Hanako
- How to Build the Loops That Just Replaced Entire Prompt Engineering
- Andrej Karpathy built the most watched AI repo of the year in one month, wrote most of it by hand, and just explained why he rejected vibe coding he used autocomplete, not agents his reasoning: LLMs have too many cognitive deficits for code that has never been written before they kept trying to force him into PyTorch's DDP container, he had a custom gradient sync implementation, the models couldn't internalize that they kept bloating his code with try-catch statements, kept using deprecated APIs, kept trying to turn research code into a production codebase he said typing English instructions is slower than just navigating to the right line and letting autocomplete finish it 3 tiers of how people interact with code right now: > reject all LLMs and write from scratch, probably wrong > use autocomplete but stay the architect, his sweet spot > full vibe coding with agents, works for boilerplate, breaks on anything novel he vibe-coded two things: a boilerplate report generator and a Rust tokenizer rewrite where he had Python tests to verify against everything else was hand-written with autocomplete the part that matters for AI timelines: the main story about AI exploding to superintelligence depends on AI automating AI research, Karpathy says that's exactly what models are worst at they know things, they don't fully know how to integrate them into your repo, your style, your assumptions his current oracle: GPT-5 Pro, copy-paste the entire repo, ask questions, often surprisingly good compared to a year ago but his verdict: the industry is making too big of a jump, it's slop, they're not coming to terms with it Bookmark & Watch, then decide where you sit on the autonomy slider
How to Build the Loops That Just Replaced Entire Prompt Engineering
Andrej Karpathy built the most watched AI repo of the year in one month, wrote most of it by hand, and just explained why he rejected vibe coding he used autocomplete, not agents his reasoning: LLMs have too many cognitive deficits for code that has never been written before they kept trying to force him into PyTorch's DDP container, he had a custom gradient sync implementation, the models couldn't internalize that they kept bloating his code with try-catch statements, kept using deprecated APIs, kept trying to turn research code into a production codebase he said typing English instructions is slower than just navigating to the right line and letting autocomplete finish it 3 tiers of how people interact with code right now: > reject all LLMs and write from scratch, probably wrong > use autocomplete but stay the architect, his sweet spot > full vibe coding with agents, works for boilerplate, breaks on anything novel he vibe-coded two things: a boilerplate report generator and a Rust tokenizer rewrite where he had Python tests to verify against everything else was hand-written with autocomplete the part that matters for AI timelines: the main story about AI exploding to superintelligence depends on AI automating AI research, Karpathy says that's exactly what models are worst at they know things, they don't fully know how to integrate them into your repo, your style, your assumptions his current oracle: GPT-5 Pro, copy-paste the entire repo, ask questions, often surprisingly good compared to a year ago but his verdict: the industry is making too big of a jump, it's slop, they're not coming to terms with it Bookmark & Watch, then decide where you sit on the autonomy slider
