[programmertote]

Same. Last week, my boss, Chief Data and Analytics Officer, dumped an AI-generated proposal (~7 pages) on how to structure semantic layer on top of our dbt models. As the Data Engineering lead, I had to read it and found a few glaring issues; left a lot of comments asking him for details where it's lacking; and proposed a few of my ideas (the path I think we should take without over complicating everything unnecessarily--esp. in the beginning).

Yesterday, one of my coworkers (Senior Dir. of Research Ops) shared with me another obviously AI-generated 5 page draft of an SOP on how to reintegrate old metrics (in the legacy SQL Server environment) into the Azure SQL while keeping everything running smoothly. She's not the most technical person, so it obviously is reflected in the doc generated.

I think we will all become AI-output-reviewers eventually. Not sure how long I can keep doing this though because the volume of materials that need reviewing seems to be growing really quickly these days....

[postsantum]

Reminds me of that guy with "people skills" from Office Space movie

"Why couldn't customers take the requirements to the software people?"

https://www.youtube.com/watch?v=RkmuI5W694o

[jadar]

Starting to wonder that we’re going to start being forced to execute on an AI output instead of sharing it with other people. If you can reason yourself into a working system, you know what you’re talking about. If not, then it’s not worth taking the time to figure it out.

[turtleyacht]

It's still risk though. Consensus distributes it. Future maintainers need to understand the design and motivations.

[jadar]

That’s not much better in my experience, from a human perspective. You inherit a system that someone designed a decade ago and all the original maintainers are long gone. One is then afraid to change anything, because of ancient landmarks and all that. But now we can actually start to piece together how these things work with AI.

[turtleyacht]

Yes, for legacy--production systems--it's totally useful to generate a high-level overview, surround an unfamiliar feature with tests, and build small harnesses that exercise real functionality.

The sooner one can find an entrypoint, the sooner one can contribute with changes.

[arewethereyeta]

set your own AI-reviewer of AI content and ping pong your way to pension.

[jochem9]

Ask them to share their prompt instead.

Calls them out on their AI bs and gives a way forward to share what they actually thought.

[sixtyj]

People will either get used to it, or we’ll start seeing automated replies like “This text was automatically generated and has not been reviewed before being sent. It’s AI garbage and I refuse to read it. Do it again. And do it better.”

[LPisGood]

Why not ask your AI to review their AI slop?

It feels like a new age version of “that which can be asserted without evidence can be dismissed without evidence.” Basically, “that which can be sent without thought can be responded to without thought.”

[walrus01]

> Why not ask your AI to review their AI slop?

If you want to be insulting about it, use a locally hosted "small" AI (under 6GB size on disk GGUF file, Q4 quantized or worse, so quite "stupid"), set to a high temperature value, no thinking mode, with a system prompt instructing it to fire off a rapid response in an absurdist writing style.

[jadar]

That’s been my experience. If I send something generated with AI to a colleague, then I get something generated with AI back. Fair enough.

[tacostakohashi]

I also use this strategy of mutually assured AI slop, with a no-first-use policy. I'll happily respond to AI slop PRs with AI slop code reviews / comments.

[isghoor]

I have found using an llm to compress the information instead instead of reading it directly saves time. It's potentially lossy but it's not like the slop you got is going to lose much value by compressing it.