1. Types of tasks and AI model tiers
What types of tasks is the on-screen AI assistant chimate.ai suited to? And which model tiers should you choose for different types of tasks?
The AI assistant's underlying mechanics allow it to handle a wide range of tasks. The service analyzes the entire screenshot: it recognizes not only text, but also images relevant to the task.
1.1 Simple requests (Simple)
In practice, the service handles tasks in math, Russian, English and other languages, competency tests, hygiene certification tests (Gigtest), quests, logic problems, pharmacology, homework and live coding well even at the most basic Simple model tier. This mode uses a fairly powerful AI model with its reasoning level set to the maximum, but without web search. We therefore state that this tier gives 70% correct answers. That figure is approximate: in actual tests, the result is over 80%, but it depends heavily on the type of task.
1.2 More demanding requests (Advanced)
The Advanced model tier is well suited to tasks that require more complex analysis, such as driving theory tests (PDD), history, medicine, literature, complex live coding, law, specialized tasks that require looking up information online, and exam preparation tasks where getting the answer right matters especially. There are also some rather unexpected examples of successful use: completing quests in online games.
1.3 Complex requests (Top)
The Top model tier is suited to the most difficult tasks and those where getting the answer right matters most. We also recommend this tier when a single request contains several tasks. It uses the most powerful AI model, with its reasoning level set to the maximum and extensive web search. In our internal testing, this tier successfully solves 95% of tasks. However, using it is not always practical: processing can take up to 60 seconds.
2. Limitations
2.1 The answer depends on information that is not public
There is a caveat here: if a definitive answer cannot be found online and cannot be worked out logically, no AI model will help. Such tasks do exist. For example, a question may depend on internal documentation for a particular piece of equipment that is not publicly available. In that case, all the AI assistant can do is smile and wave a paw: it will, of course, give a plausible answer, but unfortunately you cannot rely on it.
Other unexpected requests from users that we have not yet fully succeeded in handling include suggestions for how to continue a romantic conversation on Tinder and advice on how to play an online card game. We will get there, though.
2.2 Hiding the assistant's windows during screen sharing
The AI assistant has two windows: the main window and the overlay. Both are hidden from screen sharing, but to different degrees. The overlay is specifically optimized for use on top of another application, such as a browser window. It does not take focus when you click, which is very important when completing tasks on certain websites that try to monitor user actions. The overlay also supports transparency and has a special mode for activating buttons "quietly", without clicks. All of this makes it more convenient to use on top of another application.
The application's main window uses a different technology and has limited protection against screen sharing: in some cases, it may appear as a black rectangle in the capture. Keep this in mind and test the application in advance for your particular use case.
In this respect, CDP + Web Session mode lets you use the AI assistant even when the desktop application is minimized to the system tray: nothing gets in the way on the screen, while you control the assistant and receive answers on your phone or another PC. This is explained in more detail here.