X：Rohan Paul

Astra was probably our biggest competitive advantage while it wasn’t generally available. Since we’ve had it our productivity jumped so much that we shifted some of our plans 6 months ahead and will ship them at DevDay instead of mid next year.

An AI can complete a task perfectly and still spend your time on the wrong thing. Imagine an assistant preparing notes for tomorrow’s meeting. A message arrives that changes the deadline, and suddenly another project needs your attention first. The notes may be excellent. The assistant still needs to recognize that your priorities have changed. As AI takes on more work, these decisions become increasingly important. Which task should come first? When is an interruption justified? When should the system ask you? That’s the question behind @TodayAIofficial’s approach to personal AI: how can an agent learn what matters to a particular person and use that context throughout the day? Memory is part of the answer. Remembering a deadline helps. Connecting it to a promise you made last week, noticing that the plan has changed, and bringing it up while you can still act takes more. It also requires a way to correct the system. People change their minds. A preference from three months ago may no longer apply. An assistant should make its assumptions visible and let you update them. I think this is a useful direction for personal AI. There’s considerable value in software that can connect scattered information and help you decide where to focus. The test will be how well it handles an ordinary, messy day: catching the commitment you might miss, explaining why it needs attention, and leaving the decision with you.

I think "time horizon" is becoming one of the more useful ways to talk about agent capability. At OpenAI, as tasks get longer, success without human intervention collapses, from 86% on sub-15-minute tasks to only around 16% on the longest 64-128-hour bucket, while successful runs requiring intervention become much more common. A model can be extremely capable locally and still be unreliable over a long execution trajectory. Not tokens. Not benchmark score. How long can the system keep useful control of a task before a human has to intervene? maybe we are moving towards a benchmark, something like human minutes consumed per completed task.