TL;DR: Using AI in software product development doesn't change the process. It simply compresses the process and allows us to deliver features and conduct experiments faster than ever before. The differentiator is still going to be in solving people's problems, just as it always has been. The digital camera freed the photographic artist from spending years in the darkroom to learn the hands-on skills first, but the artist doesn't abdicate to the automated camera. Similarly, our super-capable LLM coding models free product creation from lengthy and expensive engineering cycles, which just means that the product focus becomes even more important than before as engineering ability is no longer the moat it was.
Worrying About the Wrong Thing
One of my beliefs in product development is that we are always doing the wrong thing. Whether it's a brand new product concept or a long-toothed cash cow inside a venerated company, we're still probably not doing it all perfectly.
My simple three-step mindset is to figure out what we're doing wrong. Then decide what we should do to improve upon it. Finally we have to actually do that thing. People spend entire careers inside all three of those steps. I'm going to do my best to convince you that frontier LLMs collapse a lot of the third step, have roles in the other two, but in no way meaningfully take over the whole process.
Product focus becomes even more important, as engineering is no longer a moat.
If we're building a product (that someone will hopefully buy), we're going to have to make a lot of guesses in the beginning. We're going to have to find some starting points - what is the burning and valuable problem that this product is solving? How much will someone pay for it? How do we make the person aware of it? How do we install it, get it in their hands, or allow them to use it? We use our best judgement, creative thinking, experience, and just pick a starting point for any of these answers. Then we take a look at what happened and try to improve.
My default assumption when building, launching, taking over, or improving a product is that we're doing the "wrong" thing. We might be doing it well enough for everything to be viable, but there's always room to do it better. "Wrong" might be harsh, if we're doing it well enough then it's good enough - but I've found saying it's good enough takes the fire out of the iterative improvement. Thinking about how it's still wrong gives us a reason to keep striving, so strive we will.
Once we've decided what needs to be improved we've got two more jobs - figuring out what might improve it, then implementing that as quickly as possible. Deciding what to try to improve it might be obvious to us; it might be continuing along the same pathway we've been going, continuing to optimize in the same ways. Maybe we try something completely novel and new, and try to break out of those local maxima. But however we come to that opinion, we've come to it. Ideally the change is rooted in previous iterations, feedback, data, and experience that are collectively so staggeringly overwhelming that it's clear what to do. And sometimes it's not! This is where we take our best shot, and really great teams often distinguish themselves at this point by having just the right intuition fueled by feedback, experience, and creativity to pull out some really great incremental ideas. So let's go implement it.
Our third job is implementing this change and getting it done. This is the difference between a high-agility organization and a monolithic lump that can't get out of its own way. A slow company (doesn't have to be big) that takes 37 meetings, 12 presentations, and two quarters to try out a new incremental idea is always going to lose the innovation race. A company that can roll out the change in 4 hours is going to close the loop quicker. All caveats aside - those specific numbers might not apply to your industry, if you build dams or spaceships or aircraft or medical equipment or something else where life and limb are involved. But the spirit of it is the same. And try out this statement for size and see if you can't also agree with it despite what you think are the constraints of your industry:
The quicker you can evaluate this loop, the quicker your product gets better.
And this is the spirit of the MVP as we've talked about before - it's the quickest experiment you can rip out to learn something about creating value. Are you doing the right thing? Yes? Let's do some more of it. No? Let's try something else.
So Where Does AI Come In?
Now let's add some AI to the mix. The world changes quickly, so let's establish where we are. As of this writing, October 2026, the frontier models used for writing code are all good. The benchmarks are all over the place, but Claude Fable and Opus, and GPT-6 Astra and Gemini and Deepseek and Kimi and Qwen have all got places in different coding schemes and harness architectures.
By "good" I mean they are fast, cheaper than a person, and generally capable of creating adequate results. They can do simple tasks remarkably quickly and well. Successfully completing a more complex engineering task requires a structure similar to what we do with human engineers - the overall vision or a Product Requirement Doc (PRD) is very important. Then we spin up groups of agents to write the code, other groups to test it, other groups to merge it, maybe others to release and monitor it, etc. And often we inject human architects and senior engineers in the project to establish shape, order, worry about more abstract concerns and non-functional requirements - we did the exact same thing with human engineers, but some of the functions are now displaced to LLMs.
What does this actually do? It allows us to write adequately functional software, vastly more quickly and cheaply than we could before. It's medium-grade code. It's not excellent. But it is really fast and really cheap (compared to human coders), and if you've set up smart guardrails it can be perfectly functional and well executed.
Returning to our 3-step process from earlier. If we bring our AI-enabled software development pipeline into play it has the potential to collapse that third step of the iterative process. The time to build, deploy and test a new feature tweak can be reduced to potentially a handful of minutes; it could be seconds if you really cared enough. This is good news. Great news in fact (if you care about building products and solving problems)! With some downside mixed in...
Companies that used to iterate quickly on hard problems often had large teams of really good, really smart, and really well compensated engineers. Running that iterative loop was fast because they fed and lubricated it with huge amounts of cash and equity. Now, we can run and lubricate it with cheap tokens. A $20 or $100 per month LLM plan is enough for anyone to prototype product ideas in their spare time - the implementation equation has forever changed.
LLMs can be helpful in our first two steps, but those are still the core of the product responsibility. They don't inherently eliminate or collapse deciding what matters and deciding what to do about it. The actual job of a startup is to find a set of truths about pains people have, what they'll pay to solve them, and a way to solve them. The job of an established product company is to get better at all of those things, while still holding on to the ability to do something new (which is hard). Building software really quickly just means we can now test and validate ideas even quicker, which gets us back to the real business: solving painful problems.
What Now?
The proliferation of digital cameras put photographic creation - and when coupled with the internet - distribution, in the hands of literally anyone. I'll argue that it allowed broader democratized access to the medium, while simultaneously tanking the ratio of good to bad photography available for perusal. Just a generation ago a budding photographer might spend years studying the greats who had come before, learning what works, what doesn't, and the meditative practice of years in a darkroom pondering in a slow-work way the shapes, forms, compositions that become truly striking. They probably learned in school or university or community darkrooms, alongside working professionals, mentors, and others in a slower drawn out process.
But there were two parts to this. The first was the technical skill - learning to manipulate the camera controls, whether on a smaller camera, or a medium format or large format camera. Learning how the exposure options they selected when they took the picture could then interact with changes in chemistry, timing, or technique while developing or printing to control contrasts and grain and other subtle feels. It forced them to develop an artistic concept up front, then back through a chain of related and learned skills to realize the vision. The second part is the artistic portion - learned alongside the slow absorption of the technical skill. Digital cameras super-streamlined the first part. Smart picture modes allow someone to click away with only minutes of instruction. And it's wonderful, they can capture pictures of family, friends, news events, or anything else they wish to in their lives (and now clearly the phone has taken over this role as well). But the second part is not fulfilled by the first - years of study, reflection and refinement are still required to create the types of images that, hung on a wall, would inspire people to sit and think at a museum or show. We separated the two, and erased the technical barrier on the front end - but the back end remains. The path to excellence hasn't actually changed too much. Simply pointing a camera at something and clicking abdicates the role of the artist to the camera's brain, and the state of the world at the time (composition, lighting, contrast, form, texture, gesture, etc). It's relinquishing an active role as creator, to a passive role as receiver. Anyone can be a receiver, but it's the creators who stand apart.
This is exactly what's happening with creating software. The first part - using languages, writing code, spending years in industry moving up a ladder from a junior to mid to senior engineer is all massively changing. The important part - knowing which code to write, and how, to what end, and how to architect it all at a higher level is still within the human realm. Excellence in the tactical hands-on work of writing code is vanishing, much like the darkroom technician disappeared. But the high level goal of photographic art was to communicate emotions and feelings, stimulate thoughts, and cause some of what was in the artist's head to be absorbed into ours. Similarly the goal of software products is to do something valuable for someone, to transfer some secret you know about solving a problem into the lives of the customer. The value of the product is not in how much money you paid to have it written.
We're in a time where really great new ideas can take form quicker and with less investment than ever. And where a million terrible ideas can blossom at no cost at all. Access to capital was a gatekeeper that has now vanished alongside access to a darkroom. This will allow some great ideas (and people) who had been gate-kept out to play in the various games. It's extinguishing some barriers to entry, filling in some moats, and making velocity and vision our greatest advantages.
We would never have dreamed of confusing the camera with the artist, so let's not do the same with the LLM and product. Excellence is still excellence, so let's remain in the artist's seat and continue to create. It's really the only entirely human thing we can do!
- This was neither written nor edited by AI. These are 100% Human Words(tm).