🪨 (@Varpie)

Ars Technica @arstechnica@mastodon.social

2mo

Beyond grep: The case for a context-rich AI coding harness

Augment Code's Vinay Perneti talks models, harnesses, and context.
arstechnica.com/ai/2026/07/beyond-grep-the-case-for-a-context-rich-ai-coding-harness/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_medium=social

1

0

techno156 @techno156@mastodon.social

2mo

@arstechnica Surely the use of an LLM would be far inferior to grep? One of the strong points of grep is that it runs really quickly, which is basically the inverse of a model.

It seems like using it that way would just have it blow up the context window, and reduce its usefulness.

1

0

🪨 @Varpie@peculiar.florist

2mo

@techno156 @arstechnica But one of the weak point of grep is that it only looks for what you ask for. LLMs, and language embeddings in general, can encode semantic meaning to find close results, so for instance when you look for vehicles it can return results related to cars, trucks, bikes... which can't be done directly with grep. Which increases its usefulness.

Word embeddings have been used successfully for search years before LLMs were a thing, and grep continues to be used, they just represent different needs. Sometimes, you know exactly what you want and grep is great, but sometimes you want something more broad and LLMs can provide better results, despite taking more time.

⭐1

0

1