| 8 Jun 2023 |
 cwfitzgerald | Man, the workflow of “write what you want in a comment then let github copilot fill it in” is wild. Very helpful for things like shell scripts where I know basically nothing | 22:25:01 |
 kvark | cwfitzgerald: you embraced the dark side? | 23:14:27 |
| kvark | Ralith: it has types! | 23:14:40 |
| cwfitzgerald | In reply to @kvark:matrix.org
cwfitzgerald: you embraced the dark side? coworker was using it and it was cool - and they are giving it out for free to OSS maintainers | 23:15:08 |
| cwfitzgerald | it's more useful than it is annoying | 23:15:46 |
| cwfitzgerald | especially with receptive things or boilerplate | 23:16:10 |
| kvark | Copilot will make people forget how to code. Deep reinforcement learning will make us forget how to design algorithms. What’s next? Game::run()? | 23:17:53 |
| cwfitzgerald | In reply to @kvark:matrix.org
Copilot will make people forget how to code. Deep reinforcement learning will make us forget how to design algorithms. What’s next? Game::run()? With the product that GitHub is shipping right now I totally disagree. It has zero human context to what it's doing. The core operation that it's doing is "hallucinate what code that goes here might look like". It does a good job of taking into account physical context surrounding that code, but has no idea of the purpose of the code | 23:34:23 |
| cwfitzgerald | If you look at one of the examples of one of the AI systems generating a whole running bit of code comma the code it's generating is hilariously unoriginal and boring, and has no kind of Greater structure | 23:36:28 |
| cwfitzgerald | I think the AI hypebros are doing the technology a significant disservice by toting it as this super powerful thing that can write whole programs for you | 23:37:42 |
| cwfitzgerald | It is super cool, and does have power, but it's inability to generate working code most of the time means that you need to know what you're doing to get any significant use out of it | 23:38:57 |
 Ralith | In reply to @kvark:matrix.org
Ralith: it has types! nix modules have types already, and it's not obvious that nickel is applicable much more broadly than those | 23:40:55 |
| Ralith | idk, hopefully it'll be good, haven't studied it closely | 23:43:01 |
| Ralith | just kind of wary/weary of Yet Another Configuration Language | 23:43:11 |
| cwfitzgerald | * If you look at one of the examples of one of the AI systems generating a whole running bit of code, the code it's generating is hilariously unoriginal and boring, and has no kind of greater structure | 23:45:26 |
| cwfitzgerald | Additionally, the quality of github copilot's suggestions is extremely dependent on the clarity of the surrounding code, and what exactly you're asking of it - if you're extremely consistent with naming, write lots of comments, and are doing "expected" code, it's alright | 23:48:23 |
| 9 Jun 2023 |
| Ralith | I do really like ML autocompletion for boilerplate | 00:57:01 |
 David Huculak | I loved copilot while I had temporary free access. It accidentally helped me translate some UI strings too lol. Despite growing up in Quebec it was better at writing french than me despite being apparently tuned for writing code?
I'm undecided about the question of whether the model has a sense of purpose or understanding. As far as I'm concerned it 'knows' that functions need to have inputs and outputs and that they need to get plugged into code elsewhere. Does it 'understand' what the purpose of the function is? Idk. And I think it will get more unclear once it's able to give correct output to the current adversarial inputs that are used to 'prove' that it doesn't really know what's going on | 01:25:31 |
| cwfitzgerald | I feel from it's use, it's pretty clear it doesn't actually understand anything, it's just infilling, it's just very good at judging the physical context | 01:38:05 |
| cwfitzgerald | * I feel from it's use, it's pretty clear it doesn't actually understand anything, it's just infilling, it's just very good at judging the varying importance of physical context | 01:38:13 |
| kvark | who needs good concise APIs if you can just ask ML to fill out the boilerplate | 01:39:46 |
| cwfitzgerald | even decent apis have decent amounts of boilerplate looks at a certain project of ours 😆 | 01:40:42 |
| cwfitzgerald | also lmao italicized emoji | 01:41:22 |
| 10 Jun 2023 |
 fl33tw00d | In reply to @kvark:matrix.org
Copilot will make people forget how to code. Deep reinforcement learning will make us forget how to design algorithms. What’s next? Game::run()? I agree with this 100%. I also agree with what everyone else is saying in regard to some of the brain dead boilerplate that copilot fills in being really handy. | 14:16:41 |
| fl33tw00d | Which leads me to a cool idea! They should allow you to tune the maximum allowed code complexity of samples generated by copilot using something like cyclomatic complexity | 14:18:15 |
| fl33tw00d | that way I can get the braindead autofill without forgetting the hard bits | 14:18:41 |
| kvark | that would be interesting
but I'm not worried about you or me forgetting how to design algorithms
I'm worried about the bigger picture - this becoming an art, a useless exercise for selected few.
Useless because whatever beauty you design is not going to be on par with AI's outputs. | 15:07:48 |
| kvark | Reddit exodus is a hot topic these days. What do people recommend?
I'm considering Lemmy, but it feels clunky and unpopulated rn | 19:40:32 |
| kvark | e.g. you look at the most popular server, and it's top post is basically "we are overloaded, go away". Not very tempting | 19:41:37 |
 m4b | myspace | 19:44:24 |
