifeVal tries to measure whether a model can get ordinary chores done: finding childcare, dealing with government services, budgeting, shopping, planning meals. OpenAI says it used this evaluation over the past year to track and push model improvement on everyday tasks.

That tells us where ChatGPT is headed. The company says plainly it wants ChatGPT to take actions, not just give advice, and it ties that to training models on browser use, computer use and connectors. A benchmark is often a roadmap in disguise, and this one points at an agent that acts across your household paperwork.

We think grounding evaluation in how people actually spend their time is the right instinct. OpenAI started from the American Time Use Survey and the Multinational Time Use Study, then ran its own demographically representative survey of 3,995 people to find where they wanted help. That's a more defensible starting point than a list of puzzles researchers find interesting.

The examples are appealing: shortlisting kindergartens after a move, organizing flood-prep volunteers, building a first-week meal plan for a parent newly diagnosed with diabetes. If AI can reliably handle that kind of friction, it's a genuine benefit, especially for people short on time or navigating bureaucracy for the first time.

But look at what's published. We get task scores, some above 90%, one at 100%, without a clear explanation of how those are graded, what counts as success, or how the overall suite is scored. We don't see a public task set others can run. An evaluation the developer designs, grades and reports on its own models is useful internally, but it isn't evidence we can independently weigh.

The stakes rise with the subject matter. A grocery order that's a little off is an annoyance. A diabetes meal plan, an enrollment deadline or a property deed added to a cart is not. High scores on curated examples say little about the failure cases that matter most, and we'd want to see those reported as prominently as the wins.

If this becomes the yardstick for personal agents, it's a better yardstick than trivia. We'd just like it to be one anyone can hold.

OpenAI frames LifeVal around its mission and its user base of more than one billion weekly ChatGPT users. The task areas it lists include caregiving, health, education, household administration, personal finances, shopping and access to government services.