शैडोइंग अभ्यास: They want mediocre developers... - वीडियो के साथ अंग्रेजी बोलना सीखें

I miss the good old days when technical debt

and student debt were the only debts we had to worry about in the software world.

Looking back at it, technical debt was actually quite fun.

We would take shortcuts to move faster,

ignore best practices to meet a deadline,

or delay updating a library because we were in no mood to work through the side effects.

The worst that could've happened was

that you would end up paying for it later through increased complexity in systems that nobody wanted to touch.

But ultimately, this increased complexity would increase your job security because

if If you were the only person who understood a messy part of the system,

you would suddenly become very hard to replace.

This is why for the past 20 years,

software engineering has been a pretty nice gig,

with a lot of perks and more than decent paychecks.

And I'm not making this video to tell you that all these are going to end,

but I want to discuss a new emerging kind of debt which will actually make our lives much more difficult.

Adios Money recently posted about a new concept called Comprehension Debt

and in this Monday morning review will look at the importance of being able to think

and solve problems in the generated code era.

Comprehension debt is the growing gap between how much code exists in your system

and how much of it anyone on the team can genuinely explain.

At first, this doesn't look like a problem,

since most AI-generated code looks really clean at first and it might look like all the best practices are being followed.

But underneath that, something subtle starts to shift because the understanding of the system begins to slowly degrade.

In a recent study, engineers were asked to learn a new library.

Some of them used AI to generate solutions,

others used it more as a guide to explore concepts and ask questions.

Both groups completed the task in roughly the same time,

so from a productivity perspective, everything looked identical.

But when they were tested afterwards,

the group that relied on generation performed significantly worse,

especially when it came to debugging.

And debugging is where things become very real very quickly.

If you've been involved in software for a while,

you probably know that reading and understanding code at a deep level is one of the most difficult parts of the job.

I'm not talking here about going through some random syntax filled with for loops and if statements.

I'm talking about complex implementations,

which have interactions, implicit assumptions,

and behavior that only makes sense once you follow the entire execution path end-to-end.

This part is so difficult that you'll go through code you wrote yourself six months ago

and you'd still be confused trying to reconstruct what you were thinking at the time.

When everything works, speed feels like the most important metric.

The moment something breaks, the only thing that matters is whether someone can explain what the system is doing and why.

And, if I have issues understanding my own code from six months ago

and I can proudly call myself an average developer with a weird accent and a copied YouTube channel,

I expect

that most of us will have a lot of issues understanding thousands of lines of code generated in a couple of minutes.

For a long time, writing code and reviewing it acted as a natural bottleneck in development.

It slowed teams down, but it also forced engineers to read code written by others,

to question decisions, and to build a shared understanding of how the system behaved.

The process was one of the main ways knowledge spread across a team.

With AI-generated code, the volume changes the nature of that interaction.

Code is produced faster than it can be meaningfully reviewed,

and because it is usually clean and well-structured,

it gives off all the signals that historically indicated quality.

And the AI labs are telling you to ignore the problem.

You spend your tokens generating code,

and then they provide you the tools to review that generated code for the convenient price of $15-25 for each review.

In other words, the token sellers just want you to buy more tokens,

ideally hundreds of thousands of dollars worth of tokens.

I keep trying to avoid highlighting the absurdity of some of these CEOs' comments,

but they are making it very tough for me to ignore them.

screams trying to inflate the bubble more than focusing on quantity over quality.

Sure, you could easily spin up Todo apps in an endless loop to burn through $250,000 in tokens.

That would be completely pointless and will consume a few more million gallons of water,

but at least the billionaire Jensen Huang with his geeky leather

jackets will be happy to see his Nvidia stock keep going through the roof.

But I digress.

Back to the topic of comprehension debt,

there's also a more subtle failure mode that starts to show up in AI-heavy workflows.

When an implementation changes and the associated tests are updated to match,

the conversation quietly shifts.

Instead of asking whether the code behaves correctly,

you now have to ask whether the behavior itself is correct.

The test passing no longer answered the most important question.

As we discussed in last Monday's episode,

another common reaction is to rely more on specifications.

You define the behavior in detail,

review the specification carefully, and let AI handle the implementation.

On paper, this creates a clean separation between intent and execution.

In practice, translating a specification into working software involves a large number of implicit decisions that rarely get captured in full.

Data structures, edge cases, error handling,

and performance trade-offs all shape the final behavior in ways that are difficult to fully describe up front.

This year's DORA report also highlights this new emerging comprehension crisis.

In teams where engineering practices are already solid,

where workflows are clear, and where people have a strong mental model of the system they are working on,

AI tends to make everything move faster in a way that actually compounds quality.

Developers spend less time on repetitive tasks,

decisions are made with more context,

and the system evolves in a relatively controlled way.

But in teams where things are already messy,

where architecture has grown organically without much discipline,

and where understanding is fragmented across individuals,

AI scales all of those issues at the same time.

The same acceleration that helps good systems becomes a multiplier for bad ones,

and the difference between the two becomes much more visible.

What's interesting is that the report states

that around 90% of developers are already using AI in their work and more than 80% believe it improves their productivity.

But at the same time,

100

a significant portion of them openly admit that they do not fully trust the code that AI produces.

101

So you end up in a situation where developers rely on a tool daily,

102

move faster because of it,

103

and still carry a constant layer of uncertainty about what is actually being generated.

104

The DORA data also highlights something that reinforces this dynamic even further.

105

AI adoption correlates with higher throughput,

106

meaning teams are shipping more changes and moving faster through their backlog.

107

At the same time, delivery instability increases,

108

which means those changes introduce more issues into the system.

109

So basically, you move faster,

110

create more issues, you fix those issues faster,

111

and create even more issues.

112

Your KPIs go through the roof,

113

management has nice graphics to look at,

114

clients are happy to see results,

115

and token sellers are making a lot of money.

116

Everybody is happy and Jensen Huang gets richer.

117

If I'm being honest, I spent the past couple of months worrying about the future of software engineering.

118

AI code generation is far from perfect,

119

it has improved a lot in the past couple of months.

120

However, something feels off.

121

The AI labs are pushing these coding agents hard and we have to pay very little to use their models.

122

However, their computing costs are huge and they are very far from running a profitable business.

123

This is clearly the customer acquisition phase,

124

where their pricing is so low you can't ignore them

125

and their agents are slowly becoming an integral part of the software workflow.

126

And I believe that this comprehension debt is exactly what they need to make sure you can't really walk away later.

127

This is not going to happen overnight,

128

but imagine a world 10 years from now where most of the software engineers have lost most of their critical thinking,

129

problem solving and debugging skills because they are used to delegating all of it.

130

At that point, Anthropic could increase the cost of their agents to thousands of dollars per month

131

and there would be little we could do because most of the skills are lost,

132

we have no idea what our code actually does

133

and the jobs pipeline was destroyed when we decided AI can do a better job than juniors.

शैडोइंग तकनीक क्या है?

शैडोइंग (Shadowing) एक विज्ञान-समर्थित भाषा सीखने की तकनीक है जो मूल रूप से पेशेवर दुभाषिया प्रशिक्षण के लिए विकसित की गई थी। विधि सरल लेकिन शक्तिशाली है: आप मूल अंग्रेज़ी ऑडियो सुनते हैं और तुरंत इसे ज़ोर से दोहराते हैं — जैसे वक्ता की छाया 1-2 सेकंड की देरी से। शोध से पता चलता है कि यह उच्चारण सटीकता, स्वर, लय, जुड़ी हुई ध्वनियाँ, सुनने की समझ और बोलने की प्रवाहशीलता में काफ़ी सुधार करता है।

शैडोइंग अभ्यास: They want mediocre developers... - वीडियो के साथ अंग्रेजी बोलना सीखें

इस पाठ के बारे में

मुख्य शब्दावली और वाक्यांश

अभ्यास के टिप्स

शैडोइंग तकनीक क्या है?