You can extract utility, incidentally, from its correlative pattern matching by treating it kind of like a direct translation tool from natural language to programming language. This is not how it is used in production environments now, though.

Show Post

For LLM-assisted coding, I feel that the only way to use it carefully is to use the LLM 'raw' for specifically prompted code. The agentic harnesses are built to obscure incidentality and make you believe the LLM actually understands the project.

Show Post

People think that they have prompt engineered a workflow, and they begin to trust it; they say: LLMs are unreliable in the whole, but this particular workflow works. It even feels like they have put care into engineering it. But every time it appeared to work was incidental.

Show Post

i mean you can crack an egg with a sledgehammer if you're careful with it, but that's not really what they're made for. and I think there's an argument that the slot machine nature of LLM interaction undermines the iterative process of theory formation

Show Post

@vortexegg.com has described a spectrum of "slop vs care" I find useful because it captures how most ai/LLM related problems aren't really new, but existing harms amplified in that context we can think of ai as the automation of careless production

Show Post