Is AI Actually Writing Code Yet?
A look at Kite, TabNine, and GPT-3's promise for autocomplete-style AI coding tools, and how far they really are from writing features unsupervised.
Every few weeks someone posts a clip of GPT-3 spitting out a working function from a plain-English prompt, and the comments fill up with “programmers are obsolete” takes. Having actually used the AI coding tools available right now, I can tell you: not yet, not close, but also not nothing.
The tools worth talking about today are Kite and TabNine. Both plug into editors like VS Code, PyCharm, and Sublime, and both use machine learning models trained on large amounts of public code to predict what you’re about to type. Not just the next token like old-school autocomplete — these can suggest entire lines, sometimes whole function bodies, based on patterns they’ve seen across thousands of repositories.
What they’re actually good for
In practice, the win is boilerplate. Writing a constructor, filling in an obvious for loop, guessing the next argument in an API call you’ve used a dozen times — that’s where these tools shine. TabNine in particular gets eerily good at finishing a line the moment you’ve typed enough of it to signal intent. It’s also handy as a kind of live API reference: instead of tabbing over to documentation to remember a method signature, the completion just shows up.
Kite leans a bit more into surfacing documentation and examples alongside its suggestions, which is useful when you’re working in a library you don’t touch often.
Where it falls apart
Ask either tool to write “a function that validates a user’s password and returns error messages” and you’ll get something plausible-looking that’s wrong often enough that you can’t trust it unread. These models are pattern-matchers over code they’ve seen, not systems that understand your codebase’s conventions, your data model, or what “correct” means for your specific feature. They don’t test their own output. They don’t know your business logic. Every suggestion still needs a human to read it, understand it, and decide whether to accept, edit, or reject it.
That’s the current ceiling: assistive, not autonomous. A very fast typist sitting next to you, not a junior developer you can hand a ticket to.
Why GPT-3 changes the conversation
What’s got people speculating harder than usual is GPT-3, OpenAI’s huge language model that’s been making waves this year for generating fluent text from a prompt. It hasn’t been packaged into a dedicated coding product yet, but people have already shown it can produce short scripts and snippets from natural-language descriptions, more flexibly than the pattern-completion you get from Kite or TabNine. If a model that size were ever fine-tuned specifically on code, the completions might stop being “finish this line” and start being “write this function from a description” — a real jump from what’s on editors today.
That’s speculation, not a product review. For now, the honest verdict on AI-assisted coding is: it saves you keystrokes and the occasional trip to Stack Overflow, and it’s genuinely worth installing. It is not writing your features for you, and anyone telling you otherwise hasn’t tried to ship something with it.