Hacker News Reader

Comment by akira2501 | Hacker News Reader

> I've been able to use GPT-4 extensively for coding, writing, comprehension, tone analysis, home repairs, automation, debugging vague errors, and a host of other things.

Is there any objectivity or measurements in your analysis?

> at the very least we're looking at quite striking incremental improvements

I'm not even sure what this should mean. This sounds like something I would find in a marketing pamphlet.

> "yo, why is the floor dirty? Get cleanin"

From.. possible striking improvements to the legal profession.. to.. a conversational mode embedded in my Roomba. You've actually perfectly captured the gulf between what this technology is marketed to be and what it is barely capable of doing.

View on HN

Consider this, a legal firm rather than using a search mechanism to scan 100,000 documents for specific terms - could instead say something like:

"You are an auditor/lawyer looking for suspicious conversations or patterns of deceit by employees of company Z. I will feed you documents and you will give your analysis of each document or collections of documents. Pay special attention to X Y Z. Consider these sorts of actions as deceptive (a,b,c,d). If you find anything of note, create a summary of the findings in the form ['wording indicates deceptive intent', document 37, page 93].".

You'd then feed it 20 pages at a time and for less than $100 you've scanned a mountain of legal documents in a way in which you'd normally require a team of legal assistants and 100s of hours of work.

Again, I'm not a lawyer but I can conjure this very simple example that I could implement in a month if I had the time. At this point I think you're just squeezing me for ideas.

Perhaps you should explain what the technology is marketed as that you see as lying or misrepresenting. That would make this more productive.

Key problem being, you cannot trust that what you're getting is accurate information, which somewhat negates having such a tool in the first place.

As a lawyer with 13 years experience in my previous life, I can tell you that you can't trust a paralegal either. What it does do is give you the "first pass" - if I could automate the first pass over 100,000 pages and the AI highlights say 100 examples of the content I am looking for, of which say 10 are good quality examples of what I am looking for, then we are off to a very good start.

I honestly haven't found hallucination to be a problem on GPT-4 when asking it to analyze or parse a dataset but can acknowledge it being possible (I just haven't encountered it).

I think that if we consider the accuracy rate as measured in various ways being roughly that of a human, then you're trading human mistakes for AI mistakes in exchange for dramatically lower costs and a dramatically higher speed of processing. You might even say a higher level of reasoning. In my own interactions it's been fantastic at reasoning clearly and quickly outside of complex trick questions. Most scenarios in life aren't generally full of trick questions.

This is why I've decided that I'm siding with the FTC on this product being a problem and that testing products on large swathes of the populace is a bad idea.

To some up my experience with basically all these products...YMMV.

How many tools are accurate? Every tool needs guidance and quality checks. If you turn the knob of your light switch and you stay in the dark, you need to find out whether it is the switch, the bulb or the net is broken.

I have the feeling that you are in denial. Each time something new is discovered, the goal post has moved. In previous century billions were spent on trying to automate transation to other natural languages, without good results. Today an LLM can do this mere as part of a side effect. Non-natural languages as a bonus.

> I have the feeling that you are in denial.

No, I simply disagree with your assessment of the technology and you are unwilling to admit you could be wrong. It's a very rude way to start a conversation with a stranger.

> Each time something new is discovered, the goal post has moved.

This isn't anything new, it's merely a large scale version of technology that has existed for more than 30 years powered by training methods too questionable to be seriously utilized previously. People misunderstanding any of this as "invention" is a large source of the problem, I think.

> In previous century billions were spent on trying to automate transation to other natural languages, without good results.

Why were those billions spent? What entities spent that money? Was it consumer driven?

> Today an LLM can do this mere as part of a side effect

Not particularly well, and the difference in cost between making GPT3 and GPT4 compared to the difference in their accuracy paints a bleak future for continued scaling of this methodology. It is, quite literally, all downhill from here.

> Non-natural languages as a bonus.

If you're willing to blindly trust that output. Which is why I refer to this as "snake oil." It really does appear to be the same phenomenon. People are all excited about what the label says it can do, they forgot to measure if it actually does that at a level that justifies it's cost.

I apologize for any rudeness on my part and i am ready to admit I could be wrong. Also my feeling about denialism could be wrong, it is just a description of a feeling.

Reply on news.ycombinator.com