- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
Less than a month after New York Attorney General Letitia James said she would be willing to seize former Republican President Donald Trump’s assets if he is unable to pay the $464 million required by last month’s judgment in his civil fraud case, Trump’s lawyers disclosed in court filings Monday that he had failed to secure a bond for the amount.
In the nearly 5,000-page filing, lawyers for Trump said it has proven a “practical impossibility” for Trump to secure a bond from any financial institutions in the state, as “about 30 surety companies” have refused to accept assets including real estate as collateral and have demanded cash and other liquid assets instead.
To get the institutions to agree to cover that $464 million judgment if Trump loses his appeal and fails to pay the state, he would have to pledge more than $550 million as collateral—“a sum he simply does not have,” reportedThe New York Times, despite his frequent boasting of his wealth and business prowess.
You cannot rely on a LLM to summarise accuratly
Plus, that’s not a good task for an llm because its context window would almost certainly be too short.
It would “hallucinate” because it could only “remember” a fraction of the content and then everyone would be all pissy because they used the program wrong.
I mean you can pretty simply just engineer around that. Dumping 5k pages is obviously an idiotic way of approaching the issue. But having an LLM going through 500 words at a time, with 125 words of overlap in each sequence to pull out key words, phrases, and intentions, then put that into a structured data form like a JSON. Then parse the JSONs to pick up on regions where specific sets of phrases and words occur. Give those sections in part or entirely to the LLM again; again have it give you structured output. Further parse and repeat. Do all of these actions several times to get a probability distribution of each assumption around what is being said or is intended. Build the results into a Bayes net, or however you like, to get at the most likely summaries of what the document is saying. These results can then be manually reviewed. If you are touchy, you can even adjust the sensitivity to pick up on much more nuanced reads of the text.
Like, if the limit of your imagination is throwing spaghetti against a wall, obviously your results are going to turn out like shit. But with a bit of hand holding, some structure and engineering, LLM’s can be made to substantially outperform their (average) human counter parts. They do already. Use them in a more probabilistic way to create distributions around the assumptions they make, and you can set up a system which will vastly outperform what an individual human can do.
(just asked up the thread:)
GPT-4 & Claude 3 Opus have made little summarization oopsies for me this past week. You’d trust ‘em in such a high profile case?
if you end them 100 times over the same text?
This is court, not a school project or academia. But in general I agree with you.
No this is discovery and we’re discussing how you would engineer a system to support automating it.
LLMs are still pretty limited, but I would agree with you that if there was a single task at which they can excel, it’s translating and summarizing. They also have much bigger contexts than 500 words. I think ChatGPT has a 32k token context which is certainly enough to summarize entire chapters at a time.
You’d definitely need to review the result by hand, but AI could suggest certain key things to look for.
People were doing this somewhat effectively with garbage Markov chains and it was ‘ok’. There is research going on right now doing precisely what I described. I know because I wrote a demo for the researcher whose team wanted to do this, and we’re not even using fine tuned LLMs. You can overcome much of the issues around ‘hallucinations’ by just repeating the same thing several times to get to a probability. There are teams funded in the hundreds of millions to build the engineering around these things. Wrap calls in enough engineering and get the bumper rails into place and the current generation of LLM’s are completely capable of what I described.
This current generation of AI revolution is just getting started. We’re in the ‘deep blue’ phase where people are shocked that an AI can even do the thing as good or better than humans. We’ll be at alpha-go in a few years, and we simply won’t recognize the world we live in. In a decade, it will be the AI as the authority and people will be questioning allowing humans to do certain things.
Read a little further. I might disagree with you about the overall capability/potential of AI, but I agree this is a great task to highlight its strengths.
Sure. and yes I think we largely agree, but on the differences, I seen that they can effectively be overcome by making the same call repeatedly and looking at the distribution of results. Its probably not as good as just having a better underlying model, but even then the same approach might be necessary.
Exactly. There is already one recent case where a lawyer filed a brief generated by an LLM. The judge is the one that discovered the cited cases were works of fiction created by the LLM and had no actual basis in law. To say that the lawyer looked foolish is putting it lightly…
Right, but that’s not what we’re talking about here - we’re not saying “Hey LLM, write a convincing sounding legal argument for X”, we’re saying “Hey LLM, here’s a massive block of text, summarize what you can and give me references to places in the text that answer my questions so I can look at the actual text as part of building my own convincing sounding legal argument for X.”
It’s the difference between doing a report on a topic by just quoting the Wikipedia article, versus using the Wikipedia article to get a list of useful sources on the topic.
But you can use it as a tool to assist. If it finds something actionable, you can confirm the old-fashioned way, by doing the actual reading.
until it just makes stuff up, as they have done
There are specialized LLMs that (if the document is digitized) will actually cite their references within the data they’ve been given and provide direct links. It’d still need proof reading as someone would have to check those citations but it would still speed up the process immensley.
I’ve used AI. I’ve had it make stuff up or put incorrect info into documents. I’m smart* and read through the document just like these lawyers will**. Saved me TONS of time vs just doing all the specific writing, scanning and summarizing. *citation required **smartness not withstanding
You can’t rely on people to summarize it accurately either. Humans make mistakes too. The difference is that I can ask an LLM to do the summarizing 10x, and calculate a statistical probability of a given statement being present or true in the text, at a very low cost. Just because LLM’s aren’t 100% reliable doesn’t make humans the best bar to rely upon either.