The democratisation of coding
Why coding is no longer just for coders
Earlier this week I noticed my solar inverter had stopped working. Looking at the front panel there was a large red flashing "SYS" light. What to do?
Ask AI, of course.
But what model to use? This is the type of problem that inevitably involves web searching across multiple sites - something that o3 is particularly good at - so o3 it was. o3 found the inverter manual, realised the flash pattern contained an error code and asked for information about it.
I gave o3 a video of the flash pattern. And then watched as o3 wrote code to analyze the video. It chopped the video into frames, found the indicator light and looked for changes in brightness to figure out the flash pattern. True - it didn’t find this easy - at one point it looked like it might give up.
Fortunately o3 continued and, after writing 200 lines of complex Python and thinking for nearly 7 minutes, we had an answer. There was an inverter fault and I’d need to contact the installer. OK, so it wasn’t the answer I’d hoped for; I’d hoped I could fix it myself, but at least I knew what to do next.
But consider those 200 lines of code. They were complicated. Writing that amount of code by hand likely takes at least half a day. Which, assuming an annual salary of $100k, costs $250. But now AI creates that code for essentially free. And often without us even realising.
o3 writes code all the time. Ask it to analyze a picture and it’s likely to quietly generate python code in the background. Code gets generated to help with analysis problems, maths problems… When code generation is cheap why not use it as much as possible?
It’s not just o3
Claude artefacts are another way non-coders can now get access to coding. Earlier this year I watched as someone from HR used Claude to create their own web tool to build a customized onboarding process. It’s the kind of thing that just wasn’t possible 12 months ago - building such a tool wasn’t economically viable.
The concept of no-code - tools that let users build software through drag and drop components and prebuilt logic blocks - isn’t new. But up to now it’s been limited. You could only do what the blocks permitted. Now, with a little careful prompting, Claude will build you a custom tool. The use cases are many and varied.
And then we have the heavyweight software development tools. There are two categories of these:
VSCode forks which integrate agentic AI abilities. Cursor is the best known of these, along with Windsurf - and, of course, Github Copilot. This week Amazon joined the party, introducing Kiro.
Command-line agentic tools - Claude Code, Gemini CLI, Jules, Codex, Claude MCPs.
The former take a familiar development tool - VSCode - and add AI agentic capabilities; the latter start with (approximately) an LLM chat interface and then add AI agentic capabilities.
This market is hot. Cursor is valued at $10B - about the same as EasyJet - for what is arguably little more than a VSCode fork with AI smarts built-in. In the past week an attempt by OpenAI to buy Windsurf for $3B fell apart, and Google swooped in to poach the Windsurf CEO and key leaders for $2.4B. And then Amazon, who have invested $8B in Anthropic, launched their Claude powered VSCode fork, Kiro.
Kiro differs from Cursor and Windsurf in that it introduces "spec-driven development", where detailed planning and structured components drive the coding process - perhaps a nod to the realisation that vibe coding isn’t suitable for the enterprise.
In amongst all of this, we’re starting to see the shape of the future of coding:
At the very bottom are agentic tools such as o3 which code without you even realising.
Then we have one-shot coding tools such as Claude artefacts. These can easily build simple websites - and, critically, enable non-coders to create a wide range of simple applications.
Then there are agentic coding tools; these are more capable, but more complex. They will autonomously build small applications (currently in the 1-2kloc range), but you’ll likely need to provide some oversight and debugging. And manage deployment, install, upgrade.
And, finally, the enterprise seems set to continue using VSCode forks - with some mix of AI autocompletion/ AI agents for large-scale critical codebases.
Note that the first three of those are all new - 12 months ago they didn’t exist. We’re still very early in the process of figuring out how to use them. Nor is it as simple as the enterprise will only use VSCode forks. They’ll inevitably benefit from the lower three levels as well.
While it would be unwise to let agentic coding tools loose on your operating system codebase - at least for now - we can use agentic coding tools to build tools that operate on the codebase. We can use those agentic tools to build test frameworks, to explore performance, to write deployment tools, to test theories, to understand edge conditions. These tools help our developers go faster. And that improves the reliability, the quality, the performance, the maintainability of our critical enterprise code.
And the way we approach software problems will change. Agentic coding tools can generate multiple different solutions in parallel. If it’s possible to readily evaluate the quality of different solutions then we open the door to brute force coding - implement multiple solutions and choose the best. Lest this seem inefficient, it’s worth noting this is roughly how the gradient descent used in AI training works - and it’s also the basis of the Super Grok heavy variant launched last week.
And so?
Coders have long understand the amazing problem solving power of programming. Yet until now this power has been inaccessible to the majority of people beyond crafting a few formulas in Excel.
But all of a sudden we’re moving to a world where the benefits of coding are accessible to a much larger set of people. Need 200 lines of complex Python to decode a video? No problem, o3 will do it for you automatically. Want a custom calendar app? Build a Claude artefact.
Suddenly the receptionist can build HR tools, the facility manager can debug equipment, and the small business owner can create custom applications. Professional developers gain AI assistants for their critical codebases. The gap between having a problem and coding a solution is shrinking fast - and we're only at the beginning of understanding what becomes possible when programming is no longer a bottleneck.