Практика Shadowing: JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro) - Изучайте разговорный английский по видео

Okay, so what you're seeing on the screen is an AI harness building a project of my choice.

I'm going to talk a little bit about project in some time.

But what is interesting in this is that this is the DeepSeek V4 Pro, number one.

This is extremely cheap, number two.

And this has something that I'm really liking, which is the taste feature, which is basically how I like the code to be done in the first place.

So let's talk about all of this.

The AI harness I just showed you is called as Command Code, which is a CLI tool acting as a coding agent that can write,

debug, fix, do whatever you want with your code base.

Now a super interesting part about command code is

that it has this $1 per month plan that has this $40 of DeepSeek v4 pro included in it.

And this is actually really good because if you look at DeepSeek's pricing, even the latest one, their API pricing at least, you can see that this is a model which is extremely cheap.

This one what I'm showing you is DeepSeek v4 flash

and v4 pro is also a very cheap model

while meeting a lot of expectations of frontier models like opus

and gp and the way you set up this one dollar plan

that has 40 dollars of usage

which is hundreds of millions of tokens is also super simple

all you have to do is get started with command code

i will give you all the links in the description you can see

that this is just my today's past hour usage right i have spent about 1 million tokens on deeps v4

and this has only consumed nine cents out of my hundred dollar budget

but anyway on the instructions on getting started it's super simple

you write npm install g command code inside your terminal you do a cmd login

which would use your logged in account right now

which you have subscribed with one dollar

and then you just write cmd to initiate the chat now

while command code and deep seek are working together i want to show an interesting file

that i found which gets created in your folder

if you go to this command code folder in your working directory and if you look at this taste folder

you're going to see that there's a file called taste.md.

Now this taste.md file that I'm just showing you, it's not a file that I have.

It's a file that command code is continuously learning based on how I'm instructing it.

And it's also not using everything that I'm saying it.

It has a confidence score in front of it.

So this is actually something like building your own agent skill, right?

But instead of me writing everything manually inside that skill, it's using my usage of CLI tool itself in order to figure out what sort of best practices I put.

And then according to my conversation, it can tweak these confidence scores based on what I like and what I don't like.

For example, personally, I like to configure Prettier with tabs, no spaces, no semicolons.

And it got it pretty well based on this confidence score.

I like using PNPM.

I like using Zod.

I mentioned a lot of these things in a lot of my videos, right?

So there is nothing new that you're discovering.

And for this specific project, I also asked it that use React for frontend and Cloudflare workers for backend, which it has given a confidence score of 0.7, which is pretty cool because in future,

when I ask it to add something new or, you know, make a new feature available or make a new endpoint, this taste.empty file can be basically just used directly.

And if you're working in a team, this file can also be commented on Git and your whole team can use this.

Now the project that I'm working on is basically a YouTube monitor, which I'll talk a little bit more about in an upcoming video because I'm still developing this.

But I just wanted to take a spin on how this can be done using DeepSeek V4 as a model.

This is probably going to be a small mini SaaS that I also launch in the coming time.

So stay tuned for that.

Once I have this ready, I'm going to update this here.

But one thing that I generally wanted to understand is that how DeepSeek is able to work much better

in this specific harness compared to generally

if you are using it with any other harness like cloud code

or even open code and the reason

that deep seek is much more optimized in command code is specifically

because this specific harness has been fine-tuned to work properly with deep

so i came across this post where ehemad who's also the founder of command code mentioned

that they have made deep seek v4 pro beat opus 4.7

six on ten times in their internal evals by auto repairing many of the quirks in the tool calling

and the overall explanation is slightly longer.

The core idea is performance, the raw performance of DeepSeek.

If you just consider DeepSeek as an engine, the raw performance and the intelligence of DeepSeek is actually good.

It's the harness around it which makes or breaks the...

So he says, I've been thinking about why open model bad at tool calling is almost always a harness problem, not a model problem.

See again, just to summarize before what we talked in the past as well.

AI models are things like Opus 4.7, GPT 5.5, DeepSeq V4, and so on and so forth, right?

These models are just raw intelligence.

They cannot do a lot on their own.

They can maybe answer simple questions.

You can ask them some academic questions, something that you want and they'll just spit out the text back.

What you often need is an AI harness on top of this.

So an AI harness is basically a piece of software that gives this AI model tool calls, right?

And tool calls are basically just the inversion of controls ai model will ask the harness itself

when it wants to access something for example let's say you

asked it create an html project right in order to create

an html project there are a bunch of steps first it needs to think about it

which is just the model part once it has thought about

it it needs to write file this writing the file is the harness part right

because ai cannot technically create a file on your needs to make a tool call

which would be done by the harness itself.

And I'm giving you a very simple example, but this tool calling can get very complex as well, depending on what sort of tools you are giving it, whether that's web search, executing arbitrary bash commands,

you know, sandboxing those commands.

So there's a permission system.

So all of this can get super complex.

And the idea is that the more your tool calls fail, the more AI tries to repair them on its own, the more tokens are to waste.

And the more the tokens are wasted, the less the intelligent model becomes.

We have talked about it in the past as well, that above like that's my heuristic i don't exactly know what is the actual number

or if there is one but i feel like above 100k tokens

or maybe like at maximum 150k tokens all of these models

which have like larger context windows sort of become stupid right

it they start sort of like start to lose a little

bit of context they'll start to make a little bit of mistakes here

and there so the ideal world is

that where you can get an intelligent model to do an important task in less number of tokens

that means these tool calls should ideally not fail right

because the more they fail yes you can feed it back into the model itself

and it'll repair but it's just gonna be a little messy

because you are now wasting more tokens right and

100

that is what fundamentally i feel they have fixed command code

101

especially for deep sea you can see he mentions the failure

102

modes aren't random they are a small finite compositional set across

103

deep flash pro glm quen the same four mistakes repeat of

104

almost exactly sending null for an optional field instead of omitting

105

it emitting a b as a json string instead of an

106

actual array wrapping the single argument as this where the scheme expected an array now these are not something

107

that models would not be able to on their own

108

but the idea is that

109

if your hardness can handle this in a way then you're

110

just going to save a lot of tokens doing two

111

and four and i saw that there might also be open sourcing command code

112

so you can actually check this layer of validation yourself as well once they do

113

that their github repository right now does not list it as

114

an open source project it just is there for opening issues

115

or if you want to request any features or

116

if you want to generally learn about it you can do

117

that so deep seek v4 i feel is sort of like a model

118

that you can use with command code like harness for very cheap

119

and maybe run a competition internally who gets to burn 10 million tokens as fast as possible.

120

And the funny thing is that even if you burn 10 million tokens, according to my dashboard, you will still be under $1 in cost.

121

And this does not even include the Go plan.

122

I'm actually on credits right now.

123

Your Go plan that you would have probably have even more usage because remember that their Go plan starts at $1 a month, but it includes $40 of DB for usage.

124

And they are in fact running this offer until May 31st.

125

So now is actually the perfect time to get this plan get this harness

126

and see for yourself how productive you can be with an ai as smart as deep seek

127

and a harness which is powerful enough to handle that

128

so yeah this is a pretty fun way of using deep seek v4 pro for very less cost

129

and this is one of those moments where you know you can use best of ai

130

but you don't have to break your bank last i remember this happening was in november last year

131

when you could buy a 20 claw subscription

132

and use opus 4.5 a lot and now i feel that you should give this a try, especially if you're somebody who has never used an intelligent model before and if you're only relying on free models.

133

So that's all for this video.

134

You will find all the links in the description.

135

Once again, thanks to Command Code for sponsoring this video.

136

I'm gonna see you in the next one very soon.

137

If you're still watching, make sure you leave a comment, I watched till the end below, to tell me that you were still here and let me know what do you think about the video.

Что такое техника Shadowing?

Shadowing — это научно обоснованная техника изучения языка, изначально разработанная для подготовки профессиональных переводчиков и популяризированная полиглотом доктором Александром Аргуэльесом. Метод прост, но эффективен: вы слушаете аудио на английском от носителей языка и немедленно повторяете вслух — как тень, следующая за говорящим с задержкой в 1–2 секунды. В отличие от пассивного прослушивания или грамматических упражнений, Shadowing заставляет мозг и мышцы рта одновременно обрабатывать и воспроизводить реальные речевые паттерны. Исследования показывают, что это значительно улучшает точность произношения, интонацию, ритм, связную речь, понимание на слух и беглость речи — что делает его одним из самых эффективных методов для подготовки к IELTS Speaking и реального общения на английском.

Практика Shadowing: JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro) - Изучайте разговорный английский по видео

Контекст и фон

Топ-5 фраз для повседневного общения

Пошаговое руководство по шадовингу

Что такое техника Shadowing?