Onboarding experience in v0.3.3 Alpha

[skycommand]

Hello. 😊

I tried version 0.3.3 Alpha. I'm afraid it leaves much to be desired, not just for the end user, but also for testers who need to repeatedly uninstall and reinstall it on a clean environment while it is in Alpha or Beta stages. Here are the areas in which you can improve.

• Installation: Version 0.3.3 has no installation stage. It just copies itself in `%AppData% and installs no Start Menu shortcuts. You could use Inno Setup to quickly build an installer that installs the app in Program Files, which has strong anti-malware protection and makes the app accessible to everyone user account on the PC. (Developers like me sometimes have three user accounts.)
• Onboarding: The app requires a one-time acquisition of an AI module, which is 1.78 GB. However, the app should allow the user to import this file. For a tester like me, it means I can quickly test the app even if I'm on the move and my Wi-Fi connectivity is funny. For end-users, it means the privacy that you promise them. What better proof of privacy than the app completely working offline?
• Invocation: In version 0.3.3, the invocation key is Ctrl+Alt. This means I'll have problems testing the app inside virtual machines, which reserve Ctrl+Alt for transferring mouse and keyboard control to that the of the host OS. Perhaps you could add an icon to the notification area (also known as system tray) to invoke the app via mouse only?
  • Never mind. I found the tray icon.
• Build instructions: Unfortunately, I don't know how to build a Python app from its source code. Would you please some build instructions?

As I've told you before, I have problems running v0.3.3. After downloading the AI file, I see this:

Here is what err.exe saya about it:

# for hex 0xc000001d / decimal -1073741795
# as an HRESULT: Severity: FAILURE (1), FACILITY_NULL (0x0), Code 0x1d
# for hex 0x1d / decimal 29
  ERROR_WRITE_FAULT                                              winerror.h
# The system cannot write to the specified device.

I'm trying to figure out what's the problem.

I know all of the above is a lot. And I mean A LOT! Please take your time re-reading and asking for clarification. I'll spare what amount of time I can.

[adrianium]

Hi @skycommand! Thanks so much for the detailed feedback on the errors. I've addressed each point as you suggested:

• Installation: I've corrected and executed everything exactly as you advised!
• Onboarding: I appreciate your advice, but unfortunately, I can't enable this feature at the moment. Right now, I'm focused on ensuring my skills work without distractions. If I had enabled it from the start, it would have caused chaos and some skills might not have worked. However, I'll definitely add it in the future.
• Build Instructions: I've added the BUILDING.md file for your reference.

Regarding the running issue with v0.3.3-alpha, I understand you're running from a virtual machine. I have the same problem on my virtual machine. I'm truly grateful for your time and detailed feedback. I've created a separate .exe file to address this issue, but I want to warn you that disabling AVX2 slows down the skills work. I had to make this trade-off to ensure compatibility with processors that don't have AVX2, including your virtual machine.

I really wanted to make this the main patch, but disabling AVX2 was necessary to ensure the tool works on processors without AVX2, including your virtual machine. I hope this helps!

link to no AVX exe-file:
https://github.com/adrianium/Scryptian/releases/tag/v0.3.4-alpha

[skycommand]

Oh, I see now! 😲 The error 0xc000001d is because of AVX absence. Here is the problem: The app window says it’s a WinError, so I looked up winerror.h for the error code. But… if I’ve understood you correctly, 0xc000001d is actually the STATUS_ILLEGAL_INSTRUCTION error from ntstatus.h. I’m a bit ashamed of myself. After checking everything with Process Monitor, I should have suspected that I misunderstood the error code, and it means something else.

Very well, I’m enabling AVX2 flags on my VMs and changing my virtualization app to Hyper-V. And it works. Of course, AVX2 should go into the "System Requirements" section of the README.md.

Here is the result of the Summarize skill.

And here is the result of the "Translate to English" skill.

Looking at translate_to_english.py, I see that it isn’t offline and private. It uses the Google Translate API.

That’s it for now. I have to go back to work. Cheers. 😊

[adrianium]

Wow! Before I say anything, I must mention that I admire your professionalism. When I wrote about the AVX problem due to the fact that the virtual machine’s CPUID is not working well, to be honest, I thought you would abandon the project. But, it's just wow!

I have a dilemma regarding the translator and I hope you can help me as a user. Previously, the translator skill was based on the SLM (small language model). But he was a lot of dumb and even translated terms. And I was faced with the choice of either making the skill completely local but installing libraries and loading the user’s computer even more, or turning to the Google Translator API, which the whole world already uses.

It seems to me that the essence of this dilemma is much deeper. Choose completely autonomous skills that are guaranteed to take up a lot of disk space on the user’s disk (I assume that pros like you will try to automate their process as much as possible with small model intelligence) or transfer some logic to the cloud. The question is friction versus value. Who would, for example, download 100MB+ of translator skills to translate? But if we create a hybrid model with an API, perhaps the skill would weigh in the kilobytes. Only then will friction be lower than the value. These are just thoughts out loud.

I even have an approximate solution - this is to create a skills tab on the site or something like some kind of Steam analogue. where people will consciously see what they weigh. I just understand that SLM’s intelligence is not enough for complex operations and will always ALWAYS be inferior in intelligence to cloud models. But I just assumed that SLM would act as a router to direct tasks for hybrid skills developed using Python. In my opinion, only then will AI be useful. Where you lack intelligence, Python will help. What Python can do will be able to do skills, and Python can do EVERYTHING.

I am very interested and IMPORTANT to hear your opinion on this.

UPD:
I will definitely add a point about system requirements in the future!