OpenAI introduces LifeVal: reading task scores and evidence of completion

An evaluation needs to decide which assignments to give AI. Performance on exam questions alone cannot show how much effort it saves in finding a school after moving or coordinating a family schedule. LifeVal starts with everyday tasks where people want assistance.

OpenAI says it drew on U.S. and international time-use studies and separately surveyed 3,995 people about where a personal assistant would help. The U.S. Bureau of Labor Statistics' ATUS measures time spent on work, care, leisure and other activities. MTUS collects national time diaries into a common format. American Time Use Survey · Multinational Time Use Study

%% How tasks were selected · OpenAI's account of the evaluation design
flowchart TD
  A["Time-use studies<br/>Activities in everyday life"] --> B["Separate survey of 3,995 people<br/>Where would help matter?"]
  B --> C["Build LifeVal's everyday tasks"]
  C --> D["Evaluate model work<br/>Track performance over time"]

The published kindergarten example asks for a comparison and shortlist near a new home. Its displayed response organizes three candidates and enrollment information, while leaving two preferences for the user to decide. It shows a shortlist ready for review; that response alone does not establish an actual application or confirmed enrollment.

The table separates the public example's response from evidence to examine if the user later proceeds with a real application. It is not guidance on a particular region's enrollment rules.

The announcement's model chart labels its metric “Average reward (unweighted).” Unweighted means the average gives no extra weight to selected scores. Reward here means an evaluation score, not money paid to a user. The expanded task table instead identifies its figures as the best recorded GPT-6 Astra scores, so it needs to be read separately from that average chart.

The first two averages differ by 80.6 − 80.0 = 0.6 percentage points. That is not a claim of 0.6% more success on every real-life task. The inspected announcement and expanded task table do not provide attempt counts, detailed grading rules or uncertainty intervals, so they do not establish the statistical meaning of that small difference.

OpenAI describes progress in browser use, computer use and work through connected apps. A connected app provides a route to another service's information or functions. A user needs to examine the app's supported features, data access and authorized scope as well as the model name.

These are general examples of checks for delegated everyday work. An evaluation score does not guarantee the features of a particular account or its real outcome.

The earlier HackerTube process explainer also examined evidence of completed development work. Everyday assignments likewise need a matching output: a comparison, saved schedule or submission record.