Hacker News Reader

Comment by thcipriani | Hacker News Reader

I had the same thought. People know that ChatGPT is completely fallible—prone to hallucination, etc—yet people continue to abuse it to do things it's incapable of doing—write a legal brief, create an A paper, etc. This is the same magical thinking that would lead someone to consult a wizard.

View on HN

No, that's just you dissing a tool by playing up its imperfections. Some people do what you say. Others, and I imagine most ChatGPT users that also hang out on HN are like that, recognize the tool is fallible, but also that for many tasks, it's orders of magnitude easier to check a solution than to come up with one from scratch. Even >50% rate of failure is fine - you can retry your query four times, mash the responses into something serviceable, and sand off the rough edges yourself; it's still easier and faster than doing it yourself from scratch.

Or put another way: you bet I'd go to my local wizard if magic was real and they were half-competent at wielding it. Now, magic isn't real, but GPT-4 being better at almost everything than any single human, is a trivially verifiable fact.

Or put yet another way: most software devs are fallible and half-competent at what they do; this doesn't stop the world from making software.

> GPT-4 being better at almost everything than any single human, is a trivially verifiable fact.

You are either very wilfully ignorant or deluded beyond redemption. Name one thing GPT-4 does better than a competent human — beyond sheer volume of output.

Translating any language.

That is indeed probably the strongest application (indeed, the original motivation for the development of) LLMs. But it's not one thing. It may be able to 'translate any language', but it fails to be as good as the best translator of any given language. GPT-4 does indeed have a range exceeding that of any human ever, but if you fix a domain then it won't come out on top.

So try again.

I'm no wizard of AI or anything, but Claude 3 has been exceptionally good for me at malicious compliance and bureaucratic red tape wars allowing me to bring an additional (and welcome) perspective into complex circumstances. I've also used Claude for some math and physics problems and while it can get off the rails if unconstrained or if the problem is poorly defined, it's surprisingly good and its logic matches other humans at the undergrad level. But, I'll also be the very first person to call this technology a stochastic parrot. I'm not opposed to using a screwdriver for a hammer, if it works.

Look, that's my point: for any given specific task, there are some humans better at it than GPT-4. But there are no humans that are this good at more than a handful things simultaneously. GPT-4 is. It's not word-class in anything, but it's above average in almost everything. That's a very useful quality to have.

Is it, though? The world needs quality work, not vast amounts of mediocrity.

The world runs on mediocrity. To demand otherwise is a mistake of perfectionism.

Think of the grocery store you frequent, or the barber you visit, or the accountant that does your taxes or keeps company books. They may be the best available to you, but by world standard, they're all mediocre. Hopefully, you don't hold it against them.

> GPT-4 being better at almost everything than any single human

Why do people insist on comparing GPT's output against the output of "any single human"?

Of course if you pick any human on the planet, GPT is better than them at generating something. It's better at generating poetry than I am for instance.

But it's generally not better at humans when doing the things that those humans are good at. It's generally not better than software devs at writing code, it's generally not better than artists at creating artwork, it's generally not better than doctors at making diagnosis, or better than lawyers at producing legal arguments

Comparing GPT to a single human is absurd. Compare it to the people who you're actually trying to replace with it and it comes up short constantly

At least for now

Because none of us have enough friends, colleagues, or professional contacts to have access to someone good in any random thing you may need for one-off custom job. Most of the time you have, you'd have to pay through the roof for it. GPT-4 is approximately free, so even if it's not particularly good at most things, it's still more help than any of us could have otherwise.

Reply on news.ycombinator.com