How AI coding companions will change the way developers work
allthingsdistributed.com
allthingsdistributed.com
Everyone seems to be raving about its ability to write code. What am I doing wrong?
It seems to have a lot of trouble with building code that works in a certain way. For example if you ask it to build a simple graph container with access functions that give you iterators over neighbourhoods of nodes or whatever it makes a mess. Sometimes it technically works, but the underlying graph datastructure is just ripped out of randomseocodingsite.com using nodes owning pointers to their connected nodes.
If you ask it to use a specific underlying structure like a vector of nodes and indexes to represent edges it doesn't quite know what to do.
So yes, I guess the initial implementation "works", but not well.
For wgsl and webgpu, ChatGPT4 is quite useless.
To be fair, the spec is still changing, but this means, that ChatGPT4 will mix in various incompatible bits. I hope they handtune it.
Either you already know how to do everything or you’re not asking it a wide range of questions about the work you’re doing.
It’s not there to write your code for you, although it can write snippets and short functions. It’s thereto answer questions like “how do I….”, “how does this work”, “explain this code”, “what’s the correct typescript type for….”, “what’s the maximum decimal of uint16”, “difference between and y”, “give me an example of”, “what are the arguments to”, “what causes this error”, “explain this code”, “given this data structure, how do I get the third row, first item”, “how do I Centre a div”, “create a header file for these functions”, “what’s the command to compile this code”, “what’s the regex to get lines containing foo”, “ffmpeg command to convert X to y”, “given this sql table, write a query to …”, “make 1000 rows of sample json data with fields name address phone”, “write a sed command to search and replace X with y”, “create a cmakelists to compile this”, etc etc and critically important “that didn’t work”, “another approach”, “any other ideas?”.
Then you debug the code it gives you.
I have always wanted to rap and it has wrote rap lyrics that are way beyond what I could possibly do myself. If I was Andre 3000 or Tupac though I wouldn't be impressed at all.
That is why asking experts in their domain what they think about chatGPT seems to miss the point to me.
The primary use-case of LLMs is to reduce the barrier to entry for tasks that used to require a lot of learning. In the past, some people were employed mostly just for knowing how use some library or framework. That will no longer be the case. The value of knowledge that can be found in a book (not the same as "learned from a book") has been going down with digitization (internet, Wikipedia, search engines) and with LLMs its value will be reduced to almost nothing.
For programming, you can just test if it works. That's how most people[1] check if their own code is trustworthy, absent considerations of security against malicious attackers.
In general, knowledge and authority or trustworthiness are different things. I said this makes knowledge worthless. It does not make authority or proof of authenticity worthless; probably, the opposite.
[1]: those who aren't able to always write 100% perfect code on the first try
The possible applications of the technology have known required levels of trustworthiness. As the tools surpass any given threshold, new use-cases become attractive. Already at current levels of performance there are many applications where it's easier to test or continue researching a hypothesis than to find plausible and relevant ones.
No source of information is perfectly trustworthy and we have to deal with that. This is sometimes called "critical thinking" and might become a more important skill. There is also a good chance that AI systems can be trained to do this themselves and "fact-check" the responses of others while providing sources and reasoning, which would be one tool to increase the level of trustworthiness.
If you come to 'trust' AI for critical thinking, how do you then know when or if it has been manipulated (surely it is naive to think the governments and power centres won't do everything they can to make AI say the 'right' things)?
But seeing how much trouble non-writers have in producing anything longer than a text message, I can see that it would have real value for them.
It's worth remembering how hard it is to make a start when you have no idea what you're doing, and the worst part of the learning process is when you know something is wrong but you don't know how to fix it.
ChatGPT seems to excel in a few narrow fields, but I think it can take a lot of people from mediocre/non-functional to somewhat acceptable and that's a pretty big leap.
If you literally cannot code at all, GPT-4 is a solution.
Similarly, I cannot draw at all. But I can endlessly prompt Midjourney until I get something acceptable.
People who can't code can try GPT over and over again, and hopefully eventually get something that works (even if more by coincidence, who cares as long as it works!)
I use it to code in languages I don't know.
Your own example is not proof that this time it's different - this is just a no-code tool that you happen to gain productivity by using.
It surely is not suitable for everything, but it would not be broadly in use, if it would be a toy.
Maybe there's a parallel there, that 90% of no code projects do provide complete solutions and 10% are problematic
I am not a good coder, nowhere near the level of even your average professional developer.
I’m OK, and I have been writing code of some kind or another since I was a kid in the late 80’s and have managed many technical projects and I would say I have a pretty deep understanding of how it all works.
The thing I’m best at is devops type stuff and rails but even with that I’m wildly out of practice since I’m a founder/CEO type of a non software company and I can go a week or two without even touching code.
These tools have completely changed everything for me in a way that is really disconcerting.
Previously every time I wanted to change an app I had worked on, add a feature to something, or do something like call an API to grab something interesting from another business resource and pass it along, I was in for a couple hours just getting my head into what I was doing. I’d sort of block out a weekend and by sometime Sunday afternoon I would be getting there.
Now I can just start talking to GPT like it’s an experienced developer sitting next to me. Hey so if I wanted to call this service how would I structure the query. Hey look at this curl and tell me some reasons it might be getting this error.
It definitely can’t do things for me. Without a pretty deep background in software as a concept and practical experience I wouldn’t even know where to start.
But I do have those things. And also I’m the guy who just knows the business logic and what I want.
Now I can get what I want like 30-40x faster. To be honest it feels very strange, after being sort of frustrated by translating my requirements into specs for 20 years to work with developers and wasting a lot of time it feels like someone popped a babel fish in my ear.
So there’s one anecdote.
I have a feeling most of the excitement is coming from the "citizen developer" space.
It's also useful for suggesting basic boilerplate code for something you're not very familiar with. If you're not familiar with something, you might get stuck in the phase of how to even get started.
In addition, it's useful for short but mathy code. I needed to place a bunch of items in a circle on an arbitrary plane in 3D space, and GPT-4 gave me a perfect answer for it.
Because GPT-4 helped immensely with what was my very first NN ML project.
I didn’t approach the problem with no knowledge as I’ve had a few ML projects over the decades (graphs clustering, PageRank, eigenfaces) and I’ve skimmed a few books on PyTorch and neural nets, worked through Karpasky’s zero to hero, etc.
So I knew how to ask very specific questions and I knew the general overview but I had literally never written a single line of PyTorch code.
I’m now working on the iOS app to host the model and send MIDI and OSC data to a DAW. I haven’t made an iOS app in about a decade and that was in the Obj-C days so I’ve never written a line of Swift.
I have a demo app written in Swift that accesses the camera, converts frames to the model input, and makes predictions, and now I’m ready to use this to assist with labeling many more images!
I'm working on something to make this more manageable, hopefully a lot more. You can get an email if you sign-in at https://inventai.xyz.
If something requires even a slight bit of boilerplate that has a pattern to it, it writes it much faster than I could. I find it almost uncanny sometimes how well it knows what I wanted to do next even without writing any comment telling it what to do. Not to mention the hundreds of random small suggestions like writing similar method definitions, completing all of the parameters I was going to pass to something, writing out the rest of some math formula after I typed the name of the variable for the result I want to compute, etc.
People weren't kidding about the productivity boost. On the other hand, I found trying to get ChatGPT to write large chunks of code almost completely useless - it would always be incorrect in some subtle way or failing to use the style I gave it as an example, unless the task is to write some simple python scripts.
Excerpt one thing I noticed. That the times it is subtly wrong I waste a ton of time finding the bug. It’s like looking for bugs in someone else’s code. I lack the context and muscle memory of the code written by the assistant.
I’m still not sure if it’s net positive. I wish there was an objective way to measure the effect on a personal, project and org level.
One example recently: I watched a developer scaffolding a database schema with an ORM. Plenty of boilerplate defining the fields and stuff, nice to have Copilot running, however the suggestions were baking in bad design that would work now, but almost certainly cause pain later: nullable foreign keys, text fields with inappropriate max lengths, redundant FK relationships where a relationship could be ambiguous depending on whether you joined directly on the FK or through another related table, etc.
Obviously all of that is on the developer to notice and be responsible for, but it’s really easy to just tab through the completions, see that it works (right now) and move on.
I research all the time because I find what I do fun. I read books, listen to conference talks, practice synthesizing, and project to explore ideas. (I will say there is a lot of conflicting advise out there so it’s important to try to mentalize the perspective to understand why the particular advise is offered. Anyways, I digress…)
I also have ~10 years programming as an IC.
Because of what I stated above I know the high level “strategy” of what I’ll want to do and most (if not some) of the implementation details. AI helps me with the “tactics” of implementation. I can ask AI or allow copilot to make implementation suggestions and I know enough to tailor the suggestions to my project (and to my coding preferences). Typically I still cross reference using documentation but using AI has helped reduce the amount of cognitive load required to do something interesting.
We live in exciting times as programmers. It’s only going to get better as our tooling becomes more sophisticated.
I've wrestled for many hours trying to get them to write me a short, simple function in a non-popular language and they failed again and again.
I also suspect these LLMs are going to have huge problems when APIs and language definitions change, because the data they're trained on usually doesn't have API/language version metadata attached to it, so they have no way of knowing which version is applicable to whatever you're trying to do.
They may spit out code in an obsolete API version, for example, or even give you code containing mixed API versions. This will be a nightmare to debug.
It kept apologizing and giving slightly tweaked code, which I could see was not actually addressing the issue.
Then when I asked it if the feature was actually even possible, it said No.
Oh yes, you have to run rspec init after you bundle install it. Don't forget factorybot needs this little config file. Oh didn't I mention, form_with breaks error messages? And on and on. I knew exactly what I wanted (I used to know it pretty in-depth), but just experienced constant shin-scrapes and paper-cuts getting back into it.
This is where I imagine GPT shines - though I don't know because I can't get an account.
- Use GPT-4. It's vastly better than 3.5.
- Use an integrated workflow like copilot or aider. Cutting and pasting code between your editor and a chat window/tool is tedious, and doesn't give GPT any context about the rest of your code. It forces you to do extra up front work to prepare your code to call a bite-sized function that can be easily cut/pasted. Asking GPT for revisions is slow with more cut/paste at every step.
- Coding with AI lowers the activation energy to start a new change, since you expect it to be easy.
- GPT is a champ at boilerplate, and can quickly prepare the ground for the interesting work.
- Typing a few terse, informal sentences describing the task to GPT can help organize my own thinking.
- I spend so much less time fiddling with annoying details of syntax, method names, parameter names/order, regex syntax, etc.
- It's important to break down a big change into a series of smaller, self contained steps.
- Watching GPT type out code is good "think time" for me.
- GPT is great at "busy work" like updating docs, adding small quality of life features, writing tests, etc. I accomplish 10x as much of this work now because it's so easy to just ask for it.
- GPT has astounding knowledge about available libraries and packages that can simplify your task.
- Expect to iterate rapidly with GPT on more complex code changes, offering it feedback/direction.
This will be like Solar Winds 10x by letting the attackers right into the most sacred parts of your organization… writing the code that will make or break your company.
But can’t we just filter the training set for bad things? No. The whole point of zero days is that bad folks know what they need to get inserted into your codebase before you (or companies training an AI) know it’s bad.
Dear god. Please let me know that you need to have code reviews, unit tests, and a QA staff for "critical applications". An AI doesn't replace any of that.
I need to hear you say it.
For my service to work at all, I feel like I really need GPT-4 or something better if that ever comes along.
But aside from GPT-4 trying to use APIs from two years ago, it can definitely make simple software without necessarily needing an actual programmer. Especially when you put it in a closed loop.
As the models get better with larger context windows and more up-to-date training, software engineers will probably defer more and more to the AI. And non-programmers will be able to build software with it also.
The time where this is just minor code completion or something is going to be pretty limited. In 3-5 years, I would be very surprised it was considered good software engineering to have humans writing code at all.
E.g. I used it on a project recently to translate code from one version to another, and keep in mind that the syntax was very different
A comparable example would be translating old class based React code to functional TSX with hooks.
And it does extremely well there. Other coding tools can’t even do that as there’s no way to load a whole prompt and the entire original file as a prompt.
I also remember how much just having an integrated development environment helped with the process.
And then intellisense! Wow -- I thought ANYbody could be a programmer then, and got worried for my job. :-)
I think this is just the next iteration. Yes -- I would expect AI large language models and general purpose AI to change the landscape of how developers do their work. It will be uncomfortable for a while -- but go to your user groups, listen to your podcasts, share the adventure with other people. We get to make it fun.
Microsoft is really trying to hear up how awesome copilot is going to be, but the real world doesn't operate in an MS Lab.
Our AWS rep has also said that they’ll keep our code private and won’t add them to the training set, you know, actually paying attention to legal issues like grown ups rather than creating a yawning chasm of risk.
Sandeep is being very optimistic. AI coding companions open up coding capabilities to even less experienced coders who mostly care of the code works. Whether the code is good will be questionable.
Not to mention security or other concerns.
I think it would slow me down in areas I already know about. The IntelliJ tools are already pretty good.
Based on the information available so far, here are the key takeaways from the interview on how AI coding companions will change the way developers work:
1. Generative AI is significantly impacting developer productivity and speed, particularly through the use of coding companions like Amazon CodeWhisperer, which uses a large language model (LLM) trained on open-source projects, technical documentation, and AWS services. This tool handles much of the heavy lifting associated with creating new applications and services【5†source】.
2. Large Language Models (LLMs) not only predict the next line of code but also understand the developer's intent and infer context from what's already been written, including comments, thus generating syntactically valid, idiomatic code. They also simplify mundane and time-consuming tasks, such as writing unit tests or translating code from one language to another【6†source】.
3. CodeWhisperer and similar AI tools are not meant to replace developers but to assist them in focusing on the more complex and interesting aspects of their work, such as problem-solving【7†source】.
4. The plugin of these tools is capable of understanding the developer's context, the programming language being used, and the specific task being carried out. It can provide timely code recommendations and has advanced features like a reference tracker to aid the developer in making the best decisions for their applications【14†source】.
5. LLMs are trained on a vast amount of data from the public internet, cleaned and organized to understand the vocabulary and structure of the language, and to generate meaningful sentences and paragraphs in that language【15†source】.
6. The models built on transformer architecture will soon allow for automatic translation from one programming language to another. This will be particularly useful for teams looking to transition from legacy languages to more recent or efficient ones, such as Rust for high-performance applications【16†source】.
I encountered a few issues while trying to access parts of the interview, so there might be additional important points not included in this summary.