Ivan Brko

AI is now perfectly capable of reverse engineering a tech-related book from all the public repos of people that solved the exercises. I started using AI to clarify things I didn’t understand in a book I was following along and I realized it could figure out what each chapter was teaching just by looking at those repos. You can just have AI teach you the entire book without ever buying it

Dio publike na novom Maksimiru ce imati spektakularan pogled na nekih 30% terena

claude code is getting way to slow, telling it to do something like 3 git commands and it takes 3 minutes

Fable did absolutely terrible job while attempting to write prompts for one of my agentic applications, yet to far a process that works better than writing them manually.

Shopify does breaking changes in their API releases like they're a hobby vibe-coded project 3 people use

OpenAI publishes a story meant to show how capable their model is and the takeaway for most people is that OpenAI is doomsdaying unnecessarily and that you can't use OpenAI's own models to defend yourself against cyberattacks but you have to resort to Chinese models. An own-goal

I really wonder if they displayed Veiga's contact with the ball in the graph (and they definitely have the space in the x-axis in the image), would it scale y-axis to a point where Matanovic's hair touching the ball would not even be visible.

Trying to get OpenClaw to send me a reminder in the morning and it has been failing to do so for 4 days now... Amazing

wonder how long it's gonna take till data sets containing feature estimates that agents are trained on get up to date. claude & codex warn me not to do something cause it's a big rewrite that takes weeks and then implement it in 12 minutes. it's amazing that you can just open a new branch and turn the product upside-down just to test an idea in 1 hour.

safety classifier in claude code is a great idea, but i wish they gave a nudge or something when it's not working or go back to default mode. it just halts the agent progress silently at the moment, it can't get permission from the classifier so it stops the work without notice.

most of the vibe-coded apps i see have too much features, barriers to just implementing something as soon as it crosses your mind have become non-existent.

anyone else having issues with claude code? my usage just jumped to 100% for this session even though i don't feel it should have + there were no usual warnings at 90/95%.

For a product that costs 100-200 $/month, Claude Code should have much better uptime. Can't use it for almost two hours now cause they messed up login.

Added native support for Codex CLI to Panoptes in the latest release, so it's as easy as it gets to work on the same project with Claude Code and Codex. github.com/ivan-brko/panopte… The easiest way to juggle between different projects/worktrees/agents from CLI, by far

The more I use Claude Code and Codex, the less I need IDEs. Not sure if I even need my JetBrains subscription anymore, VSCode with LSP is more than good enough to review the generated code. It's gonna be a shame if this is the end of great IDE experience.

Is there a consistent way to get claude code not to try making changes? I want to investigate something, grant all permissions so I can just come back to the report, but even in planning mode cc often just comes up with an implementation plan (after being asked not to).

Claude Code tasks are amazing. From 0 through MVP to usable project in matter of hours.

Vibe coding with Claude Code for the last couple of days, since it got included in the Pro plan. I dislike using VSCode clones for non-TS/JS projects and this is by far the best AI coding experience I've had so far when I use JetBrains IDEs.