this post was submitted on 14 May 2026
505 points (96.5% liked)
me_irl
7664 readers
2566 users here now
All posts need to have the same title: me_irl it is allowed to use an emoji instead of the underscore _
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
You have to revisit this concept every 6 months or so. Shit's moving fast, they're getting better at references and checking actual reference material before giving you an answer.
I'm not saying that everything is solidly as good as people right now, but you can't just use that blanket statement forever.
Asking one of the new models to make a paper on a 5th grade subject written like a 5th grader who get's B's did it, then clearing context or changing engines and asking it to check your work throug the lense of a 5th grade teacher would very likely net you nearly undetectable results.
You're so right I bet there's a boatload of training data on the open web about what 5th grade teachers are looking for when grading papers
Unironically, there likely is. You don’t think there are discussions on Facebook, Reddit, WhatsApp and other places where teachers discuss their ways they are finding students who are using LLMs? Cmon now, let’s not bury our heads in the sand.
We probably even have data in training about AI detection and base system prompts to take that into account when asking for educational help.
If they manage to get this all absolutely correct, we're just pumping the gas all the way to Idiocracy.
"Just 10 more data centers bro. Just another billion dollars and it will be as good as a person. Just 6 more months bro."
Yes, they are improving, but there is also deminishing returns. If you doubled the amount of resources in genAI right now you would not get something twice as good, you might get a 25% improvement, and the resources currently being used is already obscenely high. GenAI is slowly closing the gap, but the resources required to do so are rapidly increasing. Actually closing the gap in unsustainable.
Did you literally just say "if you have genAI check the work of other genAI the genAI will say it's good"?
25% wishful thinking. Valid point tho.
Yes, that is how they're getting by a large number of the previous issues. Multiple tries across versions of models with different training. Add in web searches. They're getting accuracy by cheating precision. It's expensive as fuck too.
absolutely correct. One query to local llm has a decent chance to be wrong. To bump that up, they're generating a shit ton of queries. It's eventually good for humanity overall, by the time they get it truly reasonable, the cost of the queries will be so high that when venture cap runs out, no one will be able to afford it even if it is replacing wages. Then we can go back to just using it as a tool.