What I’m curious about is how well LLMs do when they create something from scratch, because so far my experience was with letting it fix issues or add features to existing codebase where I already shaped the general architecture and put in a lot of guardrails. But what if the architecture is unclear and there is nothing letting agent know if change breaks something or not?
My only experience with tiny codebase where it did a lot of scaffolding was poor - it did what I asked for, not what I needed. If i did more of the thinking myself I would realise it’s a code that works but doesn’t solve the problem I’m after.
It’s not like this is a factory floor where you process something coming in and AI suddenly makes the process more efficient and people are idle.
Every team in tech world has infinite backlog, you don’t fire 20% the minute someone manages to close a few tickets.
They own Eufy which sells cameras with main feature being “no subscription needed”, that are very unreliable and full of ads (which isn’t being advertised as much as lack of subscription). They do also go big on labelling a lot of simple features as AI where in reality it’s something as simple as “detect a person in a photo”.
I have Eufy cameras and it’s complete garbage, sadly competition is also mostly garbage.
Bold unsustainable claims at st the core of their business, it’s not just thumbnails.
That’s my daily experience too. There are a few more behaviours that really annoys me, like:
- it breaks my code, tests start to fail and it instantly says “these are all pre existing failures” and moves on like nothing happened
- or it wants to run some a command, I click the “nope” button and it just outputs “the user didn’t approve my command, I need to try again” and I need to click “nope” 10 more times or yell at it to stop
- and the absolute best is when instead of just editing 20 lines one after another it decides to use a script to save 3 nanoseconds, and it always results in some hot mess of botched edits that it then wants to revert by running git reset —hard and starting from zero. I’ve learned that it usually saves me time if I never let it run scripts.
> they're doing as well as professionals do without oversight on production environments
The difference is that if a human does it there usually is done accountability, you’ll be asked how it happened and expected to learn from it. And if you do it again your social score goes down, nobody will trust you and you’ll be consider a liability.
If a cli tool does it the outcome is different, you might stop saying the tool or you might blame yourself for not giving the tool enough context. And if it does it again you might just shrug it off with “well of course, it’s just a tool”.
There’s also entitlement from just using it: my org uses your software and there’s what we consider is a bug so you MUST fix it as asap as possible and in the future don’t release buggy software because it costs us time and money.
My biggest annoyance with gitlab is that the UI is one huge block of text on white background with absolutely no distinction between user content (like commit messages) and the interface itself, plus you have action buttons buried in the middle of the page.
And everything loads asynchronous causing blocks to dance around when the one above loads …
There are still a few things missing from all models: taste, shame and ambition.
Yes they can write code, but they have no idea what needs does that code solve, what a good UX looks like and what not to ship. Not to mention that they all eventually go down rabbit holes of imaginary problems that cannot be solve (because they’re not real), and do where they will spend eternity unless w human says stop it right now.
Adding more LSP features to the jinja linter for saltstack that I wrote, so you can see all the bugs in your templates from VSCode (rather than waiting for CI) and do things like “rename this jinja variable everywhere it’s being used”.
I think it might be because with ad model you can sell profiling data many times over to different parties. You can’t do the same with a single charge.
> I'm not going to pay $5/month for every blog that I occasionally read
Would you pay per view? Most people (me included) would probably hesitate to say yes, because we’re used to not paying for that. But what if it meant that ad based model is gone and everything you buy is cheaper because the price does not include the cost of running ads?
IMHO in a way the constant churn of Grafana made it easier to live with it. When they break your dashboards every major release you just learn to let go. And by break it’s not just making it error and not work, it’s the constant moving of things around, refactoring the UI, replacing one component with another, all accompanied by a number of glitches every time they rewrite things. You just accept and ignore it eventually.
What’s a bigger lock in for me is metrics and promql - you just can’t ever rename a poorly named metric or you face a world of pain. Or when Prometheus releases Native Histograms to replace the old ones, and suddenly everything from rules, alerts, ad-hoc queries and dashboards needs updating.
And PromQL is so opaque, it just never give you an error unless there is a syntax issue. We need tools like https://github.com/cloudflare/pint just to know if my alert description isn’t trying to render a label that’s just not gonna be there, etc
The problem with AI isn’t new, it’s the same old problem with technology: computers don’t do what you want, only what you tell them.
A lot of PRs can be judged by how well they are described and justified, it’s because the code itself isn’t that important, it’s the problem that you are solving with it that is.
People are often great at defining problems, AIs less so IMHO. Partially because they simply have no understanding, partially because they over explain everything to a point where you just stop reading, and so you never get to the core of the problem. And even if you do there’s a good chance AI misunderstood the problem and the solution is wrong in a some more or less subtle way.
This is further made worse by the sheer overconfidence of AI output, which quickly erodes any trust that they did understand the problem.
If someone has a “readfile” cli app on gh it’s just a moment away from a issue being open:
> thanks for a great project, in our org we have a requirement that we only write files, not read them. Can you please add —-write flag so this app works for us?
The fact that someone clones your repo or uses your software doesn’t mean that you owe them anything. Every person with open source code should realise this before they start responding to feature requests.
I find the 2025 versions to be a nicer looking than pre-2025 variants, so it’s overall an improvement. But I also find the 2014 to be usually a lot better (clearer and more obvious). So incrementally it’s an improvement, but historically still worse.
> There seems to be a strong "instrument everything" culture
Metrics are the easiest way to simply expose your application internal state and then, as a maintainer of that service, you’re in nirvana. And even if you don’t go that far you’re likely to be an engineer writing code and when it comes time to add some metrics why wouldn’t you add more rather than less, and once you have all of them why not adding all possible labels? And in the meantime your Prometheus server is in a crash loop because it run if of RAM, but that’s not a problem visible to you.
Unfortunately there’s a big gap in understanding between a code editor writing instrumentation code and the effect in resource usage on the other end of your observability pipeline.
I think a lot of people, when they see an open source project that somehow fits into their current challenges, read it as the person realising the project saying “I alone now own this entire problem space” and so that GitHub repo is now just an extension of your jira board.
Well done, next maybe make the web ui usable because right now there’s absolutely no distinction between the UI itself and the user content, which IMHO combined with action buttons in the middle of the page makes for a really poor ux.
I often find Copilot summaries to be more or less an attempt at mainsplaining a simple change. If my tiny PR with a one line description requires Copilot to output a paragraph of text about it it’s not a summary, it’s simply time wasted on someone who loves to talk.
All those companies just want to provide value to users while breaking even on costs, it’s a shame they keep accidentally creating addictive products that break all user promises and keep adding more and more ways to squeeze out extra revenue.