Cory Bev
cbevil.bsky.social
Cory Bev
@cbevil.bsky.social
Software engineer, film lover, argument haver
Gotta have a soundtrack. Working out, driving, etc. I used to listen to music on my headphones while taking college exams.
December 22, 2025 at 4:46 AM
Woof
December 18, 2025 at 8:14 PM
Reposted by Cory Bev
100%. For example, what if voters aren’t asking for a line graph showing retail theft going down, but to not see any context-free, undated mass shoplifting videos in their feeds? Good luck with that if our palantíri algo overlords don’t wish it.
December 18, 2025 at 3:37 AM
You’re just getting a probabilistic traversal of some word graph that encodes an unknowable mix of right, wrong, and ambiguous text from billions of conflicting sources
December 16, 2025 at 8:05 AM
A large language model isn’t doing that. There is no objective function for human speech on well-understood topics, nevermind topics on the frontier of our knowledge.
December 16, 2025 at 8:05 AM
They’ve conflated all uses of machine learning with “ai”. Yes machine learning is Incredibly useful when there is a defined objective function and you can iterate endlessly and conform to some (local) maximum.
December 16, 2025 at 8:02 AM
I see more of a spark of inner life in the pig’s eye
December 16, 2025 at 7:56 AM
If you insist
December 16, 2025 at 7:51 AM
Reposted by Cory Bev
This is one of the reasons I hate what big tech has done to AI over the last couple of years. There are great use cases for AI / machine learning. But they chose to shove incompetent models into every available receptacle while screaming "THIS WILL DESTROY YOUR LIFE, WE OWN YOU LOL"
December 10, 2025 at 12:13 AM
So they’ll delete all neural network layers trained on unlicensed data and start new models with only licensed training data? You have a lot more faith in big tech than I do.
December 10, 2025 at 11:30 PM
How does one license the entirety of the internet?
December 10, 2025 at 8:30 PM
It absolutely is. Without the (stolen) input data, these models would not be able to generate anything at all.
December 10, 2025 at 8:29 PM
So unless a model is trained solely on synthetic or licensed data sets (arguably impossible at this scale), it is a plagiarism machine.
December 9, 2025 at 11:49 PM