Brian Cooley @blcooley

@sudoingX Anything that improves sleep. A great bed and a sleep mask are a couple things I value greatly

About 2/3 of the time it's better to just invest all at once, but you really need to consider the timeline and personal goals to know if you can handle the risk. Given your age (relatively young) and monthly income (a lot), it might have made sense to be more aggressive. You had a worse outcome this time, but I wouldn't consider the dollar cost averaging approach universally bad.

@svpino /simplify is available as a skill and it kind of does the same thing.

@svpino It already looks a lot different from what it used to, but I think whatever new roles will result will largely be filled by current day software engineers with some design/PM skills and the more technical PMs.

@hector_iosdev @AppStore I'm not accusing you of making a mistake, I'm trying to parse out why their automation made a mistake. Good luck with it.

Eventually local models will get good enough that something will have to give. I recently read something about people repurposing their OpenClaw Mac minis to run Gemma 4 31B or Gwen 3.5 27B locally. Haven't tried either, and have no idea about how good they are. But clearly local capabilities are improving. This was a total no-go not long ago.

@indiesoftwaredv A solution would be to put a trial on the shortest term/higher cost per unit time subscription. If users want the better long term deal they can change after the trial. That's how I do it anyway.

One of the things I really like about AI coding is the ease at which you can plan out large refactors of parts of a codebase, including how to do it step-by-step. The goal being that at the end of each incremental step you have a working codebase with all relevant tests passing and probably some new tests as well. This is how I'd do it manually, but it feels very painstaking that way. AI makes a lot of the tedium easier to deal with and I can focus on correctness. The AI can probably one shot it with enough planning and feedback, but the cost of wholesale changes like that is often an overwhelming PR review and a complete regression test.

@GergelyOrosz I feel like I'm doing a much better job of planning out the work. At my day job anyway. Side projects are a mess, though. I don't seem to finish anything. Always been a problem, but AI is very bad for shiny object syndrome. Too easy to start several things at once.

@luc_moetwil Mostly this startup I'm working for. It's a roller coaster and easy to get discouraged. Also my Interviewer app, though that's been going better lately.

What if it all works out? Now, there's an exciting thought!

@nbaschez Code review is like this. I give the model an out by telling it that approving the code is also a valid result and to catégorisé the severity of findings. Then if I run the model against the suggestions I allow it to ignore any it deems trivial or handled in another way

@svpino Truthfully, it works to some degree. It wouldn't be (and isn't!) my only review, but it absolutely catches mistakes that were made in code generation. I often use Claude Code with a clean context (or in gh actions) as a first review pass. Even better with a heterogeneous model.

Your existing technical debt is also agent debt when you are using AI agents for coding. Worse, agents will just build more debt on top. Probably need to be spending as much or more time refactoring to keep a project from getting out of hand.

But it used to take me days to write code riddled with bugs! Now I do it in minutes. In all seriousness I really like Jeremy Howard's approach with Solve It (?) to use AI to help you learn and level up, not just have it write code for you. Being able to get good answers to as many questions as you want to ask about a code base and use that to reify your knowledge feels a super power.

If you go back to Karpathy's original post that defined it, that's exactly what he said. But now it has kind of taken a life of its own to mean any of a number of things, including being used as a term for any AI generated code. Personally I think it's useful to separate "vibe coding" from "agentic engineering". I think eventually we are headed for a reality where it's all just called "software engineering" again and AI is just a very important tool in the toolbox. And maybe what you are engineering is agentic workflows instead of the code itself.

@CtrlAltDwayne 95%+ There are certain things it's just faster to do myself. I'm not asking Claude Code to change the value of a text literal or change a padding by 10 pixels.

If you want to keep AI from breaking existing code, you need to give it something to eval against. Lots and lots of behavioral unit tests, integration tests, screenshot tests, end-to-end tests, etc. are pretty much the only way I know of. You will also use log messages, linters, and making sure the build compiles, but it's not enough. More of a first line of defense. Unfortunately, creating this kind of eval suite for coding isn't something you can easily do without knowing how to code, so it's a real bootstrapping problem.

@_trish_xD Commodore BASIC 2.0 on a VIC-20.

That's really cool and a very nice post. It reminds me of learning all the integration methods like Runge-Kutta and Adams when I was studying optimization in grad school. The most interesting thing to me is how the tradeoff of accuracy and usefulness. A lot of times in academia you publish a paper about a 2% improvement in some accuracy or performance metric, but in the real world you can be severely limited by things like available memory or the latency of the computation itself.