Remote ArcREMOTE ARC BLOG
Could AI become the next compiler?
I've been thinking about a possibility: what if AI becomes something like the compiler was for an earlier generation of computing?
For a long time, making software do something meant learning a programming language and writing code ourselves. What if we increasingly just describe what we want, and AI writes the code, uses the tools, or even finishes the work?
Follow that idea far enough, and code might become a little like assembly language today. It won't actually disappear. Most people may simply stop needing to write it by hand.
I think we're already seeing the shift
If you look back at computing, so much of its progress has been about hiding details. Machine code, assembly, C, JavaScript, Python. At each layer, we think a little less about what the machine is doing underneath.
Claude Code and Codex can already read projects, change code and run tests. Instead of explaining every step to the computer, I can increasingly tell an AI the result I want. To me, that's more than faster programming. It may change how we interact with software.
There is a catch in the compiler analogy
A compiler doesn't have to guess what you meant. It translates instructions with defined rules. AI is different. Human requests are often ambiguous, and a wrong guess can have real consequences.
Say I ask it to clean up unused files. What counts as unused? Is uncommitted code disposable? Should deletion be reversible? A smarter model shouldn't automatically get to make those decisions for me.
So I don't think code or software engineering disappears. I think people may write less of the code themselves, while defining requirements, checking outcomes and managing risk becomes even more important.
And prompts may not be the final interface either
People spend a lot of time learning how to write prompts. But if AI eventually understands what people mean, why should everyone need to learn a complicated set of prompt tricks?
Maybe the more important questions are: what do I want, what must not happen, and how do I know the job is done? The AI should ask for missing details rather than expect one perfect instruction.
At that point, the interface might be neither code nor clever prompts. It might simply be human intent.
Where does that leave Remote Arc?
It also makes me rethink what Remote Arc should be. If models keep getting better at writing code and using computers, we can't build our whole value around simply letting AI take action. Claude Code and Codex can already do a lot of that.
The more interesting question is what comes next. Once AI knows what to do, how does it reach the user's actual computers, browser and apps? Can different AIs share one permission and execution layer? Can we tell what actually happened, and what failed?
That's why I see Connect as Remote Arc's infrastructure: a way for existing AI clients like ChatGPT, Claude and Codex to connect to a user's environment. Arc is our own first-party experience built on that same layer, not another separate platform.
But being cross-platform isn't automatically a moat. Model companies and operating systems may build execution features themselves. And do people really need several AI clients across multiple computers, or is one assistant and one machine enough? We need real usage to answer that, not a convincing story.
Security and reliability matter just as much. Running a command isn't the hard part. The hard part is knowing when not to run it, understanding failures and handling interruptions safely. Remote Arc hasn't solved all of this: terminal access isn't an OS sandbox, and Undo can't reverse everything. To me, that work matters more than simply adding more tools.
Where I've landed, for now
So I still come back to the original idea: code won't disappear, but people may write less and less of it directly. Prompts might be a transitional step too. More of our job may be to tell AI what we want, what it must not do, and then check what actually happened.
I don't know exactly where that leads. But for Remote Arc, the thing worth proving isn't whether AI can act. It's whether people can actually trust what happens after it does.