I just can't help feeling these training wheels are getting in the way of my bicycle commute.

Hacker News is a community-driven platform for sharing and discussing technology news, startups, and programming-related topics. Through user submissions and comments, Hacker News offers insights into emerging technology trends, industry developments, and entrepreneurial ventures. Readers can participate in discussions, share their insights, and stay informed about the latest advancements in technology and innovation.

Hacker News

A critical perspective on agentic LLM coding tools and "vibe coding" from a senior developer who finds them slow and requiring excessive oversight. The author argues that vocal LLM advocates may be projecting their own insecurities about programming ability, attacking skeptics as resistant to change when experienced developers claim higher productivity without AI assistance. While acknowledging LLMs are useful for documentation lookup and enabling non-developers, the author questions whether prompt-driven development delivers on its promises for experienced programmers.

The Insecure Evangelism of LLM Maximalists

So i just read an article about a person having built a library with codex and chatgpt with 80% AI generated code and now this article.
Its insane how different people achieve very different results from the same technologies.
I think it highly depends on the models that you use and the thing you try to build.
On the other hand the 80% vibe coding guy also acknowledged that he had to do lots of handholding with the LLMs.

Why does the open graph image feels familiar 😂😂 ?

A great part of the illusion of LLMs is that in software engineering often 90% of the work is the last 10% of the implementation, and this is exactly the part where LLMs fail catastrophically.

Saving this to comment on it one year from now.

We are only figuring out where the LLMs and/or we fit in this wild west. And currently the LLM capabilities change so fast that it’s IMHO impossible to even make an educated guess what this all looks like in a year or even half. So paying attention, trying everything out but also playing safe is the way to go. I personally make LLMs write all my scripts, small isolated modules and start to migrate my system design decision into max modularity and isolation. Basically local micro service galore with skills and commands for Claude Code. If LLMs become God like - I will have a head start and if not… Its much simpler to merge into a monolith than split it up. But already it’s really nice to type in it “/perf” to make LLM scrawl db (preferably with read only account 😃 ) and other potential pain points like CPU/GPU/IO and other quotas and report if there anything worth my attention. This way I don’t depend on the output to be hight quality. It just says “Hey! This metric is sus!” And I either ask it to refine the command/skill to ignore it (by explaining why it is what it is) or use this info for deeper analysis (likely also with the LLM for a first draft) for me this is the definitely baseline future use of the tek unless they make it prohibitively expensive.