undefined | Better HN

0 pointsnickpsecurity1y ago0 comments

The brain solves that problem. It seems to involve memory and specialized regions. I found a few groups building hippocampus-like, research models. One had content-addressable memory.

There was another one that claimed to get rid of hallucinations. They also said it takes 50-100 epochs for regular architectures to actually memorize something. Their paper is below in case people qualified to review it want to.

https://arxiv.org/abs/2406.17642

Like the brain, I believe the problem will be solved by a mix of specialized components working together. One of those components will be a memory (or series of them) that the others reference to keep processing grounded in reality.

0 comments

Animats1y ago

Comments on that paper? PDF: [1]

What they are measuring, it seems, is whether LLMs can be built which will retrieve a reliable known correct answer on request. That's an information retrieval problem, and, in fact, they solve it by adding "Memory Experts" which are basically data storage.

It's not clear that this helps either replies which require synthesizing disparate information, or detecting that the training data does not contain info needed to construct a reply.

[1] https://arxiv.org/pdf/2406.17642

nickpsecurityOP1y ago

On the second paragraph, there’s been work that shows whether a model has memorized or is strongly replying to certain prompts. Something like that combined with a memory-equipped model would tell you if it might contain the info.

From there, you need multiple layers building on info it contains to synthesize a reply that might be good. Alternatively, an iterative process going a few rounds through a model, re-presenting the combo of results together, and it fuses them. All based on known data or what’s in the prompt with nothing else.

This is speculative based on a few things our own minds do.

j / k navigate · click thread line to collapse