Speed Up Your AI Development Workflow by 2x

0:00 Hello everybody, I'm Nick and in this

0:01 video I want to talk about how I double

0:02 my productivity of my development

0:04 workflow by utilizing a couple of

0:06 applications that really speed up what I

0:09 think is my bottleneck and I'm sure it's

0:11 the bottleneck for many of you and it's

0:14 mainly related to typing now because we

0:17 no longer go into an application and we

0:20 just start typing everything manually

0:22 like you'll still do that but a lot of

0:24 that has now been dominated by AI.

0:27 I want to see how I can improve my words

0:30 per minute and I can type fairly fast.

0:33 Not super fast.

0:34 I can do 100 words per

0:35 minute pretty consistently, but in code

0:38 with autocomplete and so on, things slow

0:40 you down more than you think.

0:42 It's not like you're typing down an essay.

0:44 So that word per minute is actually reduced

0:46 quite a bit.

0:47 Now going back into an AI

0:49 agent like clo for example, you wouldn't

0:52 write code here.

0:52 You might paste code,

0:54 but you would write what you want the

0:56 tool to do.

0:57 So for that reason I looked

0:59 into solutions to how I can solve it and

1:02 I found two things.

1:03 First sometimes I just want to type down my thoughts and I

1:07 can go ahead and say in this demo which

1:09 is part of the demo for my online

1:12 workshop I'm running in in 7 days.

1:14 If you want to join there's a few links in

1:16 the description you can check that out.

1:17 I'm going to teach you how to do Vibe

1:19 coding for production at proper proper

1:22 scale.

1:23 something you can actually push

1:24 to production to customers and be secure

1:27 and safe about it.

1:27 And as part of that

1:28 workshop, I'm showing hey, here's an

1:30 application uh and ignore the text being

1:32 small.

1:33 The code doesn't really matter,

1:34 but this application has a few

1:35 vulnerabilities I want to showcase and I

1:37 want to show how AI can sort of fix that

1:40 by auditing it and then suggesting what

1:42 we should do about it.

1:44 Now, I can go

1:45 here and manually type, please go ahead

1:48 and run a security audit for my code.

1:51 However, can I speed that up?

1:53 And the answer is yes.

1:53 And I can do it in two

1:55 ways.

1:55 The first one is with autocomplete

1:58 and not stupid autocomplete that tries

2:02 to hallucinate things based on no

2:04 context or minimal context, but actually

2:07 take a look at what's going on in the

2:09 screen and what I'm typing, process

2:11 that, and then start suggesting things

2:13 based on that.

2:14 So for example, if I now

2:16 start typing please run, you can

2:20 actually see text is appearing.

2:21 It's small because the scaling doesn't work

2:24 well in this big screen.

2:25 When it is smaller, it works fine.

2:27 But if you notice here, it says please run the

2:30 following command.

2:31 It's suggesting.

2:32 And if I was to press tab, it will go ahead

2:35 and accept this word as something I

2:37 typed.

2:38 But I want to say please run a

2:41 and it suggest code analysis.

2:43 If I say S, it's going to say security check on

2:46 the code.

2:46 And I can press tab tab tub

2:48 tab tab tab.

2:50 And it says below.

2:51 I don't want to do it in the in the code below.

2:53 I want to do it in the code of this

2:55 codebase.

2:56 So please run security check

2:57 on the code of this project.

3:01 And I can just accept it and it automp completes

3:03 the words using the context.

3:05 Now how is that happening?

3:07 It must be some LLM

3:08 involved.

3:09 And and the answer is yes,

3:10 there is.

3:10 However, I'm not the biggest

3:12 fan of LLMs seeing what I do on my

3:15 screen and then sending it into some

3:18 server for processing.

3:19 So, all of this

3:20 is actually running locally and I'm

3:23 achieving that with an application

3:25 called co- typist.

3:27 So, cotypist is this

3:29 app over here which I'm sorry it is

3:31 small.

3:31 I'm going to zoom in a bit.

3:33 But Cotypus is an app I'm running on my Mac

3:35 which as you can see is using GMA4E4B

3:40 which is a local LLM.

3:42 At no point is a

3:44 request sent to the server for

3:45 processing.

3:46 All of this is happening

3:47 locally on my machine and I can choose

3:49 how much context I want to give it.

3:52 I can give it some information.

3:53 So I can personalize it.

3:54 For example, I've said a

3:56 few things about me.

3:57 For example, I'm Nick Jobs says I usually write in

3:59 English.

4:00 Please write in a friendly and

4:01 casual tone and so on.

4:02 And you can customize the custom AI instructions.

4:06 And then you can choose tons of things.

4:08 For example, don't use emojis on

4:10 whatever you're writing because they're

4:11 cringe or how much context.

4:13 Hey, use screenshots for context.

4:16 Do not use the

4:17 clipboard for context.

4:18 And it will take

4:20 screenshots and as I'm typing, it will

4:22 look at the context of my screen and

4:24 then do all that processing offline.

4:27 At no point does it go in OpenAI or Google

4:30 or any other service which I feel very

4:33 comfortable about because it's all local

4:35 here and then it will use that to

4:38 suggest things on everything.

4:39 It's not just for coding.

4:40 If I write an email,

4:41 this will kick in.

4:42 This will basically kick in everywhere.

4:45 When I'm on notion

4:45 and working on some notes, it will do

4:47 it.

4:47 When I'm writing a message on

4:49 WhatsApp, it will do it.

4:50 And again, all of this is local and you can choose the

4:54 model.

4:55 So, you don't have to use GMA 4.

4:57 It gives you a few suggestions on other

4:59 models that are working depending on how

5:01 good your machine is and how much RAM

5:03 you have.

5:04 You can choose to have a

5:05 bigger model.

5:05 Of course, a bigger model

5:06 means that it might be slower, but my

5:10 machine is fine.

5:10 I have half a terabyte

5:12 of RAM, so I'm way more than comfortable

5:15 with using something like this and not

5:16 stressing the system.

5:18 And I have found

5:19 that this significantly speeds up not

5:22 just my coding workflow, but sort of any

5:24 other workflow I have where I need to

5:26 type because the context is so good.

5:28 It understands what's going on.

5:30 And this doesn't just work at the terminal.

5:32 It can work anywhere.

5:33 I mean, you can limit

5:34 it.

5:35 So you can go all the way here to

5:37 app settings and say don't work on

5:40 rider, don't work on IntelligJ because

5:42 those things have their own autocomplete

5:44 usually.

5:44 So you can exclude a few things

5:46 that you don't want the thing to kick

5:48 into.

5:49 And you can also do the same for

5:50 domains.

5:51 So if you want to exclude a

5:52 specific domain, maybe your banking

5:53 domain and so on, it will know and it

5:56 won't go ahead and do anything there.

5:58 Now again, this is still fine because

6:00 all of this is local.

6:01 How do I know it's

6:02 local?

6:02 Well, first the model is local.

6:04 But to make sure that it absolutely

6:07 doesn't send anything over the wire, I

6:09 also have radio silence installed which

6:11 allows me to block certain applications

6:14 from ever calling the web.

6:15 So for example, as you can see here, codeist is

6:18 blocked.

6:18 It can't go over the web and

6:20 send requests.

6:21 You can actually go to

6:22 network monitor and guarantee that

6:25 nothing will go to the cloud on any

6:27 other server.

6:28 So this is the first part.

6:31 The second part is that look, I'm

6:34 working in an office alone.

6:36 Why do I have to type all the time?

6:38 And the answer is I don't.

6:39 Usually now what I do

6:41 is I dictate.

6:42 So I talk to my

6:43 microphone.

6:44 And in fact, I have a small

6:46 microphone over here which I always have

6:48 on my desk separated from this because I

6:50 don't want to have this in front of me

6:52 all the time.

6:53 And this can now pick up

6:54 exactly what I'm saying by me just

6:57 pressing caps lock holding it down and

7:00 space.

7:01 I'm using hyper key which is a

7:02 utility on Mac to do that.

7:04 It gives you a more flexible way to do key bindings.

7:07 And if I go and click it, you'll see a

7:08 window pop up at the bottom and it will

7:10 start writing down what I'm saying.

7:13 So just like that.

7:15 And if I just delete

7:16 it, if I show you how the workflow would

7:18 be and and you saw how fast it is, then

7:20 I can simply say run a security audit

7:23 for this application.

7:24 That's it.

7:26 And it's there.

7:27 This significantly speeds up

7:28 my workflow.

7:29 Of course, I can't talk

7:30 every time if it's late.

7:31 I don't want to

7:32 be whispering, please run a security

7:34 workflow for this application, and so

7:35 on.

7:36 But when I'm here and it's the day,

7:38 I want to do that.

7:39 And it's here and

7:40 it's super fast.

7:42 And in the same nature

7:43 as code typist this is using an

7:44 application called typew whisper.

7:46 Now with type whisper again this is a local

7:50 thing that as you could see before in

7:53 radio silence it cannot actually access

7:56 the web everything is here local and

7:58 secure.

7:59 And if I go back I accidentally

8:01 closed it.

8:02 If I go back you'll see that

8:04 I have chosen over here uh a default

8:08 engine.

8:08 I'm using Power Kit because I

8:09 found this is the fastest and most

8:12 accurate engine for this specific uh

8:14 solution I have.

8:15 But Whisper Kit is

8:16 pretty good as well.

8:17 It's just a bit

8:18 slower.

8:18 And then you can choose the

8:19 model.

8:19 Those are local models which you

8:21 can go over here on integrations and

8:25 install.

8:25 So I have Whisper Kit

8:27 installed.

8:27 I have Parit installed.

8:28 I have Apple speech as well which is

8:30 extremely fast but not that accurate and

8:33 contextually intelligent.

8:35 And I can also

8:36 have local file memory so I can remember

8:37 a few of the things I've said to use

8:39 them as context.

8:40 I can also have

8:42 snippets for example when I want a

8:43 hyphen to be converted into something

8:45 else.

8:46 And I can go here and say Nick

8:47 hyphen chops and it will chops

8:52 and we'll go here and replace hyphen

8:54 with that instead of actually typing the

8:56 word hyphen.

8:57 I'm doing the same with

8:58 things like front end design which is a

9:00 skill you would use in cloud or cursor

9:03 and so on.

9:03 You can have your own

9:04 dictionary.

9:05 You can customize everything

9:06 about this.

9:06 You can have file

9:08 transcripts.

9:08 You can have custom

9:10 hotkeys.

9:10 As you can see, I'm using this

9:12 very elaborate one, but these four ones

9:14 are actually just holding down the caps

9:15 lock button.

9:16 And you can customize

9:18 everything about it.

9:19 And again, this is

9:20 appaware.

9:21 So, it can actually

9:22 understand, hey, I'm in an editor.

9:24 I'm in a in a terminal.

9:25 I'm I'm somewhere.

9:26 And it will adapt what I'm writing to

9:29 whatever it can hear.

9:30 So if I want to be

9:31 a bit more elaborate, I can say very

9:33 quickly and that's impressive about it.

9:35 It's actually very very snappy.

9:37 I can say please run security audit about this

9:40 application.

9:40 Make sure that SQL

9:42 injection is not possible with this

9:43 application and also check for any HTTP

9:46 related vulnerabilities that might be

9:48 affecting an application like this.

9:50 So I'm speaking in my very normal voice and

9:53 it's typing way faster than I could

9:56 possibly write this.

9:57 If I try to write

9:58 this, I mean, it's not going to.

10:01 Okay, this is actually reading the

10:03 context and it's going ahead and it's

10:04 typing it with code typist, but as you

10:08 can see, it still isn't as fast as the

10:11 local LLM using my voice.

10:13 It's lovely.

10:14 It is amazing.

10:15 I know tons of you are on

10:17 Macs, and I'm certain there's

10:18 alternatives for this on Windows, but

10:20 I'm not using Windows for development,

10:22 so I I can't possibly know.

10:24 But as you can see, this can significantly speed up

10:28 your workflow.

10:29 I'm hitting 150 to 160

10:32 words per minute when I'm dictating with

10:34 my normal voice.

10:35 And this does not have

10:36 to go to the cloud, involve that

10:38 latency, and involve your recording

10:40 going somewhere, which I think is

10:42 lovely.

10:43 But now I want to know from you,

10:44 do you use any tools like this to speed

10:47 up your workflow?

10:48 Leave a comment down

10:48 below and let me know.

10:49 Well, that's all I had for you for this video.

10:51 Thank you very much for watching.

10:52 As always, keep coding.

10:54 dictating.