Hacker News Reader

My hypothesis is code completion is not a text completion problem. More of a graph completion one.

So we may have got to a local maximum regarding code helpers with LLMs and we'll have to wait for some breakthrough in the AI field before we get something better.

View on HN

But these models don't work that well even for text when you gave them a huge context. They're reasonably good at summarization, but if you ask them to "continue the story" they will write very inconsistent things (eerily similar to what a sloppy human writer does, though.)

We should be able to provide 2 fields, context and prompt so the prompt gets higher priority and don't get mixed with the whole context.

For this breakthrough to happen, big tech will need to hire software engineers again :)

But the good thing is that DeepSeek proved those breakthroughs are going to happen one way or another, fast.

Reply on news.ycombinator.com