The worst part of AI is the way it's aggressively pushed. Sometimes I have to turn off AI completions in the IDE just because it becomes extremely aggressive in showing me very wrong snippets of code in an incredibly distracting way. I hope when the hype dies down the way these tools are pushed on us in a UX sense is also dialed down a bit.
I feel you. I totally disabled AI completions as they actually were often sidelining me from my reasoning.
It is like having an obnoxious co-worker shoving me to the side everytime i type a new line and complete a whole block of code and asking me if it is good without regards to how many times I rejected those changes.
I still use AI, but favor a copy paste flow where I at least need to look at what i am copying and locating the code I am pasting to. At least i am aware of the methods or function names and general code organization.
I also ask for small copy paste changes so that I keep it digestible. A bonus point is that ChatGPT in firefox when the context gets too big, the browser basically slowsdown locks and it works as a form extra sense that the context window is too big and LLM is about to start saying non-sense.
That said AI, is an amazing tool for prototyping and help when out of my domain of expertise.
So one really big thing that can make the AI autocomplete super useful is to follow the old method from "Code Complete", Pseudocode Programming Process (PPP).
Write a comment first on what you intend to do, then the AI generally does a good job auto-completing below it. I mean you don't have to "sketch everything out", but just that the AI is using the page as context and the comment just helps disambiguate what you want to do and it can autocomplete significant portions when you give it a nudge with the comment.
I've almost fully converted to agentic coding, but when I was using earlier tools, this was an extremely simple method to get completions to speed you up instead of slow you down.
Indeed. That’s my only interaction with AI coding.
Every time Visual Studio updates, it’ll turn back on the thing that shoves a ludicrously wrong, won’t even compile, not what I was in the middle of doing line of garbage code in front of my cursor, ready to autocomplete in and waste my time deleting if I touch the wrong key.
This is the thing that Microsoft thinks is important enough to be worth burning goodwill by re-enabling every few weeks, so I’m left to conclude that this is the state of the art.
Thus far I haven’t been impressed enough to make it five lines of typing before having to stop what I’m doing and google how to turn it off again.
I dunno, I don't find it so bad. I know what I'm about to type, and occasionally the AI guesses right, so it saves me some keystrokes. I don't bother reading/understanding what it suggests if it doesn't match what I had in mind exactly.
I don't even think it saves me much time, but it saves me some keystrokes, which I appreciate due to having arthritis in my wrists.
It's funny because certain ones of my coworkers want to use AI to understand code for them, which they apparently can't do just by reading it. The suggested comments are either wrong or bland even when they just describe a single line or stanza, so I can only imagine how bad and subtly wrong they are when summarizing a directory.
"AI" autocomplete has become rather like mosquitos buzzing around my head that I have to constantly swat away. I'm likely to shut it all off soon, it's just become irritating.
My little experience with AI coding, using copilot on Eclipse, was mixed...
Context: I work with an old Java source code that uses Servlets and implements his own web framework. There is a lot of code without tests or comments.
The autocomplete, I find it useful. Specially doing menial, very automatic stuff like moving stuff when I refactor long methods. Even the suggestions of comments looks useful. However, the frequency with it jumps it's annoying. It needs to be dialed down somehow (I can only disable it). Plus, it eats the allowed autocomplete quota very quickly.
The "agent" chat. It's like tossing a coin. I find very useful when I need to write a tests for a class that don't have. At least, allows me to avoid writing the boiler player. But usually, I need to fix the mocking setup. Another case when it worked fine, it's when helped me to fix a warning that I had on a few VUE2 components. However, in other instances, I saw miserable falling to write useful code or messing very bad with the code. Our source code is in ISO8859-1 (I asked many times to migrate it to UTF8), and for some reason, sometimes Copilot agent messes the encoding and I need to manually fix all the mess.
So... The agent/chat mode, I think that could be useful, if you know in what cases it would do it ok. The autocomplete is very useful, but needs to be dialed down.
JetBrains IDEs have an option to enable AI inline suggestions on demand via a keypress. I really like it. It saves some "boring" typing, while not being annoying.
I'm pretty sure Cursor also has something similar?
Yeah, it’s just horribly wrong in my experience and a complete distraction. Code completion for functions in the project is another story and that has been around for ages.
Agents are great (in so far the models are able to complete the task). Autocomplete copilot just feels like bad UX. It's both, not super effective and also disruptive to my thinking.
I think it depends on the context. If I've been writing the same language and frameworks and code solidly for a few months, then autocomplete gets in the way. But that rarely happens, I like to keep trying and learning new things.
If I'm familiar with something (or have been) but not done it in a while, 1 - 2 line autocomplete saves so much time doing little syntax and reference lookups. Same if I'm at that stage of learning a language or framework where I get the high level concepts, principals, usescases and such, but I just haven't learned all the keywords and syntax structures fluently yet. In those situations, speedy 1 - 2 line AI autocomplete probably doubles the amount of code I output.
Agents is how you get the problems discussed in this thread: code that looks okay on the surface, but falls apart on deeper review, whereas 1 - 2 line autocomplete forces every other line or 2 to be intentional.
The most annoying is when I'm trying to think through a data structure. While I'm trying to deeply think through every member of a class, its type, relationships, etc., this zealous fellow acts like a toddler that knows no way to stay shut unless snoozed off.
