Exactly! It must be exhausting to have this huge preoccupation with determining if something has come from an LLM or not. Just judge the content on it's own merits! Just because an LLM was involved doesn't mean the underlying ideas are devoid of value. Conversely, the fact that an LLM wasn't involved doesn't mean the content is worth your time of day. It's annoying to read AI slop, but if you're spending more effort suspiciously squinting at it for LLM sign versus assessing the content itself, then you're doing yourself a disservice IMO.
You could make the case uv falls in this category (I just prefix all my pip commands with uv) though we have yet to see if astral will become a "successful business"; I'm hoping they pull it off.
People keep citing this study (and it was on the top of HN for a day). But this claim falls flat when you find out that the test subjects had effectively no experience with LLM equipped editors and the 1-2 people in the study that actually did have experience with these tools showed a marked increase in productivity.
Like yeah, if you’ve only ever used an axe you probably don’t know the first thing about how to use a chainsaw, but if you know how to use a chainsaw you’re wiping the floor with the axe wielders. Wholeheartedly agree with the rest of your comment; even if you’re slow you lap everyone sitting on the couch.
I should have said “traditional ads” but yes you’re right. Product placement is one of many use cases. You could also sponsor bounties with more general requirements in order to drive engagement in a particular sector, or incentivize comments with a particular sentiment (e.g., I have a bounty for encouraging comments on hobbyist subreddits).
One very simple use case is making repetitive edits to a YAML (or similar) file. Sure, I can record a vim macro and/or try and conjure up some way to get it done with as few keystrokes as possible and hope I don’t fat finger anything along the way. Or I can just pipe it to llm and say [make this edit], and it just works.
It's the wrong acronym. I wrote this blog post on the bike and used an LLM to fix up the dictation that I did. While I did edit it heavily and rewrote a lot of things, I did not end up noticing that my LLM expanded MCP incorrectly. It's Model Context Protocol.
Which I don't feel great about because I do not like to use LLMs for writing blog posts. I just really wanted to explore if I can write a blog post on my bike commute :)
I’ve implemented Temporal activities that pull content for each supported platform/content type and then feed that into an LLM with the bounty requirements. It’s not very elegant but for an MVP it seems fine.
Now that tool calls and structured outputs are more baked into LLM APIs, I have some refactors planned, to make the Workflow more “agentic”, but I don’t expect the user experience to change all that much; the biggest improvement will be being able to dive deeper into the content, assess images, etc.
reply