Raw LLM Responses

I can confirm the video is accurate. AI is a tool, not a replacement, and in the hands of a strong developer, it’s a real speed and capability multiplier. The biggest issue is the codebase and context problem mentioned in the video. The model may claim there’s a bug or missing code, but when you check, it’s actually there. If you push it to re-check, it will often correct itself and acknowledge the mistake. That’s a simple example, but it highlights the broader point. Without a reliable, grounded context, AI can sound confident while being wrong. What the video nails especially well is the multi-file problem. As soon as a change touches multiple files — interfaces, shared types, config, tests, build scripts, or cross-module assumptions — the failure rate spikes. It might update one file but miss dependencies elsewhere, introduce inconsistencies, or “fix” symptoms while breaking the system in a different place. A Next.js app doesn’t build when AI touches it (I know you know what I mean 😄). This is why you still need a human developer in the loop. Someone has to define the intent precisely, guide the approach, and verify the result. Even then, context-window limits mean the model can truncate, omit, misread, or apply changes offset from the intended location. Sometimes it effectively only “sees” a slice of a file, like the first or last chunk, and that’s how subtle breakage slips in. This happens because it can’t reliably keep an entire codebase in context, even with very large context windows in the million-token range, so it ends up working from partial slices of files and missing important details. Now scale that up to a real web app with 200+ files and 50K+ lines of code. Without strong tooling, scoping, and validation around it, it will struggle, and you shouldn’t trust it blindly. Developers aren’t gone or replaced. What’s happening is that mid-level devs can upgrade themselves to senior faster with AI, and seniors don’t need as many interns or juniors anymore because, with AI, a good senior can operate like “10 interns/juniors” by themselves. That’s a big part of why you see fewer junior roles. Anyone who’s used tools like VS, Cursor, or Antigravity together with Claude or Gemini will recognize exactly what I mean. If Gemini touches your mid-to-large-sized project, you can be sure it won't work/build... Regards.