تدريب Shadowing: I Spent 1.5 Billion AI Tokens on Absolute Slop - تعلم التحدث بالإنجليزية مع YouTube
B2
⏸ متوقف مؤقتاً
I spent 1.5 billion tokens on this project and put it through GPT-56-SOL on high.
251 جمل
إذا كانت الجمل قصيرة أو طويلة جدًا، انقر على Edit لتعديلها.
1
I spent 1.5 billion tokens on this project and put it through GPT-56-SOL on high.
2
I did have it spawn extra workers on high to try and knock this out in a couple days.
3
Overall, about 24 hours of consistent compute runs.
4
There were some times where it stopped because of usage errors or a reset needed to happen.
5
But I did use, I think, four resets of the $100 plan and just a ton.
6
I got lucky in opening.
7
I had a reset as well.
8
So I was able to use a ton of inference.
9
I don't know what that 1.5 billion tokens will equate to in dollars, but I would assume probably around $10,000, maybe more.
10
And I wanted to say this was terrible.
11
It made slop.
12
and I wanted to kind of showcase it because what I had was a .NET project.
13
Most people know my Windows utility that's written in PowerShell
14
and my .NET basis I've built over the last year or two off and on just kind of hacking away at it.
15
It had roughly, I think 5,000 lines of code and I wanted to expand it out so it has feature parity.
16
An idea behind Win OneShot was just to have something kind of on deck in case something happened with PowerShell, I could easily translate everything to .NET.
17
So all the functions I built was like changing registry edits.
18
Instead of doing a set item property, I could actually just force that through the .NET libraries instead of using that section of the code.
19
So I could always fix windows essentially.
20
So it's always been kind of like my fail safe or a possibility of maybe doing a standalone.
21
But I wanted to see what AI could do with it.
22
So I already had the scaffolding.
23
I already had the structure.
24
I just needed to feed it a roadmap, a spec sheet, and then have it build out of tasks for it to run.
25
And I thought, hey, let's split this up with workers.
26
And I got this whole idea because I saw something online about Bun being rewritten and ported from Zig to Rust.
27
And I was like, well, shit, if it can do that, surely 5.6 Sol High could do this.
28
This is nothing compared to that kind of port.
29
And sure, they use Fable 5, but I was like, well, 5, 6 old should be good enough.
30
And I want to say, I think we're to the spot where it's not really about the models.
31
It's about you, your ideas, your skill, your software development ability.
32
And I think this is always going to be a thing
33
that you just I don't see a way around this where you have to think through the problems, the structure, and AI just can't think for you.
34
You can tell it and direct it, but my directions was have feature parity.
35
Here's my parity project with PowerShell script.
36
Here's the scaffolding where it already has the functions in place.
37
Just add the missing functionality, whether that was, I think win 11 creator where it creates an ISO or,
38
uh, the app X removal that was just added in this last update to when you till.
39
And I was like, I, that should be relatively easy to do maybe five, 10,000 lines of code.
40
And at the end of the day, we ended up with 169,310 lines of code that was added.
41
Holy shnikes, Batman.
42
That is an absolute ton of code.
43
What does that mean when you have, well, we'll just call it 170,000 lines of code that was changed.
44
It means the project's no longer years.
45
You don't know what's up.
46
You don't know what's down.
47
When you look at the code base, it is no longer maintainable.
48
You just have so much stuff, you're better off just starting from scratch again.
49
In this instance, I'm going to show you what it created, and then I'm going to show you what it had to work off of,
50
because they're just two vastly different things here.
51
What it created was this.
52
Pretty good project on launch.
53
I did notice right away on launch, it launched considerably slower.
54
Well, to be expected with over six figures of lines added to the project.
55
But with a built binary, it's going to be menial to the end user.
56
Compute these days is so cheap.
57
And most people, you know, most software developers, even before this, had bloated projects, and it really didn't affect performance that much.
58
So I looked at it, I was like, yeah, all right, that's good.
59
Now, when it first spit out, it was all one single column.
60
So of course, a couple of reiterations needed to be made.
61
It was like, all right, let me give this an honest try.
62
Let's split it up.
63
Let's match this.
64
Did okay job after all of that.
65
We still have like a category load fail element not found down here.
66
Trying to track that down.
67
I was like, okay, that needs to be reworked a bit, but it does work.
68
And then for the tweaks page, I was like, okay, Let's see what happens there.
69
It tries to inspect all the tweaks and then tell you what the current state is.
70
Sometimes it can do this.
71
Sometimes it can't.
72
It just depends on the tweak.
73
But to try and go through and maintain this, let's say I was to release this on the world.
74
Well, people are going to go, hey, the acrylic blur is applied, but it says not applied.
75
And you're like, well, I, okay.
76
And then there'd just be tons of issues like
77
that with this iteration and you're relying on an LLM prompt to fix this.
78
At this point, the code is pretty much unreadable, unmaintainable, and that's a problem, especially if you're charging money for something like this.
79
And this is where I think a lot of people are gonna get in trouble.
80
So this is what happened with the tweaks, configuration.
81
I had to kind of move this around.
82
Initially, it made this check boxes.
83
Then I was like, no, those need to be buttons.
84
That was fixed the updates just didn't work at all i
85
think they still don't work something about state infection i was like okay well i'll go ahead
86
and click this hopefully it doesn't blow up my system it kind of works on a lot of things
87
but it's close but not and i think the fallacy here for someone
88
that never done programming before i even saw i think it
89
was a meme online where someone was like the new state of development, you know, idea, first build,
90
and then the last 10% is this massive chart where the last 10% took up more than 60, 70% of the time.
91
And honestly, this is kind of how it's always been to try and polish that last 10%.
92
The problem is, when you have so much tech debt from all this AI and iteration for big projects, it's impossible to read the code.
93
So that last 10%, if AI can't fix it, you're kind of screwed.
94
It just doesn't, you're basically relying on it 100% with no safety net.
95
So you can't go to market with something like this.
96
It's just really important that people have that or know that up front.
97
And you might be thinking, Titus, just use a different model.
98
Models will not save you from just plain bad foundation.
99
You have to have maintainable code.
100
It has to be readable.
101
And then you could pass it off.
102
I've also heard horror stories of like CEOs
103
or someone from the C-suite making something or vibing something up and then handing it off to a development staff.
104
That is the opposite of how that should go.
105
It should be the development staff makes a mock and then gives it to the C-suite to look at, not the other way around.
106
Where we're failing with AI is, or LLMs, whatever the hell you want to call it,
107
is normal people still can't make software let's just be real here let's i'm going to say
108
that and scream
109
that from the mountaintops you have to know the foundations you have to think through the projects
110
and understand what you're getting into even when i gave it proper instruction
111
and just let it run wild it completely screwed everything up now it's on git
112
so it's not a big deal let's get to and rewind to show you what I gave it, where we were at using Git.
113
And if you don't understand Git, you should not be doing any of this.
114
Git is the very first thing And it's so important to understand.
115
I can't say that enough.
116
It will save your bacon so much because this is 170 lines, 170,000 lines, thousands of dollars.
117
Let's say if it's a company, it would have been probably tens of thousands of dollars it would have spent on.
118
And now it has to be thrown away.
119
So I'm going to go directly to this private repo because I didn't want this to be public.
120
If you're interested in the code, I could make it public so someone can look over it.
121
But I'm going to go back to before this all started.
122
So here's the commit history for win one shot.
123
I'm going to just grab this because I haven't pushed any of these changes.
124
And we can just discard pretty much all of it.
125
So the commands get reset, dash dash hard, and this.
126
It's going to reset everything from here.
127
So once we do this, say goodbye to the project.
128
And then you're going to see how the project changes.
129
Look at the current structure.
130
And here we go with the git reset.
131
So now we're back to this.
132
And if we look at what we have, we have a sources, test tools, all this.
133
These are untracked.
134
We could actually get rid of these.
135
We're going to just do a git clean FD.
136
It's going to push out all these tests, everything that was wired up from it.
137
So now we're working off of a clean from May.
138
I'm going to do a clean and then the build really fast.
139
It looks like there was a bunch of artifacts left over.
140
So essentially we can use LLMs.
141
I'm not saying I'm not anti-LLM.
142
We can use it to troubleshoot this.
143
Like, why didn't this reset happen?
144
And it said there's a bunch of leftover sources in the OBJ directory.
145
And then we looked through, it was still looking and referencing some of those sources from the source directory that was created by the LLM,
146
which was causing the errors.
147
So now that that's done, it probably will wipe out that.
148
We're going to clean those generated paths and restore and build.
149
And now we should be able to build the full release.
150
Now it's successful.
151
Now we can do a .NET run.
152
And here is the project as it was before the LLM took over.
153
Obviously quite a bit simpler, obviously not so much bloat.
154
I did kind of like the sidebar.
155
So while this seems like a massive waste, it did give me some ideas.
156
Like I really did like that sidebar pop out.
157
I might actually take that, implement it.
158
That's actually a pretty good thing.
159
Obviously still have to work on the config page.
160
Updates page, I was kind of thinking through a few ideas here for like default security and disabling.
161
This all kind of has a few different changes here.
162
Obviously it's not reading my Windows version.
163
This is actually Windows 11, not Windows 10.
164
So this is not without its own bugs, because it was hand coded by me.
165
But you got to look at the code at the end of the day, like this is what was there.
166
It's so much easier to just kind of right there's my views.
167
I don't have five or six different sub projects.
168
It's very maintainable in this structure.
169
So I look at the spec sheet and all the stuff that the LLM built over that.
170
And the problem isn't the models.
171
The problem is me, my ideas, how I instructed it, and then just letting it run wild without proper checkpoints.
172
Bigger projects need those checkpoints.
173
You need to understand before it just completely goes off the reservation
174
and then makes 150,000 lines of code when it should be 10,000 lines of code.
175
Well, I don't even know how many lines of code is in this entire project.
176
Let's find out.
177
Here we go.
178
Roughly about 10,000 lines of code.
179
So you see the difference there.
180
And to manipulate this to be better,
181
like removing the checkboxes, making those icons and those cards like I did in WinUtil, you can do all of that here.
182
And the lines of codes change for that kind of implementation.
183
We're talking maybe 100, 200 lines of code.
184
Maybe it's in that ballpark.
185
And it's so important to understand you can't just let these things run wild.
186
So what this should be is go from here and say, hey, let's improve the cards.
187
Instead of check boxes, let's add cards with icons.
188
And then slowly go through and change the things that you need changed.
189
And then go from there.
190
That's probably the biggest takeaway from this is I think we're in an era where you have these extremes.
191
You have people that don't have any business making software, trying to vibe up something and then handing it off to a professional.
192
That should never happen.
193
Professionals need to make it using these tools that can make it faster, but they are the ones to commit that structure,
194
do those checkpoints, make sure it's doing it properly and making sure it's maintainable in the future. So important.
195
And then the second thing is There's also people, the tech bros out there that just say, hey, make me an app, go.
196
And then just sit there and reiterate, spending billions upon billions of tokens to make something that kind of works.
197
But as soon as they go to market with this, it's going to fall apart.
198
And then it's not going to be maintainable.
199
And then all of a sudden they sold something that just doesn't work at all.
200
And they're just like, oh crap, that was false advertising.
201
And it just runs down this entire terrible cycle.
202
So out of my last video, 10 projects for AI to do that I've done, I wanted to make a video today saying,
203
here's where AI failed miserably.
204
And I don't blame AI for this.
205
I blame me for it.
206
I should have checked in.
207
It should have been a phased approach where after each phase, I'm checking the code and going, nope, stop, stop.
208
That's just bad.
209
No, we don't need to do that.
210
We don't need to write 20,000 lines of pester tests for something that should never even be tested.
211
We don't need to do all these other things
212
that it ended up doing because it was just not the right way of doing it.
213
So it ended up implementing it incorrectly.
214
Not that it didn't work, but it's not done how I would want it being done.
215
So hopefully this all makes sense to somebody.
216
Don't one feel like, hey, you're a normal person and then you're just gonna be able to develop software that you never developed before.
217
Small software, sure.
218
Dashboards, go for it.
219
I think normal folks should do that all day long.
220
I do it.
221
It's great.
222
Like just look back.
223
Websites, wow.
224
Fantastic at websites.
225
But when it comes to big projects, I very much caution you.
226
You need to think of it as a pair programmer code review.
227
It's fantastic at those things, fixing really like nuanced bugs.
228
Like why the hell is it doing this weird glitch right here?
229
Tell me why.
230
And then it'll spit out and go, well, you idiot, you're utilizing it this way improperly.
231
And you should be pulling this library.
232
And you'd be like, Oh yeah, that, that makes sense.
233
That's where it's at with great, like big projects.
234
It's fantastic.
235
It's a great tool for it, but it will never replace the programmer.
236
It will never replace the developer.
237
I think it'll cut down or make that developer a lot faster
238
and it'll make them be able to do a lot more with less, but a replacement is not.
239
And that's what this video was.
240
This is just kind of a state of 5.6 Soul.
241
You could say, hey, Fable 5 would have done a better job.
242
Maybe, probably.
243
I mean, I couldn't imagine doing a worse job than this.
244
But I don't think it's a model issue.
245
I think it's a me issue, an ideas issue, a foundation issue, an instruction issue.
246
You have to think through these projects.
247
You have to think through these problems before it encounters them.
248
So it sets it up with a structure that can be maintainable and usable and verifiable.
249
It's so important.
250
anyways I hope this helps somebody with
251
that let me know your thoughts down in the comment section and I'll see you in the next one
تنزيل التطبيق
Everything you need to speak fluently
AI PronunciationScore every sentence
IPA PracticeMaster every sound
VocabularyBuild your word bank
Vocab GameLearn while playing
حول هذا الدرس
أنت تتدرب على اللغة الإنجليزية باستخدام "I Spent 1.5 Billion AI Tokens on Absolute Slop" مع تقنية الـ Shadowing.
ما هي تقنية التظليل الصوتي؟
التظليل الصوتي (Shadowing) تقنية تعلم لغة مدعومة علمياً، طُورت أصلاً لتدريب المترجمين الفوريين المحترفين. الطريقة بسيطة لكنها قوية: تستمع لصوت إنجليزي أصلي وتكرره فوراً بصوت عالٍ — كظل يتبع المتحدث بتأخير 1-2 ثانية. تُظهر الأبحاث تحسناً كبيراً في دقة النطق والتنغيم والإيقاع وربط الأصوات والاستماع والطلاقة.