要点
- Task map rooted in American Time Use Survey + Multinational Time Use Study, plus a demographically representative survey of 3,995 respondents on where people want assistant help.
- Domains: caregiving, health, education, household admin, personal finance, shopping, government services.
- Example model scores on demo workflows: kindergarten enrollment 93.1%, flood volunteer coordination 98.0%, home search 100%, remote game watch 96.3%, diabetes meal prep 90.4%.
- Framed as tracking browser use, computer use, and connectors over the past year toward an agentic personal assistant.
- No public task list, grader, or third-party baseline appears in the post.
OpenAI’s LifeVal post argues that AGI progress should be measured on the chores people actually want help with, not only academic knowledge tests. That framing matters. The evidence published with it is thinner than the ceiling scores suggest.
[1]What shipped
OpenAI says it built LifeVal to track and drive model improvement on everyday tasks over the past year. The task catalog starts from the American Time Use Survey and the Multinational Time Use Study, then adds a demographically representative survey of 3,995 respondents on where people most want a personal assistant. Resulting areas include caregiving, health, education, household administration, personal finances, shopping, and accessing government services.
The post shows example conversations and scores for current models: enroll in the right kindergarten after a move (93.1%), manage volunteers ahead of a flood (98.0%), find a family home with budget and commute constraints (100%), set up remote game-watching with a grandparent (96.3%), and meal prep after a new diabetes diagnosis (90.4%). The narrative links these lifts to a year of work on browser use, computer use, and connectors inside ChatGPT, moving the product from chat toward an agent that takes actions.
Mission language is explicit: for more than one billion weekly ChatGPT users, help should mean getting things done, not only advice. LifeVal is presented as a step toward measuring that.
[1]Evidence on the table
Hard, checkable pieces in the post are the survey size (3,995), the named time-use source surveys, the listed life domains, and the five demo scores. Method detail for sampling, task construction, grading, tool access, and failure taxonomy is not published here. There is no linked public leaderboard, task suite, or independent lab result. All performance numbers are OpenAI-run.
[1]Steelman and gaps
Steelman. Academic benchmarks miss the distribution of daily time: childcare logistics, government forms, health routines, money admin. Anchoring an eval in time-use research and a representative poll is a better product compass than another exam set. If models can operate browsers and connectors safely on those workflows, that is closer to “helpful assistant” than a higher MMLU point.
Gaps: scores sit near a ceiling on selected demos, which weakens discrimination and invites cherry-picking concerns. Without public tasks, graders, and tool stubs, outsiders cannot falsify the claim or compare models. The post does not report error modes, human baselines, cost, latency, or how often the agent must hand control back. “100%” on a home-search demo is a product screenshot class of evidence until the suite is open. Conflict of interest is structural: the same lab ships the assistant and the scorecard.
[1]Who should care, and what to watch
Eval and safety teams should separate “everyday task taxonomy” (useful) from “vendor KPI with closed items” (weak science). Regulators and enterprise buyers asking for agent readiness should demand public tasks or third-party audits before treating LifeVal-like numbers as procurement truth. Rival labs can adopt the time-use framing without accepting OpenAI’s ceilings.
Watch over six months: whether OpenAI releases LifeVal tasks or a redacted public subset; whether independent groups publish everyday-agent suites with human baselines; whether ChatGPT product notes cite LifeVal when shipping browser/connector changes.
[1]Close
LifeVal’s durable idea is the taxonomy: measure assistants on the hours people actually live. The durable risk is treating closed, near-ceiling demos as proof that everyday AGI has arrived. Six months on, either the suite opens enough for outsiders to disagree with the scores, or LifeVal remains a compass for OpenAI’s roadmap rather than a benchmark the field can share.
[1]