要点
- OpenAI (2 Nov 2026) introduced LifeVal to measure and drive progress on “tasks that matter in everyday life,” stressing that a personal assistant should take actions—not only give advice.
- Task discovery used the American Time Use Survey, the Multinational Time Use Study, and a demographically representative survey of 3,995 respondents; domains include caregiving, health, education, household admin, personal finances, shopping, and government services.
- Example current-model scores cited: kindergarten enrollment 93.1%, flood volunteer coordination 98.0%, home search 100%, remote game-watching setup 96.3%, diabetes meal/care week plan 90.4%; OpenAI says the past year improved browser use, computer use, and connectors.
LifeVal shifts the eval story from academic knowledge boards to everyday workflows users say they want help with—childcare, government services, groceries, budgets—and frames the goal as an agentic personal assistant that gets things done.
Falsifiable read: if third parties cannot reproduce the ~90%+ example scores under comparable task definitions, or if highs mainly reflect narrow scripts/human assist rather than end-to-end autonomous completion, treat 2 Nov 2026 mainly as product direction and training priority—not delivered general everyday-task SOTA; if independent reruns and failure disclosure hold, raise the weight of “daily-task evals driving model improvement.”
[1]Official facts: dated 2 November 2026; title LifeVal; subtitle on measuring progress on everyday tasks. OpenAI cites more than one billion weekly ChatGPT users and lists childcare, government services, budgeting, school prep, recipes, groceries, and workouts.
Method sketch: start from ATUS and MTUS, then a 3,995-person survey on where people most want personal-assistant help. The post shows several “Open conversation” workflows with percentage scores and says latest models score very highly on many LifeVal tasks; over the past year it trained for faster, more capable, more aligned browser use, computer use, and connectors.
[1]Evidence is an OpenAI official post with survey N and example scores, but without a full task card, grading rubric, human baseline, confidence intervals, or third-party reproduction bundle. Scores such as 100% on home finding especially need a definition of success (shortlist? cart? compliant completion?).
Steelman: anchoring tasks in time-use research and a representative “want help” survey is closer to consumer-assistant products than academic boards alone; pairing that with browser/computer use and connectors matches an “can act” story. Beyond the steelman, self-graded, cherry-picked highs bias toward success paths and underweight long-tail failures, permissions, legal liability, and bad data.
[1]Strongest steelman: if everyday-task evals truly close the loop into training, ChatGPT can move from “can chat” toward “can finish a job,” which ordinary users feel more than another knowledge bench.
Steelman risk: example scores cluster at 90–100 without comparable-model controls or failure modes; compliance and truthfulness risks in government, medical, and housing tasks are thin; “1B+ weekly users” is a scale claim, not automatic external validity for LifeVal.
[1]For product and eval readers, LifeVal signals OpenAI elevating actionable everyday tasks as an internal north star; for policy and safety readers, expanding browser/computer use widens the blast radius of errors and misuse. Watch whether task specs go public, whether independent reruns are feasible, and whether high-scoring tasks ship with auditable tool-call traces.
[1]Near term, do not read the example percentages as a settled everyday-AGI progress bar: judge by public task specs, third-party reruns, and real-user completion/undo rates. If those three are missing, LifeVal is mainly a training and narrative dashboard; if they arrive together, it can become a reference axis for consumer-agent evaluation.
【No.2】
[1]