On .NET Live - Make your first NuGet library with .NET 10 and AI
dotnet
0:00 Heat.
0:01 Hey, Heat.
0:01 [music] [music] Heat.
0:27 Heat.
0:28 N.
0:32 [music] Heat.
0:44 Heat.
0:46 N.
0:45 [music] Hey, hey, hey.
0:55 [music] [music] [music]
1:27 [music]
1:35 Heat.
1:35 Heat.
1:37 [music] Heat.
1:43 Heat.
1:44 N.
1:48 [music] Heat.
1:53 Heat.
2:08 [music]
2:16 Heat.
2:16 Heat.
2:21 [music] Heat.
2:25 Heat.
2:52 [music]
3:04 Heat.
3:04 Heat.
3:16 Welcome.Net friends.
3:18 You know what time it is.
3:19 It's on.net live.
3:21 This is 60 minutes of unscripted.net entertainment.
3:26 And our mission here on this show is to empower all of you,
3:29 our .NET community members, to achieve more.
3:32 I'm your host, Scott Addie, and I'm joined by co-hosts Myra Wenzel and Cam Soer.
3:39 Hello.
3:39 Today we are joined by guest Weey Lynn who will talk
3:42 to us about making your first Nougat library with .NET 10 and AI.
3:48 We would you like to introduce yourself to the audience?
3:51 Thank you.
3:52 And uh I'm Wade and uh I'm the software's uh open source contributors
3:59 and uh right now I also as a MVP now and uh my project
4:04 like mini is like pictures uh is D foundations purchase and the other
4:10 project like mini world PDF uh especially mini PDF is today's uh target.
4:17 This project is uh I create uh two months ago and uh
4:26 is very different from the my uh previous projects like mini.
4:32 So I need to read a lot of open
4:36 uh source XTML document and read great projects like
4:40 POI and uh E+ or library office and uh
4:47 after time pass I will forgot a lot of details.
4:51 So I need to use issue tracking.
4:55 But right now by modern net uh foundations
5:00 and with the copil AI we can do different
5:04 things and I will show you uh today how to quit from Lo to like this project.
5:14 we can get uh wrong uh products productions projects.
5:23 Okay.
5:23 Thank you.
5:24 All right.
5:25 Um we need to get one one uh thing out of the way
5:28 before we jump into that topic that you just described for today.
5:32 Cam, would you like to cue the the video?
5:36 All right, folks.
5:37 Microsoft Build is coming back to San Francisco this year,
5:41 June 2nd through the 3rd.
5:43 You can expect interactive sessions, meetups,
5:46 and workshops from some of the biggest names across.NET,
5:49 VS Code, Visual Studio, GitHub, and more.
5:52 As a reminder, registration is open.
5:55 If you were to go to aka.mbuild, you can register.
6:00 We are excited to connect with you both
6:03 uh or either in San Francisco or virtually,
6:05 and I will actually see you there this year.
6:08 Um, nice.
6:10 All [laughter] right.
6:11 [clears throat] So with that uh behind us, thanks Kim.
6:15 Weey let's get into the topic for today.
6:17 Let's let's start discussing that.
6:21 Yes.
6:22 Uh about today discussing I think uh the first
6:27 uh about a lot of developers would like like
6:32 me I would like have the projects can shares
6:35 on the world and help the uh Google's developers.
6:39 So uh if you would like like me uh today
6:43 I will show you how to do it by my experience.
6:48 Okay.
6:49 And uh okay this is the purchase and I can show you this from uh mini sales.
6:59 a lot of users feedback uh they would like
7:02 to have the functions to convert ESL to PDF
7:07 but today like docker uh docker solutions or like others
7:14 uh library solution then first one is is not free
7:18 or we need to install a lot of things
7:21 and is wer so it means more resource usage and more
7:27 time so I create this chargers and uh I
7:32 will show now how to create it and uh let me into the code and the first ones I
7:41 used the uh copilot copilot and how I use copiler
7:48 to develop this project is first one is I
7:52 quit here and then this is the project structures here
7:58 And uh here and uh you can see here
8:04 the project started and this is the I project how
8:08 to run it's first ones we need to uh have
8:13 the function to test uh like big LM will uh
8:21 ask to check what you do for example Right now I have the two uh cell and uh it
8:35 converting to image and I let compilers to discrete
8:44 what is the difference between the two image and then
8:50 system will describe it and then web it to prompt
8:59 and send it to the compiler to say I have this difference and how you can fix it
9:08 and system will go to wrong and to try to fix
9:13 it like a borrowed here here and the top
9:19 one is my pro the system generations it's very quick
9:23 overview about The detail we will talk later and uh
9:28 the images generate the prompt and the prompt sent
9:32 to the compiler and to run like our um
9:36 VS code GitHub compilers like that the function like
9:45 that and the system will try to fix it and is
9:50 not enough because if we uh a lot of experience.
9:56 I need to uh sit in front of my computers and uh
10:02 I need to wait uh finish until the next step and then
10:08 I I check the result and type the next step description then
10:14 run again and then you have to sit sitting on your uh seat.
10:21 But right now in my experience now I I like this midnight I can
10:28 uh run the program and I go to sleep and the next day morning I
10:35 will come here to check to check the github this for example like that is
10:45 still running project and uh if I get the d Here and check here.
10:59 Here the system will uh pick the diff
11:05 difference and I will check the logic and about
11:10 the previous experience I have read a lot of document
11:14 to know the LM is is correct or not.
11:20 But now we can like that and just said uh please please uh teach
11:31 me as a student uh teach me what you do what you do anything
11:38 else and you will show me what he did and I can review
11:44 the logic and then quick to understand what he do is correct or not correct.
11:51 Then uh we will go to the next section is not uh through my key point.
12:01 I still have to sitting on my seat.
12:06 How to resolve it?
12:08 So the solution is if you can make a self
12:16 involving uh test function it can resolve loop for example like
12:23 that here is my assistant here you can you can check
12:32 here we have the scope if you have the uh provision experience
12:39 you will know list word s as I you will have
12:43 the uh description and the gift code between the two images
12:48 by programs uh and compare their binary or others uh um
12:54 key point yes and uh we can get distance uh stored.
13:03 So if we set a as expect results scope
13:09 for example like here I said 99 code is enough.
13:15 If bolded, you will need to react to our compiler
13:23 system or our uh self evolving system to check again
13:28 what you generate images and give the different description again
13:38 and send to the compiler to say please fist again.
13:46 But if you have the experience like uh experience like that if
13:52 sometimes uh right now AI is not smart but sometime they
13:58 will like blocking in a circle time and the one time
14:03 and one time again like a circle and then is broking is block.
14:09 So we need to set up your mess attempts account.
14:16 If you don't do it, you you will waste a lot of tokens on this system.
14:22 So my purchase is that I like that.
14:27 Right now I said two time attempts.
14:31 If over two times I will use the reviewer ideas.
14:40 My uh key uh the primary LM is copilot opers to develop
14:50 and uh give description but if like a bonet recycle bonet I will join
14:59 review like GPT called D right now you can use GPT 5.5 it
15:06 is in the latest model I think is better but I didn't try on.
15:12 So I will give the reviewer to check what is the pro do and uh
15:22 he will say oh your key point is there and uh why you don't
15:28 uh fix it this key point and because is uh reviewer idea is different
15:35 model so the f percentage is more lens one model to involing by himself itself.
15:47 And the we can have the uh standard of the evolving cycle ten.
16:02 we can get the basic uh self update AI functions to develop your program now.
16:13 Okay, this is the first versions I make v1 version and uh
16:22 here why the laptop I have used the skills the first one
16:31 I use the this part I can uh have a great press
16:35 through uh resolve my basic uh or word convert to pdf function function.
16:43 But but until advantages uh function like
16:53 uh char or like uh formulas computations
17:01 uh lm will not uh have don't have enough uh knowledge to resolve it.
17:11 Why?
17:12 because uh normals uh LM like power let resolve
17:17 a lot of uh questions uh in the world.
17:22 So is very uh prefer to general but if
17:27 you would like to uh through specific uh professionals
17:34 questions you have to change your own your own
17:41 uh data source and here we can use the skill.
17:49 how we can use skill to resolve the question.
17:55 Here I use the two great projects.
18:00 First one is library office and the second one is POI and my uh other is my own
18:06 projects to change the uh my own standards
18:13 knowledges data source and how I can store it.
18:18 I will use very right now very famous uh tools of obsidian
18:25 here I can show you here why obsidians can resolve the problem.
18:35 First one is obsidian is like a uh our spice uh memory link like the samples.
18:44 I have learned in her inherent test problem
18:49 and the previous experience system will uh load a lot
18:55 of data from u Google browsers or Google Chrome
19:00 or uh GitHub or LM teams itself to resolve it.
19:08 But if you use the obsidian,
19:12 you will just be back here only one two three four five six seven uh
19:24 valid and uh merge to maybe one prompt or merge into ISG feedback to our maybe
19:34 co-pilot assistant and uh it were very helpful to reduce the LM uh get question
19:45 and then because our contest is linked
19:50 and very specific for one a professional park.
19:55 So, I I I want to double click on this just real
19:58 quick quick here uh because this is actually very relatable to work
20:02 that I'm doing um at Palumi lately with uh LLMs and reviewing uh
20:09 docs PRs and um I I love what you've done here with Obsidian.
20:15 I don't know if folks are familiar with Obsidian,
20:16 but it's a uh it's a notetaking app
20:19 that just uses markdown files in your file system.
20:22 And it does this really cool graph that he's showing us here where
20:26 it links all it shows you how all the documents are linked together,
20:29 which I think is like really great for uh
20:32 for you know skills like he's talking about here when you're
20:35 building an LLM skill and you can see where it's
20:38 going to use other skills and and tie tie back.
20:42 [clears throat] Yes.
20:44 And uh yes I think uh yes like you said uh previous versions I use I quit
20:53 my own LM by Lama but uh but [clears
20:59 throat] it's not better than obsidian in my 305
21:04 because uh my own training because I don't have
21:08 enough powerful computers because of the cost and um
21:14 uh and it is uh just vector and uh you have to do a lot of fightings.
21:23 So about the obsidian how I do it to link to each others here uh you can use uh
21:33 if you don't want to create your own uh program
21:38 by coding you can use this one and you just
21:42 copy this and uh coron this and run it
21:46 and for example like that I have coroned Other I have
21:54 colon library office code is C++ project and the P1
21:59 project like here it is Java project and uh you
22:03 use this one and tell him uh oh first
22:09 you have to install uh the Obsidian skill but is
22:13 easy you just need to ski search and you can
22:16 very quick to start then you use the skill then
22:21 system said uh please uh prepare and read one
22:29 by one about the source code and to uh to insert
22:37 it by Obsidian COI to insert the markdown like we
22:46 said to uh Obsidian and it's not The important part,
22:52 the more important part is here actually here uh top one is
22:59 about implementation detail but this part
23:04 is most impaired important because if uh
23:09 a smart LM will uh to say merge how they can link
23:16 to the together and is very uh hard work uh previous experience we don't have
23:28 the LM before and they can do it very quick and uh we
23:32 just need to keep it wrong and wrong wrong and and uh use
23:37 my uh uh token and then uh return to our Here we have
23:50 the data obsidian data source and uh we have the skill about skill here.
24:00 You just need to uh create a skill like uh let me stop first.
24:11 Here is our skill here.
24:19 Uh skill is very uh easy to use function.
24:24 For example, like that you just uh have to use
24:30 uh a skill markdown to discrete how can use.
24:39 If you have the experience about semantic,
24:43 let me show you the semantic uh or lance is experience.
24:54 I can show you how the detail about the working process.
25:03 Here is semantic pocket this one.
25:08 And then and we can this project is uh my projects about
25:13 the QM system and by AI and uh we can use the function
25:19 coin like here this description like if you users input uh
25:26 a normals report uh function and uh please run this property function point.
25:35 This is mapping to our skill description here.
25:41 And about the kernel, this kernel function is mapping to the this name here.
25:51 And here you can see the very different parts.
25:54 Here it is just the uh general contest language.
26:00 But here we can use the uh specific programming contest and uh
26:08 it is different because uh like obsidian skill they just put contest
26:15 here say uh I put a link about Obsidian DI document link
26:24 and you just need to query first and then uh the page
26:31 will tell you how to use it and about some other uh
26:37 detail is enough for uh maybe 80% uh uh uh uh you
26:47 your test but if you don't you would like to uh s
26:53 like my uh mini PDF uh issued you you can do it.
27:00 You cannot do it like that because it is system if you use
27:05 the normal language content system will get
27:09 and get will uh means dynamic context.
27:14 So if we would like to do the specific
27:18 uh functions like that I have to do like that.
27:23 I haven't do the screen like the screen folder here.
27:29 I can tell system you have to uh use
27:35 the red gas to search first and the searchware.
27:40 Searchware is here.
27:42 I tell you you have to search the the the name here is about
27:53 here and uh about the topic mapping
27:57 where is the table function is table function
28:00 is mapping the uh w xpf table here is here you can check here it's
28:11 mapping So system uh don't need to guess what your uh proper uh target is.
28:19 It can step it it can step a lot of token
28:24 and the second it have better uh biggest function target function.
28:34 Okay then return to our workflow.
28:38 Here we have the skill and to by sweep and by skill markdown
28:44 and get the data then return to our uh proc and she will go
28:53 to the next part to use like I said GPT reviewer to check if
29:00 they bark two times if don't and the lamb scores over uh 1.9 99.
29:09 They can go to the next part like my system here.
29:15 One part finish or block two times then skip to next one here.
29:23 Okay.
29:25 And you can see here uh like uh biologically is very
29:31 per uh perfect uh flow but you can check the score.
29:39 It improve is not huge.
29:42 Maybe every cycle only can uh increase just 1% or even 0.5%.
29:53 Why?
29:53 because uh right now uh about the MN like here
30:00 I show you like here if you don't uh check
30:10 and uh to understand the detail about your program for example
30:15 like that system where we remove your Uh maybe uh
30:24 the second cast they have finish the auto scale function
30:32 but next you remove it because you just want to uh
30:38 increase your scope because LM just have a target and they
30:43 just want to uh match the your target by any way.
30:50 So if you don't check it,
30:51 it will remove it and you your you your your token is west the second test.
31:00 So you have to check then like like uh the second page
31:09 like that you can this is block a lot of times uh
31:18 hold nine just block here because uh system don't understand the border
31:24 because uh in system is very stimul they can very easy to understand.
31:31 This border is uh many maybe five piece px uh the left
31:39 one is one px is not easy to understand and descript the description.
31:46 So in the morning uh I have to check and uh to go into the code and uh to see
31:57 the detail why he [snorts] broke here and I have
32:02 to read uh like I said please teach me uh why you
32:10 block and please uh describe uh uh link I can
32:17 read and then I read and fix it by manual because
32:23 uh in my experience you face it is manually is uh
32:28 more quick than AI to recycle wronging a lot of things.
32:34 Yes.
32:37 Okay.
32:39 And uh this is about the whole process about the system and until
32:44 now is this there any questions and the next part I will explain.
32:53 Thank you.
32:54 So let me let let me see if I if if I if I've got the the lay of the land here.
33:01 So, you've got your your your open- source PDF library,
33:06 and you've got um uh acceptance tests built into your CI/CD
33:13 workflow that take they they do the they do a automated comparison
33:19 of the before and after of the PDF output and then uh uh
33:26 they they put like some deterministic values on that using the SSIM comparison.
33:31 And that's where you compare two images and and how how different are
33:34 these images and you then let the um the AI make improvements iteratively.
33:43 Is that is that correct?
33:47 Yes.
33:47 Okay.
33:47 So yeah, that's really interesting.
33:50 my So, we're actually doing something like that um
33:53 at Palumi actually where we're using it's called a it's called
33:56 a brownie and ratchet where you you uh let
33:59 the agent make improvements and then you compare it against some
34:03 determinative standard and if it's better you accept it
34:08 and if it's worse you don't and and over time
34:13 the it tends to push toward toward um I I
34:17 don't want to say perfection but a a more perfect product.
34:23 Yes, that's brilliant.
34:24 Brilliant.
34:24 I love it.
34:25 Thank you.
34:26 Thank you.
34:27 And yes, and uh this is my I guess like
34:32 you said and this is my first time to try it.
34:35 I I think maybe it can uh better than better uh after maybe one month
34:44 by one month or it will be maybe
34:47 worse because according to the project more larger
34:50 and uh right now the content size cannot
34:54 enough to understand what the project is maybe
34:58 it will more worse and you have to wor uh uh spend more time to manual core.
35:05 Mhm.
35:06 That is my test to check how the LM can do it.
35:11 Uh not yet what I can do not right.
35:18 Yeah.
35:18 But I see on you get that you are publishing like several
35:22 versions of this this package and like making iteratively better and better.
35:30 Um um and I also saw that you
35:32 have like several other packages that you maintain there.
35:37 [laughter] So you've been busy.
35:43 Uh yes.
35:43 Uh because uh if I just a person uh I cannot handle
35:49 projects like like but we we we are a team like samples.
35:54 Yes, we have a team partners contribute.
35:58 That's great.
36:00 I wanted to go back to something Cam mentioned.
36:02 Um, now I want brownies.
36:04 Thanks, Cam.
36:05 But [laughter] was it brownie and wrench?
36:08 That was the term.
36:10 Brownie and ratchet.
36:12 Ratchet.
36:12 Okay.
36:13 And what is that again?
36:14 It's an iterative improvement.
36:16 Yeah.
36:16 So, so basically you you you apply some sort of a deterministic value
36:22 to score each change and if the change is better, you accept it.
36:27 And if it's worse, you don't.
36:31 Interesting.
36:36 [laughter] Yeah.
36:39 Yeah.
36:39 I've been I've been doing a lot with with agentic code lately.
36:42 I know.
36:43 Uh it's a very people I think are
36:45 are it feels like the the community is divided.
36:50 Half half of the community is like, "Oh my god,
36:52 I don't want to hear any more agentic
36:54 anything." And then the other half is agentic everything.
36:57 And um unfortunately I'm in the agentic everything side.
37:01 Um [laughter] like I I I don't even read my own email anymore.
37:04 I I have a rule in Outlook to just archive everything.
37:09 Market is red and archive everything and then uh twice a day I have it
37:13 do a digest of what came in and just let me know if anything was important
37:17 and then any emails from me go straight to trash.
37:21 Exactly.
37:21 That's exactly right, Scott.
37:24 Wonderful.
37:25 All right.
37:25 I I don't see any questions in the chat.
37:27 Um should we move on to the next piece way?
37:31 Okay.
37:31 And uh about the next step uh about uh if we have enough uh projects and uh
37:39 you would like to uh uh push to the Nougat and let the uh word dyner to use.
37:52 you can uh use like a pipeline like that.
37:57 I can show you here is about the workflow here.
38:04 And uh the first if if you would like to have a lot
38:09 of people to uh cooperate to each other to improve the project,
38:16 you have to have a good uh CI/CD especially CI uh because is
38:25 more easier to then you understand is the uh PRs or code modifications.
38:34 is good or not.
38:35 very quickly.
38:37 And uh here like here you just need to set up your unit test and uh
38:48 see and you will get the result is pass or not in AI uh not
38:56 you can very quick to do it like you just need to set uh like
39:02 that please generate uh uh CI/CD the SDI
39:08 and uh here to create the process board.
39:12 Especially I would like to set uh here I I have the wider screen
39:19 about the detail because uh like I said I installed a lot of tools
39:27 like obsidian and uh uh copilot DOI and uh library office source code
39:37 or PI source code in my computer but online did H doesn't have the environment.
39:45 So what what I have to do to check if somebody have uh standard
39:50 PR and I would also like have the uh images uh checking different functions.
39:59 I just uh put here and uh while the a very uh very quick
40:11 function just said uh please use this uh AI you you can um uh
40:20 you can set up the AI function in your own uh workflow function
40:27 by your AI here like me here I use the open AI API and to check
40:35 if somebody send the PR and check by CI workflow and check the two
40:41 picture is uh similar or not similar is enough a score or not you
40:47 can write script here it's very helpful yes then if you pass the unit
40:55 test then you can very quick to uh have a CD to debour your project.
41:02 Like here I I have a web to uh skill like here just a markdown very easily.
41:17 And how how I can use it right now?
41:21 I can like here and here and here.
41:27 And uh I just said uh please generate less tech GitHub
41:35 tech last ver version and uh until current commit version and uh
41:42 please let the uh versions numbers um at uh one point
41:49 uh 0.1 and then I just play quick and you will hit
41:55 the steel uh uh release skill then system will run and to auto
42:05 compare the last tech until now uh my commit to generate
42:12 the contest why I what is the features why is the bug
42:18 why the uh uh others it's very different from the our previous
42:26 time because in previous uh we have to w by manually
42:32 here and sorry uh we have to write manually and we
42:42 have to remember what we did and uh one by one
42:47 to check here and to write down what we have changed here
42:53 is uh maybe it's not is text but about the uh
43:02 like the burrow one it will be a uh hot task
43:09 for example the mini PDF I use the two modern dominate functions
43:16 first one is bracer bracer here we can like Sorry, sorry, sorry.
43:30 Is [laughter] online demo here.
43:35 We we create I I create a online demo here is uh let me change to English here.
43:42 It is browser client only.
43:46 So uh system would not uh store any data
43:53 in the server or not just store in your client.
43:57 It's very uh it is very different from my experience
44:05 because uh in uh before uh before we have to uh
44:11 basically have a uh converting server and uh you have
44:16 to install a lot of things and right now we can use
44:19 the bracer and just here to just select a PDF
44:25 and convert And uh she uh system will download the phone
44:29 surely because uh prompt or just have a few uh font
44:35 because the uh license is so you [clears throat] have to
44:40 and the lamb you download and the system uh will
44:45 uh here I use is a I saw that [laughter]
44:53 and here we can hear to it convert and then
45:00 we don't need uh any server function in any server yet.
45:07 Great.
45:07 So it's using the static uh functionality from blazer.
45:11 Yes.
45:12 From blazer.
45:13 Mhm.
45:13 So it it use the uh because it compile to uh
45:21 uh binary resource and the chrome uh uh the browser
45:25 will uh unpack it and uh loading in memory
45:30 and uh we can use very native functions on our server.
45:35 So it is is the images that uh but is you need to install a lot of uh font.
45:45 So for example like uh Microsoft license bond I cannot
45:50 provide uh client installation because it is not il is illegal right
45:58 so so [clears throat] I provide a cloud API about the cloud API uh this mean
46:06 we need to have a function uh server uh server function and uh here I use
46:13 the uh our Ashur's Ashure uh FP service and return to our CD we about the braers
46:26 or Asure service we need to have a lot of code before but right now we just
46:35 like that uh tell the uh compiler and let compiler to set up the uh initialized
46:45 version here and lamb here release publish here
46:58 I can here to set up our process yaml
47:03 and uh very easy and it on our GitHub server and set up the docker and uh
47:10 to build uh uh to install the d to build and publish and use the code.
47:18 Please do not use the uh please use environment key
47:25 key or you will waste your money here and uh
47:29 we can set up here to secret key then you
47:33 will publish our brace project and azure a service to system.
47:43 lamp.
47:44 We can check here our action here.
47:59 Our CD deploying here system will use
48:12 our secret code to log into Azure to deploy our Azure service and the traers uh
48:21 program lamb every time it just automatic codies.
48:27 So you just need to focus while you coding just
48:33 focus on that and the lamp system can uh deploy
48:37 the system and then you can easily to use the latest
48:42 version maybe like online demo or your AOT COI version.
48:50 Here you can uh pack to to uh different service.
48:59 Oh about the sorry about the AOT I don't tell I also let the uh I
49:06 also create uh AOT functions first time I
49:11 think is very difficult uh hard test because uh
49:17 if you would you if you would like to use the AOT this means you just need
49:23 uh you use the C++ function to compile
49:27 your C# charges to uh binary uh executions program.
49:35 It mean you cannot use a lot of uh third party library.
49:40 You just have can use native uh uh methods.
49:46 Let's mean huge test developing coding test but it
49:52 did it and it it run very well.
49:55 So uh like here uh after the CD system will uh
50:04 compile uh the source code project then to Linux and Mac
50:11 OS system and the user just can download it or they
50:16 just they don't want to download it they can use
50:20 the uh uh ste uh your cmd like that to send
50:27 it to install uh screen lamp you can install it
50:34 and use it use the mini PDF CI to use it
50:39 it's very uh powerful and very uh I cannot image default
50:47 yeah well and I I seems like you're having
50:50 a lot of fun I think with the with all
50:51 the CI/CD engineering I know that That's something I
50:55 that's something I didn't expect with all of the agentic
50:59 coding lately is all of the work I'm doing
51:01 in GitHub actions lately um to to make stuff
51:05 like this work and it's it's really uh I
51:08 I really like what you've done here uh way.
51:10 It's it's uh it's very clever and I think the more you do the more it's like
51:15 oh this could be better and like if you you think about
51:18 improvements you can make to your workflow that would speed things
51:22 up or improve some some part of the review process or etc.
51:28 I don't know.
51:28 At least I keep having um like for for the for the projects that I own
51:34 internally here at Microsoft like I keep thinking
51:38 about ways that we can make things better.
51:42 Yes.
51:44 And um I I'm curious though with your versioning strategy like I
51:48 saw that your mini PDF for example is at 0.26, right?
51:53 So like when do you feel like it's going to be ready for like a 1.x version?
52:00 Yes.
52:01 And uh Yes.
52:02 Because uh if if more uh sample like the pictures here image here and uh
52:11 they pass the 99% it can give me the uh uh brief to to make
52:20 sure the project is uh or not because uh yes about we we we
52:30 feared uh some spark or some issues uh inside the deep uh coding part.
52:41 So I just can let enough sample to let me improve my brief.
52:52 Yeah, thank you.
52:53 So I'm looking at the the release history here.
52:56 So uh it seems like you you started this project back in February.
53:01 There have been numerous releases since then, so you've been quite busy.
53:06 Um, if someone wanted to get involved and contribute to this project, uh,
53:10 are there any any areas where you're looking for help from community members?
53:16 Yeah.
53:17 Yes.
53:17 Uh, I think about the project functions.
53:21 I think uh today I would like to share
53:26 the process and uh if some would like to uh uh
53:33 together to uh improve the project uh they can provide
53:38 their token because uh I used a lot of token here.
53:42 uh for example I use uh basically I use the opus it is uh uh 7.5 yeah
53:53 every day every request so I I buy I I right now I have three account about uh
54:02 of the pro uh GitHub copiler pros so every month
54:10 I I need to as a um a little money.
54:16 So So So you you need people to come
54:18 and and help spread out the spread out the token spend.
54:23 Yes.
54:24 And [laughter] uh it's a it's a true it's
54:29 a true problem we have right now with tokens.
54:36 Yes.
54:35 And uh I think uh but thanks uh Microsoft
54:40 they provide one account I can free to use
54:43 uh because uh MP but uh is and uh I would like to say uh like about the copilots
54:55 because uh lot of people would like to uh
54:58 like to use proco code or code uh the CI
55:04 they they are better but in uh copiler we
55:08 can use uh very powerful uh powerful function like that.
55:16 Mhm.
55:15 Uh we we can use uh difference model.
55:20 Yeah.
55:20 like yes yes for example uh about the imagic difference descriptions uh
55:29 I will use the three model to to to help first one is
55:36 cloud opers and the second one is GPT and the second one is
55:40 gven it is very uh easy to use in uh compiler uh now
55:47 yes so I I think is a very good part of my experience.
55:54 [clears throat] Just a note.
55:59 Awesome.
56:00 All right, we've got about five minutes remaining in the show today.
56:04 Are there any resources you wanted to share
56:05 with the audience before we wrap things up?
56:09 Okay.
56:10 Uh I think uh let me think.
56:15 uh I think right now uh if about uh AI development I think is
56:24 enough but if you don't uh you don't want to spend time to understand
56:32 the detail uh I think about a small part is okay but about
56:39 a uh a bit uh bigger larger part I think it is not smart enough.
56:48 Now my experience maybe others people have more
56:52 powerful workflow or others uh function methods to resolve
56:58 it but is in my experience that so uh is the seniors or or our develops value.
57:10 So we can understand the detail and uh to know what the uh LM would like
57:19 to teach us and uh where we can find the solution and understand what they say.
57:28 So I think I I would just like to uh make a note here.
57:34 [laughter] Right.
57:38 Sounds great.
57:39 Well, uh, Weey showed us a ton of, uh, new stuff today.
57:43 I I learned a lot.
57:44 Um, please do go out and check out that mini PDF pro uh, um, project on GitHub.
57:51 Um, sounds like Way is open to contributions there
57:55 and willing to accept your tokens with open arms.
57:59 So, [laughter] please do use your tokens on that project.
58:02 Uh, we thank all of our viewers for tuning in today.
58:06 As a reminder, this was the on.net net live show.
58:09 Uh you can check out other uh streams we've done in the past.
58:12 They've all been recorded.
58:13 Those are out at doot.netlive.
58:17 We will be back next week and we'll be
58:19 joined by guests Zo and Tunes who will join
58:22 us and talk about a project called Outbox Kit
58:26 which is all about implementing the transactional outbox pattern.
58:31 That's a mouthful.
58:32 Hope to see all you next week.
58:34 Thanks again Weey.
58:35 We appreciate your time today.
58:37 Thank you.
58:38 Thank you.
58:39 Good day.
58:40 Bye.