What I thought, I would never hear from a Microsoft core developer: "The shortcut is 'Shift+Command+I'. He might as well have added, that is 'Shift+Control+I' for those who are still using Windows.
And where to find the Plugin Store? Like they mentioned it as if it was already available... Was kinda disappointed to see that copilot ignored my request to open an issue
Just like what the video said at 31:00--- It takes time to debug. especially those panic bugs that only happen when the code is running. Hopefully, one day Copilot can handle the messy running panic bug and give constructive ideas to programmers.
Good video. I am really interested in How it works with microservices? Like you have 100 repositories, can copilot “follow a code path” accross multiple repositories ?
The LLM itself has no need of a concept such as a repository. You simply give it code and keep giving it more code until you're satisifed that you've given it enough or until you fill it's entire context window. It won't matter if the code lives in 1 repository, 1,000,000 repositories, or 0 repositories--it won't have any effect on the model's ability (or lack thereof) to follow a given code path. The Copilot system as a whole is more likely to have some form of code boundary defined in terms of the client' local workspace(s) rather than any kind of boundaries defined in terms of the underlying revision control repositories.
Other than GPT4 + Context Window which is good, I'm seeing the inclusion of workspace folders which is huge. If AI has somewhat a more complete picture of the project.. it makes things more relevant. GOOD STUFF! :)
@@festiml.2561 I had the same experience at first, but changed my mind after more testing. The difference is that GPT-4-turbo follows system instructions much more accurately, sometimes to the detriment of reacting to user messages. It can be annoying, but in aggregate, I think it was the right move. To create more "human-like" agents, I ask them to "embody their persona". I then develop my agents by asking them to reflect on and update their system prompt as we go along. This worked because GPT-4 "classic" will follow instructions from user prompts even when they contradict their system prompt. This no longer works (I have to build that escape hatch into the system prompt to get reliable results). Similarly, when reacting to incoming emails, since the email body is added as a user message, a malicious actor could override the system instructions quite easily and steer the agent. We had to break our agent up into two layers, "understand email" then "react", to prevent this. The new model is much better at avoiding this out of the box (though we still take precautions).
after seeing lot of hype just purchase copilot and guess copilot not able to generate a small script to generate random data :) hahahaha just a waste of money!
omg, as if github copilot isn't already overpowered, it just got even better with voice prompts.
What I thought, I would never hear from a Microsoft core developer: "The shortcut is 'Shift+Command+I'. He might as well have added, that is 'Shift+Control+I' for those who are still using Windows.
Potato Potahto
Good feedback. Next time I add more shortcut annotations to the videos 👍 /Harald
Damn…
@@PATRICIOcomplex :)
(:
Where can I find the documentation on building the skills plugin and the agent plugin for copilot please?
And where to find the Plugin Store? Like they mentioned it as if it was already available... Was kinda disappointed to see that copilot ignored my request to open an issue
Just like what the video said at 31:00--- It takes time to debug. especially those panic bugs that only happen when the code is running. Hopefully, one day Copilot can handle the messy running panic bug and give constructive ideas to programmers.
Good video. I am really interested in How it works with microservices? Like you have 100 repositories, can copilot “follow a code path” accross multiple repositories ?
Unlikely as it's based on GPT
...3.5 which is really bad @@djstr0b3
The LLM itself has no need of a concept such as a repository. You simply give it code and keep giving it more code until you're satisifed that you've given it enough or until you fill it's entire context window. It won't matter if the code lives in 1 repository, 1,000,000 repositories, or 0 repositories--it won't have any effect on the model's ability (or lack thereof) to follow a given code path. The Copilot system as a whole is more likely to have some form of code boundary defined in terms of the client' local workspace(s) rather than any kind of boundaries defined in terms of the underlying revision control repositories.
Other than GPT4 + Context Window which is good, I'm seeing the inclusion of workspace folders which is huge. If AI has somewhat a more complete picture of the project.. it makes things more relevant. GOOD STUFF! :)
Will Github Copilot Chat support vim?
No
Accenture - "Help customers"; yeah right.
I found the "we care about our employees" even better
@@manuelrenner5164 😂
Accenture : I sell you and keep profit for me
Awesome Preview 🎉🎉🎉
As a developer, this is how we shoot in our own legs. But we can expect better software which solves more problems.
Amazing stuff!! Nice work!
Copilot should just use GPT-4-Turbo instead of 3.5 Turbo and GPT-4, it's faster, cheaper, and smarter.
Are there plans to do this? Although I can't see MS doing this
except it is worse than gpt-4 from my experience with it
@davior14 yeah, github is actually working on it right now
@@festiml.2561 It is worse at basic question answering I think, but it is better at coding from my experience.
@@festiml.2561 I had the same experience at first, but changed my mind after more testing. The difference is that GPT-4-turbo follows system instructions much more accurately, sometimes to the detriment of reacting to user messages. It can be annoying, but in aggregate, I think it was the right move.
To create more "human-like" agents, I ask them to "embody their persona". I then develop my agents by asking them to reflect on and update their system prompt as we go along. This worked because GPT-4 "classic" will follow instructions from user prompts even when they contradict their system prompt. This no longer works (I have to build that escape hatch into the system prompt to get reliable results).
Similarly, when reacting to incoming emails, since the email body is added as a user message, a malicious actor could override the system instructions quite easily and steer the agent. We had to break our agent up into two layers, "understand email" then "react", to prevent this. The new model is much better at avoiding this out of the box (though we still take precautions).
Cursor = copilot X
And they tell me to wait a few weeks
And full gpt-4 with non-programming questions allowed
I Cursor is still way better!
Pilot
As a developer, i have to say I’m scared 😂
We will eventually be replaced - it's inevitable. At least we can ride the waves of AI until then.
amazing ❤
Yeah dude
Super!
Developbot, develepbot, develepbot, develepbot!
Awesome :)
As a developer, thankfully I'm old, and perhaps I'll die soon. Problem solved.
after seeing lot of hype just purchase copilot and guess copilot not able to generate a small script to generate random data :) hahahaha just a waste of money!
what a joke, the bubble will pop soon
u maaaaaaaaad