>>16690233
>Even traditional software behaves in unexpected ways unintended by its developers
NTA, but this always, always boils down to the developers not realizing exactly what they were telling the software to do - software bugs aren't magic, a traditional program is doing exactly what it's told to do barring hardware-level faults.

>>16690181
>"in context" learning
You're probably aware, but this isn't learning - it's expanding the input scope with previous output. As I said in >>16690180, we haven't even come close to figuring out how to do active learning. More specifically, we have no real solutions to the vanishing gradient problem - if we did, they would have been jumped on for active learning.

We also don't know how to train feed-forward networks to do anything except approximate a probability function. You can approximate any function within the training window in this way, but the approximation can't actually be used to extrapolate outside of the training window - to my knowledge, it's actually much worse at extrapolation than bog-standard regression models are.

...We also only know how to train feed-forward networks. Recurrent networks are trained by training a feed-forward network and then "folding" it back into a recurrent network.

Research into any of these problems feels like a more promising avenue than "just make the chatbot bigger," because despite some cope about it we really HAVE run out of quality training data and models have effectively plateaued. The current grift is cherry-picked and carefully crafted "metrics" to make newer models look like bigger improvements or more specialized models to look more generalist than they actually are. We haven't actually moved far from what was the cutting edge in late 2023. There was obsessive industry mass-adoption in 2024, but it is losing a bit of steam because end users are starting to realize the models are hitting diminishing returns before hitting "worker replacement" levels of competence.

Back | Go to board

Board:

/sci/

Thread:

what are humans going to do once robots learn to do all the jobs?