Turn Up the Gain

12 September 2026

An opinion piece based on my experience using coding models and working in software teams.

AI coding is an amplifier. Calling it an accelerant only tells half the story.

It makes writing code faster, obviously. But the more interesting question is what you're amplifying. Good judgement? A useful product? Or a confused organisation's ability to produce a mountain of expensive bollocks?

All three can benefit from faster typing. Only two are worth having.

More code was never the goal

You could produce a hundred times more code and still deliver nothing useful. That's why the multiplier, whatever it happens to be for a particular task, isn't the thing I care about. I care about how quickly I can turn an idea into something people can actually use.

In my experience, the useful part goes well beyond generating code. In a distributed system, an engineer might know one service inside out and have a passing understanding of the others. Give a capable model access to the relevant codebases and it can help trace a feature across those boundaries. It can bring together context that would otherwise mean digging through unfamiliar repositories or waiting for someone else's availability.

That doesn't make it omniscient. It does make it useful.

And the change over the past few months has been striking in my own work. Previously, I spent a lot of time turning tickets into comprehensive technical plans before handing them over. I had to spell out the approach, the constraints and the implementation details to get a solution I was happy with.

Increasingly, I can start with a loosely defined feature idea and get a competent proposed implementation. There is still judgement involved in deciding whether it belongs in the product, whether the assumptions make sense and whether the result works. But I am spending less time explaining every step of how to write it.

That's a substantial change to the job.

Some of this looks like grief

I understand enjoying the coding itself. Staying up late, being stuck on something difficult, finally cracking it and watching it work. There is real satisfaction in that.

But the user wasn't waiting for me to have a breakthrough. They were waiting for the thing to work.

My suspicion is that some of the hostility towards AI coding comes from that distinction. If the part of the job you love most is personally solving the puzzle, having a machine do more of it can feel like losing something. If your main satisfaction comes from getting useful software into people's hands, the same development can feel liberating.

Those aren't mutually exclusive. You can enjoy both. But when someone dismisses all AI-assisted code as slop, regardless of what it does or how it's been checked, I struggle to hear a serious engineering argument.

It sounds like denial. A bit of Luddism dressed up as concern for standards.

Perhaps grief is the better description. Grief for a world where every line was manually typed, hard-won technical knowledge carried a particular status, and struggling through the implementation was an unavoidable part of creating something.

I can sympathise with missing that world. I don't think missing it is a good reason to pretend the tools don't work.

Humans have been writing slop for years

I've worked at enough companies and read enough code to have very little patience for the idea that human authorship is a mark of quality.

I've seen plenty of manually written shit. Needless complexity. Fragile integrations. Features that technically satisfy a ticket while completely missing the point. Entire systems whose maintenance costs bear no sensible relationship to what they do.

In the work I'm seeing now, I don't find human-written code inherently better than AI-assisted code. The distinction that matters to me is whether someone competent has understood, reviewed and taken responsibility for the result.

The security discussion deserves the same scrutiny. “An LLM deleted my production database” is a headline, not enough information to understand a failure. What access did it have? What was it asked to do? What executed the change? Where was the review?

I wouldn't assume every account is fabricated, or that a destructive action must have been explicitly requested. But treating a dramatic anecdote as proof that the whole approach is worthless is garbage. Show the circumstances and the failure. Then we can have a useful discussion about it.

The organisation gets amplified too

A bad engineer now has a way to produce bad solutions much faster. A good engineer has a way to explore, build and check useful solutions with less implementation work.

That's the argument for amplification.

And engineers work at the coalface of organisations. What reaches us is shaped by the quality of the decisions above and around us. Confused priorities, contradictory requirements and an inability to decide what matters don't disappear because implementation gets cheaper.

My expectation is that dysfunctional organisations will use some of this new capacity to produce more dysfunction. More half-considered features. More systems nobody needed. More changes driven by whoever spoke last in a meeting.

A model can't decide on an organisation's behalf what is worth doing. Someone still has to understand the problem and own the outcome.

I'm here for what I can build

I want a shorter distance between having an idea and putting a working solution in front of someone who needs it. If AI helps me do that, I'm using it. I don't need to personally type every line to feel that the work counts.

Writing this on 12 September 2026, the pace of change in my own workflow has made arguments from even a few months ago feel dated. At times, it feels like years of progress are being compressed into months. I don't know whether that pace will hold, or where the next few months will leave us.

For now, I'm excited by what I can build.

What this means for us beyond the work is harder to sit with. We build identities around being useful, knowing things and being able to do what others can't. I can welcome a tool that takes more of the work off my hands and still wonder what happens as that list keeps growing.

What this means existentially for humans remains to be seen.