Building at the Intersection of AI and Human Potential

I spend most of my working day with AI. I write code with coding agents, I use models to read documentation I'd otherwise skim, and I make a YouTube channel, The Synapse Times, that's all about the field. So I'm not writing this as a sceptic standing outside. I'm writing it as someone who is in it, likes it, and still thinks a lot of the conversation around it is pointed in the wrong direction.

The question I keep coming back to is simple. When someone uses the thing I've built, do they come away more capable, or just more dependent?

Two ways to build the same product

Take almost any AI product and you can build it two ways.

The first way aims to replace the person. The pitch is "you don't need to understand this any more, the model will handle it". It feels great in a demo. It also leaves the user worse off the moment the model gets something wrong, because they've been encouraged to stop paying attention and they no longer have the context to spot the mistake.

The second way aims to extend the person. It does the tedious parts, shows its working, and leaves the person in charge of the judgement calls. It's harder to demo because the human is still visibly doing something. But the person using it gets better over time instead of slowly losing the skill.

I think the second kind is what's worth building. Not because the first kind is evil, but because it's fragile. A system that only works if nobody checks it isn't a finished system.

What I've learned from coding with agents

Coding agents are where I've felt this most directly. They are genuinely useful. I can hand off a refactor, a test suite, or a first draft of a feature and get something back in minutes that would have taken me an afternoon.

But I've also noticed what happens when I get lazy with them. If I accept changes I haven't read, I end up with a codebase I don't fully understand. Then, a week later, something breaks and I'm debugging code that is technically mine but that I've never really thought about. The time I saved comes back as interest.

The approach that works for me is boring. I stay clear about what I want before I ask. I keep the tasks small enough that I can review the result properly. I read the diff. I treat the agent like a fast, keen junior engineer: very helpful, sometimes confidently wrong, and not the person who signs off.

Used that way, the agent makes me faster and I still understand my own system. That's the whole point.

Capability is the thing that compounds

There's a reason I care about this beyond my own workflow. Skills compound. Someone who understands their tools can adapt when the tools change, and in AI the tools change every few months. Someone who has handed over their understanding has to start again every time.

So when I think about what "human potential" means in this context, it isn't a slogan. It's a practical test:

Does the product teach the user something, or hide everything from them?

Can the user tell when it's wrong?

If the product disappeared tomorrow, would the user be better or worse off than before they started using it?

I don't always pass that test with what I build. But having it written down keeps me honest.

Where the hype gets in the way

A lot of AI marketing promises that you won't need to think. I understand why; it sells. But I think it does damage, because it sets people up to trust systems in exactly the situations where they should be most careful.

The more honest message is less exciting: these tools are powerful, they're uneven, and they reward people who know what they're doing. That's not a reason to avoid them. It's a reason to learn them properly.

I'd much rather tell people that and have them build real skill than oversell something and have them quietly give up the first time it lets them down.

What this means for what I build

At Hepteract Group, this is the filter I try to apply to everything. The tools should do real work, not just look clever. They should keep the person in the loop where it matters. And they should leave people more capable than they were, whether that's an engineer, a small business owner, or someone just getting started with technology.

It also shapes how I make content. On The Synapse Times I'd rather explain how something works, and where it falls short, than just announce that everything has changed again. People can handle nuance. They're usually grateful for it.

The bit that stays human

I don't know exactly where AI is going, and I'm wary of anyone who says they do. What I'm fairly sure of is this: judgement, taste, and responsibility don't get automated away just because the typing does. Someone still has to decide what's worth building, notice when it's wrong, and own the outcome.

I want to build tools that make more people good at that. That's the intersection I'm interested in.