I’m posting this with the sincere hope that the R1 can be saved from becoming an expensive dust bunny. The current state of reliability—or, rather, the frequent lack thereof—is fundamentally undermining the device’s utility and the entire promise of the Large Action Model.
Frankly, demanding that a user repeat themselves is a waste of breath and thought process. It’s highly non-productive and forces the very context switch we are trying to avoid.
Here is the breakdown of the critical issue, which occurs regardless of signal strength (whether I’m walking across a parking lot or near sufficient Wi-Fi):
-
The Silent Drop: The R1 will acknowledge a request, begin the process of searching or interacting with a service, and then simply silence. There is no error message, no feedback, just a void.
-
The Amnesia Trap: If I then ask the R1 what it came up with, it often replies with a generic “searching what I can come up with on the latest news,” or some similar phrase. This response confirms it has not only failed the task but has completely moved on, exhibiting total amnesia regarding the original, complex request.
-
The Productivity Killer: The task is so thoroughly dropped that it does not even retain the request within the Rabbit Hole. I am forced to start the entire request over. The only reason I continue to test is the frustrating reality that asking a second time often works, proving the initial failure was a transient bug, not a capability issue. This is why I—and likely many others—have resorted to simply opening a phone and using more reputable chatbots.
The Path to Resolution: Error-Resilience and Context Memory
The current behavior of immediately forgetting the user’s input upon hitting an error state is the most egregious flaw. A forward-thinking, user-centric device must be designed for error resilience.
The Proposal:
-
Retention on Error: If a request fails, times out, or encounters a known blockage (like a CAPTCHA, which ties into my previous post), the R1 must retain the core query in a state of active suspension.
-
Retry on Re-engagement: If the user then taps the button and asks, “What did you find?” or simply “Try again,” the Rabbit should recall the last unfulfilled request and attempt it once more, rather than defaulting to a new, generic search query.
This lack of task persistence is a fundamental flaw that is actively turning the R1 into a source of frustration instead of a source of seamless interaction. You’ve built a powerful concept; now, please give it the foundational reliability it needs to truly save us time, not waste it.
Addressing this memory loss is crucial to fulfilling the promise of the device.
For context, this issue mirrors the context-dropping failure I previously highlighted regarding web security blocks:
Proposal for CAPTCHA Handling Improvement
Thank you for your attention to this critical bug. We hold high standards for this product, and we are ready to use it—but it must meet us halfway on basic operational reliability.