zakk

software engineer building products with agents. dad x2.

[zakk @hexednobility]

3 years ago, many predicted massive SWE productivity gains from LLMs. As a sr swe at big tech, I wrote this post to explain why they were wrong.

I was largely right. Even now there is little evidence in 2026 the industry has writ-large seen 30% productivity gains.

But that's changing, fast. It’s now abundantly clear - software development and human computer interaction its self is amidst a profound transformation.

And I've never been more excited.

I’m reviving this account to talk LLMs, engineering and second order effects.

Follow for the signal.

[zakk @hexednobility]

.@bchesky and @Jason recently claimed that AI will boost software developers' productivity by ~30% across industry.

As a developer and IC at a major tech company, let me debunk that.

1. Software developers don't spend all their time coding.

Typically, coding occupies less than half of a developer's time. The rest goes to parallelization costs (meetings, documentation, alignment, planning) and operations (monitoring metrics, configuring alarms/pipelines, debugging). This limits the time that can be saved. Even with improved coding efficiency, the overall productivity gain would be modest.

2. AI might expedite coding but not code reviews.

Every competent tech company reviews all code before production, often by multiple devs. If I use AI-generated logic, I must first understand it such that I can defend it in code reviews. Review time for others remains unchanged. Some will choose to ship AI-written code without understanding which will lead to inevitable disaster.

3. AI models must be fine-tuned for in-house frameworks, libraries, and practices.

While the base models can write code using open-source frameworks, they are useless when it comes to in-house development. Overcoming this issue requires fine-tuning which demands capital and—ironically—more software developers. Additionally, fine-tuning AI models to a company's internal software and documentation isn't a one-time task. As frameworks, libraries, and best practices evolve, the AI model requires continuous updates and fine-tuning, consuming significant resources and development time.

4. AI excels at teaching, not executing. Experienced developers will see far less value.

Learning a new tech stack is where AI is most impactful. ChatGPT is the most patient and knowledgeable teacher you could ask for. For a tech stack I've mastered over the years though, AI offers minimal value. It might help with boilerplate, but I'll rely on my own expertise for critical code.

While I see significant value in AI as a software development tool, its immediate impact is overestimated, particularly in large tech companies.

Execs, VCs, and others without boots-on-the-ground engineering experience are swayed by demos and Twitter thread bois. It's exciting and novel, but the proof isn't in the pudding.

Startups might see more impact. They lack in-house frameworks and prioritize shipping over quality. They might hire less experienced developers who benefit from AI as a learning tool.

But a 30% boost? No way. Not this year.

[zakk @hexednobility]

@charliermarsh Much more interesting.

How much code is written by a keyboard is not an interesting statistic, LLM-written is the same in this case.

[zakk @hexednobility]

@charliermarsh “no one reviews code shipped to prod anymore” is a much more provocative and interesting statement

[zakk @hexednobility]

@rstchristopher The term “AGI” makes it difficult to have a rational discussion about this.

I don’t believe we have made more than an inch of progress toward AGI but I do believe in LLMs ability to supercharge productivity in a way that is very likely undervalued right now.

[zakk @hexednobility]

@vrexec Eating some labor sure but we are still greatly constrained by people that have your experience.

You have to know what you want and you have to know what output is good.

[zakk @hexednobility]

There is some gray area between vibe coding and reviewing every line of code too.

I have an agent orchestrator app I built for my self (web UI + execution engine). I haven’t read a line of the code.

Yet I dictated the architecture and code base design. I can competently address bugs and evolve the system with new features.

It’s not especially elegant or performant but as a piece of personal software I’m really happy with this in-between space.

[zakk @hexednobility]

@dexhorthy Speaking my language. The nature of LLMs makes it seem easy but regardless of how “intelligent” the models become there is clearly skill required.

I wrote about the core skill of “Knowing what you want” here

[zakk @hexednobility]

https://t.co/sfqc6tQwfn

[zakk @hexednobility]

@housecor In my experience, clearing the context then replacing with precisely the narrow set of context (usually a markdown file) needed for the next task is always better

Its just a question of whether its worth the time

[zakk @hexednobility]

@karakhanyanS No, the code still matters because software is not static.

Bad, redundant, slow code can work and achieve your short term goal and still make for a bad product that cannot evolve.

[zakk @hexednobility]

LLMs will become more powerful but they won’t magically become easier to use as advancements are made

Like any tech, applying LLMs is a skill that takes time and practice.

Wrote more in depth about how to be effective with LLMs here

[zakk @hexednobility]

https://t.co/sfqc6tQwfn

[zakk @hexednobility]

@housecor Agreed that beta/staging environments are fairly useless though I would add

Ad-hoc prod environments too

Every dev and agent needs to be able to spin up a prod-like environment to run manual and automated tests

[zakk @hexednobility]

@TechEmails This is like when Marlboro lights were the industry leading safest cigarettes for youths

[zakk @hexednobility]

@mattpocockuk “AI is bad at design” is actually just “non-designer software engineers are bad at design”.

Skills can help fill in gaps but you still have to know what you want and be able to precisely describe it.

I wrote about this here https://t.co/zHVjBWaeUx

[zakk @hexednobility]

https://t.co/sfqc6tQwfn

[zakk @hexednobility]

@lucas__crespo “AI-native mac” please leave this garbage to windows, mac did nothing wrong

[zakk @hexednobility]

@kupsas101 @omooretweets @cursor_ai Now I wouldn’t be thrilled about it but at the end of the day I would accept my pay being tripled

[zakk @hexednobility]

@BoringBiz_ You have to know what you want it to produce. AI is useful for doing the legwork but immediately falls apart if you want it to do something strategic.

Wrote a bit about how to use LLMs effectively here

https://t.co/zHVjBWaeUx

[zakk @hexednobility]

https://t.co/sfqc6tQwfn

[zakk @hexednobility]

@stevekrouse @tempo Wow surpassing @VicVijayakumar in the native AI fax service market

[zakk @hexednobility]

This app is undefeated when the wife rattles off a marathon list of tasks, packing list and all of her hopes and dreams in one breath

[zakk @hexednobility]

@thorstenball It’s all about how you use it. Lots of discourse on here from unserious and unskilled folks to start with.

The experienced engineers with strong foundation are leveling up all around me.

[zakk @hexednobility]

@Chase_Swartz3 As a parent I find it obnoxious when parents say things like this. Of course being a parent is morally superior but you sound like a jerk