I finally turned off GitHub Copilot yesterday.

David Chisnall (*Now with 50% more sarcasm!*)

I finally turned off GitHub Copilot yesterday. I’ve been using it for about a year on the ‘free for open-source maintainers’ tier. I was skeptical but didn’t want to dismiss it without a fair trial.

It has cost me more time than it has saved. It lets me type faster, which has been useful when writing tests where I’m testing a variety of permutations of an API to check error handling for all of the conditions.

I can recall three places where it has introduced bugs that took me more time to to debug than the total time saving:

The first was something that initially impressed me. I pasted the prose description of how to communicate with an Ethernet MAC into a comment and then wrote some method prototypes. It autocompleted the bodies. All very plausible looking. Only it managed to flip a bit in the MDIO read and write register commands. MDIO is basically a multiplexing system. You have two device registers exposed, one sets the command (read or write a specific internal register) and the other is the value. It got the read and write the wrong way around, so when I thought I was writing a value, I was actually reading. When I thought I was reading, I was actually seeing the value in the last register I thought I had written. It took two of us over a day to debug this. The fix was simple, but the bug was in the middle of correct-looking code. If I’d manually transcribed the command from the data sheet, I would not have got this wrong because I’d have triple checked it.

Another case it had inverted the condition in an if statement inside an error-handling path. The error handling was a rare case and was asymmetric. Hitting the if case when you wanted the else case was okay but the converse was not. Lots of debugging. I learned from this to read the generated code more carefully, but that increased cognitive load and eliminated most of the benefit. Typing code is not the bottleneck and if I have to think about what I want and then read carefully to check it really is what I want, I am slower.

Most recently, I was writing a simple binary search and insertion-deletion operations for a sorted array. I assumed that this was something that had hundreds of examples in the training data and so would be fine. It had all sorts of corner-case bugs. I eventually gave up fixing them and rewrote the code from scratch.

Last week I did some work on a remote machine where I hadn’t set up Copilot and I felt much more productive. Autocomplete was either correct or not present, so I was spending more time thinking about what to write. I don’t entirely trust this kind of subjective judgement, but it was a data point. Around the same time I wrote some code without clangd set up and that really hurt. It turns out I really rely on AST-aware completion to explore APIs. I had to look up more things in the documentation. Copilot was never good for this because it would just bullshit APIs, so something showing up in autocomplete didn’t mean it was real. This would be improved by using a feedback system to require autocomplete outputs to type check, but then they would take much longer to create (probably at least a 10x increase in LLM compute time) and wouldn’t complete fragments, so I don’t see a good path to being able to do this without tight coupling to the LSP server and possibly not even then.

Yesterday I was writing bits of the CHERIoT Programmers’ Guide and it kept autocompleting text in a different writing style, some of which was obviously plagiarised (when I’m describing precisely how to implement a specific, and not very common, lock type with a futex and the autocomplete is a paragraph of text with a lot of detail, I’m confident you don’t have more than one or two examples of that in the training set). It was distracting and annoying. I wrote much faster after turning it off.

So, after giving it a fair try, I have concluded that it is both a net decrease in productivity and probably an increase in legal liability.

Discussions I am not interested in having:

You are holding it wrong. Using Copilot with this magic config setting / prompt tweak makes it better. At its absolute best, it was a small productivity increase, if it needs more effort to use, that will be offset.
This other LLM is much better. I don’t care. The costs of the bullshitting far outweighed the benefits when it worked, to be better it would have to not bullshit, and that’s not something LLMs can do.
It’s great for boilerplate! No. APIs that require every user to write the same code are broken. Fix them, don’t fill the world with more code using them that will need fixing when the APIs change.
Don’t use LLMs for autocomplete, use them for dialogues about the code. Tried that. It’s worse than a rubber duck, which at least knows to stay silent when it doesn’t know what it’s talking about.

The one place Copilot was vaguely useful was hinting at missing abstractions (if it can autocomplete big chunks then my APIs required too much boilerplate and needed better abstractions). The place I thought it might be useful was spotting inconsistent API names and parameter orders but it was actually very bad at this (presumably because of the way it tokenises identifiers?). With a load of examples with consistent names, it would suggest things that didn't match the convention. After using three APIs that all passed the same parameters in the same order, it would suggest flipping the order for the fourth.

#GitHubCopilot #CHERIoT

Mx Autumn :blobcatpumpkin:

@david_chisnall your experience is how I expected mine to be if I had actually given the technology a chance.

Machine learning has been useful for decades, mostly quietly. The red flag against LLMs for me (aside from the authorship laundering and mass poaching of content) was there scramble by all companies to shoehorn it into their products, like a solution looking for a problem; I’ve yet to see it actually solve.

David Chisnall (*Now with 50% more sarcasm!*)

@carbontwelve I used machine learning in my PhD. The use case there was data prefetching. This was an ideal task for ML, because the benefits of a correct answer were high and the cost of an incorrect answer were low. In the worst case, your prefetching evicts something from cache that you need later, but a 60% accuracy in predictions is a big overall improvement.

Programming is the opposite. The benefits of being able to generate correct code faster 80% of the time are small but the costs of generating incorrect code even 1% of the time are high. The entire shift-left movement is about finding and preventing bugs earlier.

Alaric Snell-Pym

@carbontwelve @david_chisnall this also matches my expectations, and I've seen people mention studies in teams showing no productivity gain, too.

So I'm intrigued by the few people who DO report that LLMs help them code, though (eg @simon ). Is there something different about how their brains work so LLMs help? Or (cynically) are they jumping on the bandwagon and trying hard to show the world they've cracked how to use them well, to sell themselves as consultants or something?

Simon Willison

@kitten_tech @carbontwelve @david_chisnall I'm actually getting more coding work done directly in the Claude and ChatGPT web interfaces and apps vs using Copliot in my editor

The real magic for me at the moment is Claude Artifacts and ChatGPT Code Interpreter - I wrote a bunch about Artifacts here: https://simonwillison.net/tags/claude-artifacts/

Here are all of my general notes on AI-assisted programming: https://simonwillison.net/tags/ai-assisted-programming/

David Clarke :tinoflag:

@david_chisnall @carbontwelve this is what has been gnawing at the back of my brain. The purveyors of LLM's have been talking up the latest improvements in reasoning. A calculator that isn't 100% accurate at returning correct answers to inputs is 100% useless. We're being asked to conflate the utility of LLM's with the same kind of utility as a calculator. Would we choose to drive over a bridge designed using AI? How will we know?

David Chisnall (*Now with 50% more sarcasm!*)

@zebratale @carbontwelve Calculators do make mistakes. Most pocket calculators do arithmetic in binary and so propagate errors converting decimal to binary floating point, for example not being able to represent 0.1 accurately. They use floating point to approximate rationals, so collect rounding errors for things like 1/3.

The difference is that you can create a mental model of how they fail and make sure that the inaccuracies are acceptable within your problem domain. You cannot do this with LLMs. They will fail in exciting and surprising ways. And those failure modes will change significantly across minor revisions.

Stephen J. Anderson

@simon @kitten_tech @carbontwelve @david_chisnall How would you avoid or deal with the issues that David encountered? Specifically, subtle bugs that the process of debugging make the whole process less efficient than writing it yourself. Is there one of your notes that deals with that already?

Glitzersachen.de

@david_chisnall @zebratale @carbontwelve

"do make mistakes" I wouldn't call that a mistake. The calculator does what it should do according to the spec how to approximate real numbers with a finite number of bits.

It's (as you explain) a rounding error. A "mistake" is what Pentiums with the famous Pentium bug made.

But maybe it's my understanding of English (as a second language) that is at fault here.

Pendell

@glitzersachen @david_chisnall @zebratale @carbontwelve the calculator /is/ doing exactly what it's been programmed to... and it is programmed to make specific and defined "mistakes" or errors in predictable and clear cut ways in order to make the pocket calculator run on as little power as possible.

An LLM, likewise, is also doing exactly what it was programmed to do... and that is to spew regurgitated nonsense it read off the internet.

Simon Willison

@utterfiction @kitten_tech @carbontwelve @david_chisnall you have to assume that the LLM will make weird mistakes all the time, so your job is all about code review and meticulous testing

I still find that a whole lot faster then writing all the code myself

Here's just one of many examples where I missed something important: https://simonwillison.net/2023/Apr/12/code-interpreter/#something-i-missed

Simon Willison

@utterfiction @kitten_tech @carbontwelve @david_chisnall but honestly, the disappointing answer is that most of this comes down to practice and building intuition for tasks the models are likely to do well vs mess up

Manipulating some elements in the HTML DOM with JavaScript? They'll nail that every time

Implementing something involving MDIO registers? My guess is there are FAR less examples relating to that in the (undocumented, unlicensed) training data so much more likely to make mistakes

Martijn Faassen

@kitten_tech

@carbontwelve @david_chisnall .

Note that how @simon reports using this to generate little projects is an entirely different mode of working with them. I have used copilot for a few years now and like it myself, which is mostly context sensitive autocomplete.

A Q&A session to create code for a CLI tool or web app is a very different way of working I started exploring more recently. It's surprisingly capable for little projects and requires a different approach.

Simon Willison

@faassen @kitten_tech @carbontwelve @david_chisnall Steve Yegge calls it CHOP, for Chat Oriented Programming https://simonwillison.net/2024/Jul/12/the-death-of-the-junior-developer/

pasta la vida

@pendell @glitzersachen @david_chisnall @zebratale @carbontwelve floating point finance calculations is a common mistake...

pasta la vida

@pendell @glitzersachen @david_chisnall @zebratale @carbontwelve programmers and CPU designers are just a tad sensitive and insecure when someone points out the calculator makes a mistake and isn't mathematically perfect