I have noticed that some commands, even even typed in the Terminal, disable the Terminal, with no seeming connection from the wording itself.
Secret Menu - Enabling the Terminal
Secret - Disabling the Terminal
Also there is a huge issue with Context loss.
For Instance, the Context Test ‘Repeater’ which I created and use for testing memorability from such systems.
After the second (sometimes overlapping and its after three, but rarely!) Prompt following the Information you wish it to keep in mind, or instructions. The Terminal Rabbit loses context to the VIp-Info or Instruction Set. Is this intended?
Same in Italian language (also if for now it is not supported) but I have seen the behaviour you described, more different terms triggers the setting actions, but this particular one that disable terminal it’s really annoying!
The Terminal Issue for example is, when in a prompt the words ‘Secret’ are included, it triggers the terminal to shut off/ turn on, the same is when ‘Secret Menu’ is included.
Also it gets stuck sometimes and repeatedly opens settings, when a prompt includes the word ‘Menu’ , no matter what the actual task was
With Context loss:
The Repeater System is designed to check how many consecutive answers an Agent can provide, while adhering to the context of the given task.
It seems that the agend is able to repeat his orders, or follow orders from a task for only two to three consecutive answers, after that, it forgets its original task
I just want everyone to know that the issue with certain queries triggering the settings menj was now fixed as of the latest update! Thanks team for listening and quickly updating! Thanks @simon
Re: The repeater test you created: this is a super neat idea… but I don’t know if it accurately represents the capabilities… the intention triage system (the thing that made settings open way too often in the past) can route your request to various configurations/LLMs…
IE - when you ask “go to amazon and order the items on this list” and use vision mode to parse a shopping list, their are at least 3 different systems in play that only have the appropriate amount of context for their task:
the intention triage just needs to know what you said
the vision agent just needs the picture and instructions from the triage agent
The web agent needs the full task, the list, and context of everything it does from then on (the sites it visited, the buttons it clicked, the original instructions, the content of the list, etc.)
So, depending on what you are trying to achieve (other than static testing/benchmarking) will influence the result.
I hope that makes sense TLDR; We’re always working to improve things - but in this case, I think it’s probably working as expected (memory wise)
Simply says “feedback received” it will trigger it to reproduce I think you should able to get it. Another trigger is “Go to put a feedback on Amazon”
Many times if the word “feedback” is inside a context, will trigger a functionality that everytime says that the service is under maintenance, this could be reproduced sayng “Come to put a feedback on eBay” or “Come to put a feedback on Amazon”