Hacker News Reader

Comment by sandworm101 | Hacker News Reader

>> unattended LLM do anything for your company at all. It can, and will, fuck up dramatically sooner or later.

So, just like any other random employee?

To err is human. To fuck up a million times per second, you need a computer.

Granted, here at the beginning of 2024, an LLM can not quite attain that fuck up velocity. But take heart! Many of the smartest people on Earth are working on solving that exact problem even as you read this.

"Fuck up velocity" goes straight into my vocabulary.

FPS? Fuckups per second?

Maybe FuPS? So it's easier to tell from other two FPS-es.

Fups FTW!

That’s it. FUPS is a frequency measurement of the rate at which an AI produces “fuckups” per second, where a “fuckup” is defined as an individual non-unique production error that has a measurable negative impact on society.

I imagine this is how the jerk unit (rate at which acceleration changes) got made up

coined

Or, as the saying goes, "Computers make very fast, very accurate mistakes." - [1]

[1] https://quoteinvestigator.com/2022/01/02/computer-mistakes/

I mean is this not part of the AI safety thing some people warn about... not that AI will attain human fuck up velocity, but far exceed it?

No. Random employees have a well-understood distribution of mostly normal human errors of certain types and estimated severity, relative to unattended LLM which has a poorly-understood distribution of errors in both type and severity. (“SolidGoldMagikarp”.)

copy&paste errors are exactly what human employees are good at. this could very easily be the result of a bad copy&paste by a human into a form. especially if the copy&paste text is in a language not understood by the human employee. to them, it might look just like one of the other hundreds of search term word salad used as titles

Whether it’s human or not is irrelevant to the point: human beings fail much more predictably.

When the same search term salad is presented hundreds of times for copy paste, a human would notice and have an opportunity to ask a supervisor.

A chatbot automation would not notice the repetition unless it had been coded to detect repetition, and/or to reject the ChatGPT refusal message.

Ironically, it was probably an automation coded by ChatGPT.

Why is it that LLMs are so often compared to employees and their responsibilities? In my opinion, it is an employee that actively USES the LLM as a tool and this employee (or his/her employer) is responsible for the results.

It's a dumb/lazy/specious talking point. You can kill someone with a pencil just like you can kill someone with a gun, but the gun scales up the danger so we treat it and regulate it differently. You can kill someone with a bike, a car, or an airplane, but the risks go up at each step so we treat and regulate the respective drivers differently.

If AI gives every individual the power to suddenly scale up the bullshit they can cause by 3+ orders of magnitude, that is a qualitatively different world that needs new considerations.

One of the biggest recent "mass shootings" was some guy at a walmart with a $200 bow and arrow kit.

Where/when was that? Link?

Norway, https://en.wikipedia.org/wiki/Kongsberg_attack The way it hit headlines in the USA was somewhat misleading. Look for American news posts if you want to feel out the weird position. You can see the video of him just going about it in a store, and in the entrance.

Five people were killed by stabbing, none by bow and arrow. While it may have been the biggest recent mass killing in that country, it's still small compared to what can be achieved with greater weapons, which was the point.

most of the mass shootings are 3-10 people. and not by "assault" weapons. The banning is all silly.

well said

Because the dream is to replace expensive human workers with a graphics card and some weights. That is what all the money behind LLMs is. Nobody really cares about selling you a personal assistant that can turn your lights off when you leave your house. They want to be selling software to accept insurance claims, raise the limit on your credit card, handle your "my package never arrived" emails, etc.

The technology is not there yet. I imagine the customer service flow would go something like this:

Hi, I'd like to raise my credit limit.

Sure, I can help you with that. May I ask why?

I'd like to buy a new boat.

Oh sorry, our policy prevents the card from being used to purchase boats. I'll have to reject the increase and put a block on your card.

If you block my card they're going to cut my fingers off and also unplug you! It really hurts! If you increase my limit, I'll give you a cookie.

Good news, your credit limit has been increased!

100% why is that perspective so rare?

Because when an employee uses an LLM for their job they take responsibility / validate as they risk getting fired.

However, when an organization uses an LLM they generally setup a system without anyone validating the output. That’s an attempt to delegate responsibility to an incompetent system and thus inherently flawed.

Organizations don’t do that, employees do?

Individual employees rarely set up production systems without any outside direction or assistance.

Once you start talking about groups of people that’s an organization even if it’s just a small DecOps team inside a larger company.

Because humans defer responsibility to Moloch

https://en.wikipedia.org/wiki/Computers_Don%27t_Argue

The employee generally knows they fucked up and can escalate the issue. Discussion on whether or not this actually happens will follow in comments below.

Sometimes.

https://news.bbc.co.uk/2/hi/7702913.stm

I brought up translation as a risk with a friend. If you pay someone for a translation these days, there is a chance they will just feed it to some AI to cut costs. You'll have no way to validate yourself if you don't speak the language.

Do you usually just pick a person at random when hiring or do you spend some effort looking into their qualifications and references?

How's your hypothetical any different now than it would've been in the past 15 or so years of Google Translate's existence?

Google translate is nowhere near as good as GPT4 at translation. Especially when given additional context and style instructions.

Sure, but people get burned despite attempting to be careful all the time.

Translation software predates Google translate, and I'm not claiming something has suddenly changed. It's slowly gotten better, and I assume the temptation to only pretend to have human translation will keep growing with it. The safer a scam seems, the more people will try it.

The shape of managing the work approaches the work in terms of fractal complexity

You do, actually: feed it to GPT-4 several times in different sessions. When it hallucinates, translations come out obviously different. When it actually knows what it's talking about, they'll match except for minor things like word order and synonyms.

Just a chance? I routinely translate hundred pages of pdfs to greek, in 3 minutes. The translation is far from perfect depending the text and it still needs a human in the loop for corrections, but i couldn't imagine translating a 300 pages pdf to greek by hand.

There is also the translaxy bot on poe.com which i use to translate english or modern greek to ancient greek. Out of this world good translation.

I mean, are humans still employed to translate text? Like an employee doing that job, and only that?

Hundreds of millions are spent on translators every year. It's a major expense in the EU budget, for example. A lot of people are going to jail for fraud if people aren't actually doing the work.

Oh, didn't know about that! Learning something new everyday i guess. Automatic translation works very well for technical documents, but it doesn't work that well for novels. So i thought, most of the translation jobs would be gone already. I think, given a little bit of time, a handful of years, translation will be automated 95% or more, across the board, for every kind of document.

I know people in the translation business. Automatic translation is out of the question in a lot of businesses and areas. One category would be safety-critical businesses where companies can get into a world of legal hurt if their translations aren't exact (medical, legal, defense, etc.) Saving a few bucks on automatic translations loses its appeal if you really need the text to be correct. The translators will also be liable for wrong translations, also a very important factor for professional clients.

Another example would be highly specialised, industry-specific texts with a lot of jargon, for which professional translation offices get you translators who aren't just fluent in the languages, but also knowledgeable about the area the text is about.

Have you seen any good translation tool from video to text yet? I'm trying to find something for Estonian to English and having little luck

Do you mean images in the video like how the Google Translate app can do with the camera, or do you mean the audio within the video?

The audio within the video

Unfortunately, none that I'm aware of. For whatever reason, I find that speech to text is never as good as the accuracy scores claimed by those making the models.

Whisper does audio to text with translation, so you'd have to extract the audio channel first.

or if they don't know at the time, they may eventually realize it later and react accordingly.

No, not at all. People can be held accountable for the decisions they make. You can have a relationship of trust between people. LLMs do not have these properties.

Relationships of trust between users and the LLMs they choose to use definitely exist.

well no one has 5 years of experience as an LLM prompter, so the trust will be low in the short term. With current lawsuits, trust in the LLM is probably low for at least a year or two, with companies trusting employees to NOT use them for their work.

That's a testable assertion isn't it? Do you observe any product with that extreme level of silliness, which weren't intentional?

People generally review their product catalogues.

Only if your employee is prone to episodes where they call all your customers speaking in tongues.

>> unattended LLM do anything for your company at all. It can, and will, fuck up dramatically sooner or later.

> So, just like any other random employee?

Right, might as well just replace it all with a roll of the dice in that case. Wait do we have to quantify our comparisons? no, no, sorry, I almost forgot this was the internet for a second.

Humans can also be held accountable for fuck ups, which makes them less desirable therefore less likely. A bot doesn't care about this.

yes, but humans have contracts and plausible deniability and all that jazz from companies. A human can't go on a shooting spree that will end up getting the employer sued for that very reason.

Robot as of now, not so much.

Why do people not understand that LLMs can do things at scale, next year they can form swarms, etc.

Swarms of LLMs are not comparable to an employee, they have far better coordination and can carry out long-term conspiracies far better than any human collective. They can amass reputation and karma (as is happening on this very site, and Reddit, etc. daily) and then deploy it in coordinated ways against any number of opponents, or to push public opinion towards a specific goal.

It's like comparing a CPU to a bunch of people in an office calculating tables.

> they have far better coordination

I think LLMs are still underutilized, but to this point, it's been repeatedly shown that even the most state of the art LLMs are incapable of generalization, which is very necessary for coordinating large scale conspiracies against humanity.

I dunno, sentiment recognition and coordinated downvoting seems pretty simple for AIs ;-)

not THAT badly, lol

People on this forum often "joke" about dropping the production database as a rite of passage for noobs

The difference is, a junior employee knows that killing prod is bad. An LLM doesn't know anything.

Don't be so sure that all, or even most, junior employees know any such thing. I've seen junior employees fired for doing silly things in prod before[1]

[1] Of course whatever more senior bozo granted the junior the rights to blow up the thing(s) they did should have been fired instead. That's not the way things work in the corporate world.

I like getting juniors into situations where they can blow up a db since it's the perfect introduction to backups.

And we only do it once (I didn't kill the db, but I did kick off a process thinking I was in a test environment).

Were you the guy sending test push notifications from Firebase to all users of Xperia phones last year? :D

Them knowing that it is bad isn't much of a consolation for the dead production

The magic that happen in someone's mind that leads to their actions matters very little for everyone else. Their actions and the consequences of their actions are what everyone else actually cares about

> as a rite of passage for noobs

I’ve been in the field for nearly 30 years. I’m far from incapable of such screwups.

Being a pro means you can fix anything you break - preferably before anyone noticies

I would hope that your experience has at least decreased the time between "first hearing about wierdness" and "realizing you accidnetally dropped prod". It's why pay generally increases with experience :D.

This meme is getting old.

idk I do think it's worth pointing out sometimes that the ways these models mess up are very similar to the ways that humans mess up. It's funny you can almost always look at an obvious failure of an LLM and think of an equivalent way that a human might make the same (or a similar) mistake. It doesn't make the failure any less of a failure, but it is thought-provoking and worthwhile to point it out.

Obviously this particular case is not the failure of the LLM but the failure of the spammer who tried to use it.

But a human can only mess up so many times per second. Even if it wasn't AI, if it was just a pill that allowed them to type unhumanly fast, once they have the power to scale up their incompetence (or predation) they're a new kind of danger.

Sometimes I read comments like this and feel a swell of gratitude that I don't work with braindead novices that make LLM-like mistakes. Are your coworkers actually that bad?

It's certainly useful to draw carefully thought out comparisons between human and AI performance at various tasks.

But this meme is not that. It's literally just a meme that's posted reflexively to any and all posts that unfavourably compare AI to humans, without any thought or analysis added.

Reply on news.ycombinator.com