Gemini-2.5 pro is a cowboy coder (but still faster than a Google engineer)youtube.com1 points·by sjmog·작년·2 comments
sjmog·작년·discussAI feels a lot to me like improv. It’s fun to watch, sometimes, and in experienced hands it can even be great - but it’s missing the crafted core of what makes something really good - the “thing” that great art becomes transparent before.
sjmog·작년·discussyesterday I posted a video testing claude-4 sonnet solving an https://simstack.io long-horizon swe challenge unaided (https://news.ycombinator.com/item?id=44424468). for comparison, here's gemini 2.5-pro.I noticed that 2.5-pro is way more cavalier, skipping backups, and trying "more stuff more quickly" than claude's more cautious approach.
sjmog·작년·discussforgot to mention that the fastest time to solve the first challenge was 43 minutes (cloudflare engineer)second was a google engineer, 48 minutesclaude + o3 (and a bit of help) does it in about 1h15