Coding is over. Get over it.

I’m a software engineer at JPMorgan Chase. I have more than 15 years of experience in building software and I used to love it. These are my own observations, not the firm’s position. JPMC is a very large company, and what I observe from my vantage point not be true at all. AI is pervasive company wide, and the how people work with AI varies a lot (we’re all learning to use it…). There are some guardrails in place that are very sane for a large bank, but I’m fortunate to be in a place that lets me experiment a lot. Costs are high, but the firm can afford it, and it may be a good investment to let people learn how to use the tools first, even if expensive, and think about costs later. It’s hard to know right now if the investment will pay dividends or not.

Return on Investment and Small Models #

With high costs, there is a high pressure as well to show they are worth it. The problem I have is that I don’t know how you’d measure the ROI even if you wanted to. Are engineering teams generating more value than they’re spending on AI? Hard to say. Would the same work have been cheaper without it? Probably. Would it be done much slower? Maybe. Could it have been done just as well with a cheaper model? To my knowledge nobody has actually run that experiment.

The performance of a single engineer has always been difficult to measure, and AI hasn’t changed that — it’s just added a new variable. Whenever I hear “we built this with AI in a 50% more efficient way,” I say “BS”. Unless that number is backed by something more rigorous than a number you made up, you don’t have a clue.

My current opinion is that AI is a major enabler. It has let me do things I genuinely couldn’t have done without it. While I’ve had extensive experience building apps, the infrastructure side was still a new field for me, particularly terraform. With AI agents I have more confidence in what they output and their assessment of the plan diff than what I would have alone, with my limited expertise. Whether that was worth the cost, I still don’t know and maybe never will.

A trend worth keeping up with are cheap, small models, which are getting genuinely good. GPT 5.6 Luna, with max reasoning, was a big surprise to me. I predict we’re not far from a point where AI inference cost becomes close to irrelevant for many use cases. A rounding error rather than a cost item. If that happens, I think the bigger revolution won’t be in coding at all. It’ll be everywhere else. All that can be automated, just by asking your computer…

That said, if anyone tells you with confidence what the future looks like from here, they’re guessing. Including me. Nobody actually knows.

The paradigm shift is hard to overstate #

I need to keep reminding myself of this: Model capabilities are evolving so fast that what was true last month may not be true today. It’s hard to keep up to date even when you work with it every day. Recently the head of Claude Code, Boris Cherny, said to delete all your AGENTS.md Claude.md files and just see what the models can do nowadays. I agree, many of the rules and nudges I used to need to give to my agents are no longer necessary. The pace of change seems to be accelerating and it is scary.

I considered myself an above-average coder but AI now codes better and faster than me. I say that with something close to grief. For 15 years I’ve been grinding and perfecting my craft towards excellence. That’s no longer necessary. My process for learning a new domain or technology used to be modeling it in code: build to learn. That’s not really how I learn anymore, and I miss it. I’m finding it hard to change but It no longer makes economic sense to write all the code by hand, so increasingly I don’t.

Not long ago I used to use AI as a great typing assistant, to make the speed of getting ideas from my head to the text editor. These days I only need to state the goals, and the app/script/answer just magically appears. I fear that my capacity to read and really understand code will atrophy since AI can now do research and answer any specific question I have in seconds, without any effort from me. I used to say that lazy developers are good developers because they will write automations for grunt work. Not sure if lazy is good these days because lazy developers will just be conduits for “AI Brains”.

I used to go deep on one problem at a time but now the expectation, implicit or explicit, is to parallelize across several threads at once. That kind of parallelization drains me in a way that sequential deep work never did. I think I could go back to depth-first work if models were fast enough to keep pace with a single train of thought. They have a higher throughput than me but the latency per turn is still high enough that single-threading feels wasteful. While I’m very happy with the kinds of things that AI has enabled me to do (it really is a force multiplier), I’m not sure I will be able to keep this kind of multi-threading in a sustained way. Maybe I’ll adapt. I don’t know…

Coding is over. get over it.

I will miss it

What hasn’t changed #

There are good developers, mediocre developers, and bad developers, and the good ones are still hard to find. The best developers I’ve worked with are never the single-minded coding obsessives. They can clearly explain their POV, and rally others to join them on a mission. They are well-rounded people with a good balance of agreeableness and assertiveness. This was true before AI, and it’s more true now, because coding itself is no longer a scarce thing. Software development IS NOT CODING so I’m still hopeful about the future of my profession. AI will certainly be a “better than average” programmer. The baseline will be much higher. Everyone needs to skill-up to stay relevant.

My concern for the near future is that it is possible that the jobs that AI is taking first are the ones that were entry level positions, where you’d go to actually learn the trade. Learning will start to be something you must want to do, and work hard towards. To my children I’ll try to make them knowledgeable all around, knowing both the hard and the soft sciences. History as well as biology. Philosophy as well as math.