[bearjaws]

I just used it to compare two smaller legal documents and it completely hallucinated that items were present in one and not the other. It did this on three discrete sections of the agreements.

Using ctrl-f I was able to see that they were identical in one another.

Obviously this is a single sample but saying 90% seems unlikely. They were around ~80k tokens total.

[carlosbaraza]

I have the same feeling. I asked to find duplicates in a list of 6k items and it basically hallucinated the entire answer multiple times. Some times it finds some, but it interlaces the duplicates with other hallucinated items. I wasn't expecting it to get it right, cause I think this task is challenging with a fixed amount of attention heads. However, the answer seems much worse than Claude Opus or GPT-4.

[akomtu]

Everyone is trying to use Language Models as Reasoning Models because the latter haven't been invented yet.

[fnordpiglet]

That’s not needle in a haystack.

I would note that LLMs handle this task better if you slice the two documents into smaller sections and iterate section by section. They aren’t able to reason and have no memory so can’t structurally analyze two blobs of text beyond relatively small pieces. But incrementally walking through in much smaller pieces that are themselves semantically contained and related works very well.

The assumption that they are magic machines is a flawed one. They have limits and capabilities and like any tool you need to understand what works and doesn’t work and it helps to understand why. I’m not sure why the bar for what is still a generally new advance for 99.9% of developers is effectively infinitely high while every other technology before LLMs seemed to have a pretty reasonable “ok let’s figure out how to use this properly.” Maybe because they talk to us in a way that appears like it could have capabilities it doesn’t? Maybe it’s close enough sounding to a human that we fault it for not being one? The hype is both overstated and understated simultaneously but there have been similar hype cycles in my life (even things like XML were going to end world hunger at one point).

[HarHarVeryFunny]

That's a different test than needle-in-a needlestack, although telling in how brittle these models are - competent in one area, and crushingly bad in others.

Needle-in-a-needlestack contrasts with needle-in-a-haystack by being about finding a piece of data among similar ones (e.g. one specific limeric among thousands of others), rather than among disimilar ones.

[1970-01-01]

I've done the same experiment with local laws and caught GPT hallucinating fines and fees! The problem is real.

[tmaly]

Imagine if they started using LLMs to suggest prison sentences

[Aerbil313]

Interesting, because the (at least the official) context window of GPT-4o is 128k.

[davedx]

> Obviously this is a single sample but saying 90% seems unlikely.

This is such an anti-intellectual comment to make, can't you see that?

You mention "sample" so you understand what statistics is, then in the same sentence claim 90% seems unlikely with a sample size of 1.

The article has done substantial research

[dkjaudyeqooe]

That fact that it has some statistically significant performance is irrelevant and difficult to evaluate for most people.

He's a much simpler and correct description that almost everyone can understand: it fucks up constantly.

Getting something wrong even once can make it useless for most people. No amount of pedantry will change this reality.

[davedx]

What on earth? The experimental research demonstrates that it doesn't "fuck up constantly", you're just making things up. The various performance metrics people around the world to measure and compare model performance is not irrelevant because you, some random internet commenter, claim so without any evidence.

This isn't pedantry, it's science.

[lopuhin]

And also article is testing on a different task (Needle in a Needlestack which is kind of similar to Needle in a Haystack), compared to finding a difference between two documents. For sure it's useful to know that the model does ok in one and really bad in the other, does not mean that original test is flawed.

[bckr]

Yeah I asked for an estimate of the percentage of the US population that lives in the DMV area (DC, Maryland, Virginia) and it was off by 50% of the actual answer, which I only realized when I realized I shouldn’t trust its estimate for anything important

[KeplerBoy]

Those models still can't reliably do arithmetic, so how could it possibly know that number unless it's a commonly repeated fact?

Also: would you expect random people to fare any better?

[bckr]

It used web search (RAG over the entire web) and analysis (math tool) and still came up with the wrong answer.

It has done more complex things for me than this and, sometimes, gotten it right.

Yes, it’s supposed to be able to do this.

[chrischen]

Arithmetic just happens to be something we can easily and reliably verify, so it becomes painfully obvious when LLMs are just stringing together some words that sound like the right answer.

[kylebenzle]

What you are asking an llm to do here makes no sense.

[potatoman22]

Why not? It seems like a natural language understanding task

[marshray]

You haven't seen the promotion of the use of LM AI for handling legal documents?

It's purported to be a major use case.

[cmrdporcupine]

You might be right but I've lost count of the number of startups I've heard of trying to do this for legal documents.