251–260 of 333 posts
Earlier quoted context omitted.
Once it writes the code, take that into a new session to fix a bug. Repeat with new sessions. Don’t let it read the buggy code, it will just get worse.
How do you not let it read the buggy code but also take it into a new session?
Watching the examples and other people's demos, I get the sense that this product completely ignores what makes good software. It has no "big picture" plan or contextual awareness required to make those nuanced yet critical decisions in the far ends of applications. Those cases where you badly need domain expertise, understanding of how the user interfaces with the product, awareness of existing technical debt, and s…
The velocity is good enough to scale back the size of the team, it's good enough to mandate a pace that is only possible with AI doing automated submission and review of PRs, it's good enough to not have any kind of formal design work. Just ship.
After several months of this, any additional bug fix by the AI is just adding to the existing morass of tech debt created by the AI. Slowly but surely, the AI will stop being able to submit a PR that doesn't result in a regression. The people who actually care about software quality aren't on new projects; those are run by juniors and copilot. Nope, they're going to be stuck on maintenance of this garbage.
And meanwhile, that new team working on new feature X is moving so quickly... better give them a raise!
The people who suffer here are, of course, the non-marginal user, who will have to contend with every new feature breaking every other existing feature that they like in the software.
Earlier quoted context omitted.
I had ChatGPT output an algorithm implementation in Go (Shamir Secret Sharing) that I didn't want to figure out. It kinda worked, but everytime I pointed out a problem with the code it seemed more bugs were added (and I ended up hating the "Good catch!" text responses...) Eventually, figuring out why it didn't work made me have to read the algorithm spec and basically write the code from scratch, throwing away all of…
The skill in using an LLM currently is in getting you to where you want to be, rather than wasting time convincing the LLM to spit out exactly what you want. That means flipping between having Aider write the code and editing the code yourself, when it's clear the LLM doesn't get it, or you get it better than it does.
Earlier quoted context omitted.
I had ChatGPT output an algorithm implementation in Go (Shamir Secret Sharing) that I didn't want to figure out. It kinda worked, but everytime I pointed out a problem with the code it seemed more bugs were added (and I ended up hating the "Good catch!" text responses...) Eventually, figuring out why it didn't work made me have to read the algorithm spec and basically write the code from scratch, throwing away all of…
An alterinative to this workflow that I find myself returning to is the good ol' nicking code from stackoverflow or Github. ChatGPT works really well because the stuff you are looking for is already written somewhere and it solves the needle-in-the-haystack problem of finding it, very well. But I often find it tends to output code that doesn't work but eerily looks like it should, whereas Github stuff tends to need a…
I only find current LLMs to be useful for code that I could easily write, but I am too lazy to do so. The kind of boilerplate that can be verified quickly by eye.
Earlier quoted context omitted.
I wonder if AI tools is going to kill editor diversity. Every one of these tools VS Code is the first class citizen and other editors are an afterthought. Sure people can write their own Emacs package but that's only if the tool developers enable it and the experience is usually not as good as the official VS Code version. I can also see the future where not using VS Code is a signal you are less reliant on AI tools…
This is a great fear of mine as a (neo)vim user. I've repeatedly encountered situations already where it was assumed everyone was using VS code, and I only see that getting worse as more and more integrations get built into IDEs.
Earlier quoted context omitted.
This is a great fear of mine as a (neo)vim user. I've repeatedly encountered situations already where it was assumed everyone was using VS code, and I only see that getting worse as more and more integrations get built into IDEs.
There are lots of hackers bringing VsCode-only extensions to Neovim. I brought Copilot Chat to neovim back when it first became available to me and the community is still alive and well, though with a different maintainer. Tbh AI tools aren’t that useful, and those that are will get ported. Currently trying to bring Cursor to Neovim but that has proven much more difficult
Do you know why Github didn't make an official version? I'm definitely going to give yours a try, and I don't really care whether it's "official" or not, but rather I'm curious because looking for official neovim support is a useful signal of their priorities/intentions.
Earlier quoted context omitted.
Creating this using Open Source LLMs would be like saying you tried A5 Wagyu by going to Burger King, respectfully. I think benchmarks are severely overselling what open source models are capable of compared to closed source models.
I really don't think they're being over sold that much. I'm running llama 3 8b on my machine, and it feels a lot like running claude 3 haiku with a much lower context window. Quality wise it is surprisingly nice.
Github is likely using a GPT-4 class model which is two (massive) steps up in capabilities in Anthropic's offerings alone
Earlier quoted context omitted.
I had ChatGPT output an algorithm implementation in Go (Shamir Secret Sharing) that I didn't want to figure out. It kinda worked, but everytime I pointed out a problem with the code it seemed more bugs were added (and I ended up hating the "Good catch!" text responses...) Eventually, figuring out why it didn't work made me have to read the algorithm spec and basically write the code from scratch, throwing away all of…
Once it writes the code, take that into a new session to fix a bug. Repeat with new sessions. Don’t let it read the buggy code, it will just get worse.
Watching the examples and other people's demos, I get the sense that this product completely ignores what makes good software. It has no "big picture" plan or contextual awareness required to make those nuanced yet critical decisions in the far ends of applications. Those cases where you badly need domain expertise, understanding of how the user interfaces with the product, awareness of existing technical debt, and s…
Having something round up the likely areas of the codebase that needs touching feels magical. It doesn't always succeed! But it feels pretty magical to get that boost when you're new to some part of a codebase (which, real talk, code I wrote > 1 month ago, I must page back into memory).
Making it easy for me to progressively add context for the model is an accurate analogue for how I think as a developer when tackling a task. I have to build a mental model of how things work. And then a plan for how I'm going to change it.
Maybe for the kinds of tasks you usually tackle, it won't have value. But the amount of context it's attempting to bring to bear on whatever task you give it is categorically more — and better — than any other tool I've seen. I have seen (and been the author of) spaghetti. Could I make CW generate spaghetti? Surely. That's why it's a tool for developers, not a substitute for developers.
Earlier quoted context omitted.
Don't forget that this is marketing. "You'll be the best cook if you buy the Mega Master Automated Kitchen Appliance (with two knives included)" That line is marketed at me, who does not know how to cook, they're telling me I'll be almost a chef. You'll hear Jensen say that coding is now an obsolete skill, because he's marketing the capabilities of his products to shareholders, to the press. It might well be that in…
All the famous chefs didn't become famous from their cooking. They became famous because of their charisma. Jamie Oliver looked really good on camera. AI will never be able to bullshit the way humans can.