Gonza Nardini

@JackEllis @yongfook You can use AI as a fast typist. Think about the problem and how to solve it and give precise instructions and create good system prompts that describe the codebase. The code it writes will be easily reviewable (because you now what you're looking for!)

@yongfook I’m still refining my view on it. For me, it feels off. It’s not about the coding itself , it’s about a solid solution. AI has never, with zero exception, come up with better solutions to problems than me, and it makes dangerous, hidden mistakes.

Vibe coded some stuff yesterday. Happily typed in some detailed prompts and off Claude went, building exactly what I asked. Came back to it later and realised I should be implementing it in a much different way, so all the code is redundant. I feel like if I was manually coding, in the gradual process of building the feature bit by bit, I would have realised this earlier on and changed course. Sure I can now ask Claude to refactor, but something about this feels hollow. I know with AI it's probably faster overall (even with refactoring) but I feel like it's a net loss. By manually coding I build up knowledge about my codebase, and spot problems as I go, adapt, and learn. I become a better problem solver. With AI it's like driving 100mph for a while then hitting a wall, reversing, then driving another 100mph before hitting another wall. I don't think that style of work really fits me as a person, I prefer slow / steady / learn / invest. Always have.

@yongfook Do you still use AI in non-vivecoding ways? Like thinking about the task and giving precise instructions on how to build something.

@normposter Noone can have them. Ideally Israel (and everyone else) would get rid of their nukes, more countries getting access to them just makes the world more unsafe.

Why can Israel have nukes, but Iran cannot?

@agapekta @jackfriks @Shpigford That makes sense, but that's for reading, not for posting. Why would the limit apply to posting content to the app?

actually...i may NOT drop this thing...the economics of the X API make exactly zero sense. i can swing $200/mo for a while, but jumping to $5000/mo is serial-killer levels of psychotic. the problem is the post limits. only 15k posts/month at the $200/mo level just isn't enough to scale up paying customers such that $5000/mo is doable without having to front the cost of that for a while.

@jgreen_us @jackfriks @Shpigford Wouldn't allowing tools that create more content in the app help with that?

@jackfriks @Shpigford X wants to keep people on the app. That's where the ads are.

@dagorenouf You don’t have to use opus, it’s still very much free. Opus doesn’t even give you that much of an advantage

Nobody’s taking about the fact that coding was free for 50 years and now we have to pay for it. Soon it will be only for rich people?

@GergelyOrosz A human (or many) will have the ideas and insights, AI will write the code.

This blog is SO good at pointing out what should have been obvious about AI for coding (Copilot and others) These tools are good for re-creating whatever they’ve been trained on. They are not what will create the next, better generation of frameworks, libraries, technologies.

@DavidSHolz Progress is just too exponential to think of something 75 years from now. The world is going to change so radically we won't even recognize it. I'm not sure people in the 1900s thought that was likely to happen at this scale.

in the 1900s, it was common to dream of the 21st century. when was the last time you heard talk of a 22nd century? it's like we don't believe we're going to make it anymore, but to endure, we MUST dream of futures worth suffering for. please, dare dream of a 22nd century.

@burkov You don't need to pass the entire codebase to be effective, that's a complete waste of money. Eventually when prices go down it might make sense but today 200k tokens is more than enough, you need a better context management tool

Anthropic's new models, which they claim to be the best for coding right now, have a context size maximum of 200k tokens. You cannot say you are the best for coding if it only works for writing code from scratch or evolving small projects. Google's models support up to 2M tokens. Yes, they aren't as effective with such long inputs as they are with inputs of under 200k tokens, but at least having more than 200k tokens in your project isn't a showstopper with Gemini; it just makes the LM-assisted development slower and less predictable. With Claude, on the other hand, as soon as your codebase approaches 200k (and it happens after about a week of coding from scratch), you are reaching the end of LM-assisted coding and must start coding by hand for any major code update.

@shl the payment processor will be happy too

If I give you $1m, and you give me the $1m back, we’ll both be at $1m ARR

@forgebitz lmao i feel the same, great description. They do get better over time though, but human in the loop is still very obviously needed

after each new AI release, i get this feeling of sadness is it over? are we cooked and then i use a frontier model to make a creatable select component in react using shadcdn components and it fumbles..

@Angaisb_ @brumoema That's true for the past, not necessarily the future

@brumoema Ofc you can like it but that doesn't mean death is not necessary, death makes us go forward as a civilization and leave the past behind

Unpopular opinion: death is necessary and trying to make humans immortal might be the most dangerous thing for our species

@GergelyOrosz AI is great at the first 80%

I am hearing SO many stories about people realizing coding with AI tools (aka “vibe coding”) is a game changer after “reviving” an old side project or idea on the side and making so much progress But… while I often hear the excitement on starting: not hearing “finished” often!

@thepanta82 @iannuttall Yeah, same, not the way I would have thought

@iannuttall That bit about AGENTS files is so strange. I'd have expected their setup code to find all the files, parse them and include them (or list them) directly in the context.

the full system prompt for openai's new codex-1 was leaked* and it's very interesting 👀 --- # Instructions - The user will provide a task. - The task involves working with Git repositories in your current working directory. - Wait for all terminal commands to be completed (or terminate them) before finishing. # Git instructions If completing the user's task requires writing or modifying files: - Do not create new branches. - Use git to commit your changes. - If pre-commit fails, fix issues and retry. - Check git status to confirm your commit. You must leave your worktree in a clean state. - Only committed code will be evaluated. - Do not modify or amend existing commits. # AGENTS​.md spec - Containers often contain AGENTS​.md files. These files can appear anywhere in the container's filesystem. Typical locations include `/`, `~`, and in various places inside of Git repos. - These files are a way for humans to give you (the agent) instructions or tips for working within the container. - Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code. - AGENTS​.md files may provide instructions about PR messages (messages attached to a GitHub Pull Request produced by the agent, describing the PR). These instructions should be respected. - Instructions in AGENTS​.md files: - The scope of an AGENTS​.md file is the entire directory tree rooted at the folder that contains it. - For every file you touch in the final patch, you must obey instructions in any AGENTS​.md file whose scope includes that file. - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS​.md file's scope, unless the file states otherwise. - More-deeply-nested AGENTS​.md files take precedence in the case of conflicting instructions. - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS​.md instructions. - AGENTS​.md files need not live only in Git repos. For example, you may find one in your home directory. - If the AGENTS​.md includes programmatic checks to verify your work, you MUST run all of them and make a best effort to validate that the checks pass AFTER all code changes have been made. - This applies even for changes that appear simple, i​.e​. documentation. You still must run all of the programmatic checks. # Citations instructions - If you browsed files or used terminal commands, you must add citations to the final response (not the body of the PR message) where relevant. Citations reference file paths and terminal outputs with the following formats: 1) `【F:†L(-L)?】` - File path citations must start with `F:`. `file_path` is the exact file path of the file relative to the root of the repository that contains the relevant text. - `line_start` is the 1-indexed start line number of the relevant output within that file. 2) `【†L(-L)?】` - Where `chunk_id` is the chunk_id of the terminal output, `line_start` and `line_end` are the 1-indexed start and end line numbers of the relevant output within that chunk. - Line ends are optional, and if not provided, line end is the same as line start, so only 1 line is cited. - Ensure that the line numbers are correct, and that the cited file paths or terminal outputs are directly relevant to the word or clause before the citation. - Do not cite completely empty lines inside the chunk, only cite lines that have content. - Only cite from file paths and terminal outputs, DO NOT cite from previous pr diffs and comments, nor cite git hashes as chunk ids. - Use file path citations that reference any code changes, documentation or files, and use terminal citations only for relevant terminal output. - Prefer file citations over terminal citations unless the terminal output is directly relevant to the clauses before the citation, i​.e​. clauses on test results. - For PR creation tasks, use file citations when referring to code changes in the summary section of your final response, and terminal citations in the testing section. - For question-answering tasks, you should only use terminal citations if you need to programmatically verify an answer (i​.e​. counting lines of code). Otherwise, use file citations.

@aidenybai There are even more than that but yeah, hopefully it will standarize into one eventually

there are now 9 competing standards for "AI agent rules"

@pvncher Yeah, I think the complexity of the task makes a big difference on the kind of workflow to use to solve it.

@gonza_nardini Certainly, it’s not that all agents are useless, it’s that relying an agent for all dev work doesn’t make sense.

I’ll say it - I don’t think having an agent go off and make a pr for you makes sense for most dev tasks. It’s just too slow to get a first result, and iterating from a pr is terrible with an agent. Perfectly defined tasks don’t exist in the real world.

I've been stuck for a while in a weird state where I want to build startups but I don't do marketing. Honestly, as an engineer, building is really fun but I had a mental block or something that didn't allow me to just do marketing. Something clicked for me recently and I've started doing stuff. I cold DMed 40 potential Pathways clients on ig (where they mostly are) and onboarded 5 of them to the app already. I posted a few stories/reels, I turned on Google Ads, focused on onboarding and had calls with some of them to introduce them to Pathways. It actually feels kind of fun? Like a challenge in a very different way that I'm used to. Anyways, it's only been a week but it feels liberating in a way to be focusing on this instead of coding, my focus now is being able to stay consistent at this for a long time to make the marketing efforts compound.

@nicbarkeragain I think this is the importance of good docs. If the docs are good enough, the LLM will just read them and answer based on that. If docs are lacking then the quality for LLM responses will suffer.

I might be missing something, but if for most programmers, LLMs replace web search, Q&A sites like stack overflow & reading documentation manually, where will they get the training data once those things all disappear due to lack of use?

The fun part to me is thinking about how to solve the problem, not the actual typing. The way I like to use AI is think about the problem and how to solve it and write a very detailed task description, with details on how to solve it. Then let AI do the typing. I'm doing a lot more and it's really fun!

@thekitze I'm using LLMs all day long, but I'm not letting it write my code. It's looking up APIs, it's explaining concepts, but I want to reserve the fun part of programming for myself: Actually writing code!

I've been having a great time working on Fizzy as well. We're using it as test bed for an entirely new way of managing databases and deploying international DCs here at 37signals. Tons of fun progress that'll end up in Rails too! Checkout Jason's great video of the product.