Comment by jll29 | Hacker News Reader

Comment by jll29 | Hacker News Reader

Parent Full thread

Read Paul Tetlock's research about so-called "experts" and their inability to make good forecasts.

Here's my own take:

- It is far too early to tell.

- The roll-out of ChatGPT caused a mind-set revolution. People now "get" what is possible already now, and it encourages conceiving and persuing new use cases on what people have seen.

- I would not recommend any kinds to train to become a translator for sure; even before LLMs, people were paid penny amounts per word or line translated, and rates plummeted further due to tools that cache translations in previous versions of documents (SDL TRADOS etc.). The same decline not to be expected for interpreters.

- Graphic designers that live from logo designs and similar works may suffer fewer requests.

- Text editors (people that edit/proofread prose, not computer programs) will be replaced by LLMs.

- LLMs are a basic technology that now will be embedded into various products, from email clients over word processors to workflow tools and chat clients. This will take 2-3 years, and it may reduce the number of people needed in an office with a secretarial/admin/"analyst" type background after that.

- Industry is already working on the next-gen version of smarter tools for medics and lawyers. This is more of a 3-5 year development, but then again some early adopters started already 2-3 years ago. Once this is rolled out, there will be less demand for assitants-type jobs such as paralegals.

My dentist already uses something called OverJet(?) that reads X-rays for issues. They seem to trust it and it agreed with what they suspected on the X-rays. Personally, I’ve been misdiagnosed through X-rays by a medical doctor so even being an LLM skeptic, Im slightly favorable to AI in medicine.

But I already trust my dentist. A new dentist deferring to AI is scary, and obviously will happen.

I had a misread X-ray once, and I can see how a machine could be better at spotting patterns than a tired technician, so I'm favorable too. I think I'd like a human to at least take a glance at it, though.

The mistake on mine was caught when a radiologist checked over the work of the weekend X-ray technician who missed a hairline crack. A second look is always good, and having one look be machine and the other human might be the best combo.

Unfortunately what's likely to happen in our current society and medical system is:

1. Instead of being expected to read N x-rays per hour/day/whatever, radiologists will soon be able expected review 3N or 4N AI diagnoses in the same time.

2. With far less time to spend on each case and an AI that is right most of the time (but imperfect) humans won't be great at catching AI mistakes, even if they're trying.

3. Pay and prestige for radiologists will drop, leading to a lack of quality people entering the field, exacerbating the problems in #2.

4. Eventually administrators and/or politicians will decide that (using made up numbers) very cheap AI-only review with 90% accuracy is "good enough", even though the more expensive AI+human combo can yield 93% accuracy or very expensive human-only review can yield 97% accuracy.

> A second look is always good, and having one look be machine and the other human might be the best combo

For now I agree. 2-4 years from now it can be 20 ultra strong models each trained somewhat differently that converse on the X-ray and reach a conclusion. I don't think technicians will have much to add to the accuracy.

I regularly get to hear veterinarian rants. AI is being forced on them by the corporate owners in multiple fronts. The pattern goes:

Why aren’t you using the AI x-ray? Because it too often misdiagnoses things and I have to spend even more time double checking. And I still have to get a radiologist consult.

Why are you frustrated that we swapped out the blood testing thingamabob with an AI machine? Because it takes 10 minutes to do what took me 30 seconds with a microscope and is STILL not doing the full job, despite bringing this up multiple times.

Why aren’t you relying more on the AI text to speech for medical notes? Because the AVMA said that a doctor has to review all notes. I do, and it makes shit up in literally every instance. So I write my own and label the transcription as AI instead of having to spend even more time correcting it.

The best part is that the majority of vets (at least in this city) didn’t do medical notes for pets. Best you’d often get when asking is a list of costs slapped together in the 48 hours they had to respond. Now, they just use the AI notes without correcting them. We’ve gone from zero notes, so at least the next doctor knows to redo everything they need, to medical notes with very frequent significant technical flaws but potentially zero indication that it’s different from a competent doctor’s notes.

This is the wrong direction, and it’s not just new doctors. It’s doctors who are short on time doing what they can with tools that promised what isn’t being delivered. Or doctors being strong armed into using tools by the PE owners who paid for something without checking to see if it’s a good idea. I honestly do believe that AI will get there, but this is a horrible way to do it. It causes harm.

> Text editors (people that edit/proofread prose, not computer programs) will be replaced by LLMs.

This is such a broad category that I think it's inaccurate to say that all editors will be automated, regardless of your outlook on LLMs in general. Editing and proofreading are pretty distinct roles; the latter is already easily automated, but the former can take on a number of roles more akin to a second writer who steers the first writer in the correct direction. Developmental editors take an active role in helping creatives flesh out a work of fiction, technical editors perform fact-checking and do rewrites for clarity, etc.

> Read Paul Tetlock's research about so-called "experts" and their inability to make good forecasts

Do you mean Philip Tetlock? He wrote Superforecasting, which might be what you're referring to?

We're ten years and a trillion dollars into this. When we were 10 years and $1T into the massive internet builout between 1998 and 2008, that physical network had added over a trillion dollars to the economy and then about a trillion more every year after. How's the nearly ten years of LLM AI stacking up? DO we expect it'll add a trillion dollars a year to the economy in a couple years? I don't. Not even close. it'll still be a net drain industry, deeply in the red. That trillion dollars could have done some much good if spent on something more serious than man-child dreams about creating god computers.

> - Text editors (people that edit/proofread prose, not computer programs) will be replaced by LLMs.

It has been a very, very long time since editors have been proof-reading prose for typos and grammar mistakes, and you don't need LLMs for that. Good editors do a lot more creative work than that, and LLMs are terrible at it.

Name a better duo: software engineering hype cycles and anti-intellectualism

We were the stochastic parrots all along.

And because we are very smart, so must be it.

Isn't your post anti-intellectual, since you're denigrating someone without justification just for referencing the work of a professor you disagree with?

Video VFX artists are already suffering from lower demand.

Reply on news.ycombinator.com