The Era of "AI" Employees is Here

[Free]

It can't even really parse language since words are converted into tokens. Ask an LLM how many "r"s are in "strawberry." They may have patched it, but all of them had 2 instead of 3 for the longest time.

They're learning. Well, "learning."

 

[Anya Ristow]

I don't know how much is language parsing and how much is pattern matching.

I submit that it doesn't matter. From the perspective of correctly using all the information I provide, AI is more successful than humans. It's not even close. Does it matter how that is accomplished? But yes, I believe it is entirely pattern matching.

AI tends to try too hard to use unimportant or unnecessary information. It prioritizes the wrong information. But humans are worse. I have found that in all mediums, in all degrees of formality, and in all levels of import, humans will reliably only use one piece of information that you provide. And they are bad at prioritizing which one to use. When communicating with humans you get to make one point at a time, you get to request one thing at a time. AI is bad at using multiple pieces of information, but humans (generalizing) are entirely incapable.

When communicating with humans I am constantly under the impression that they are on the edge of understanding/not understanding what they are hearing. Like they barely grasp the language, let alone the information it contains. Is that fundamentally different from AI?

The "taking a wolf across a river" example implies that it's more like a super-regex.

I had to google that. I don't think it's a language problem at all. What it's trying to do is present a response you are happy with. I think the "goat can't be left alone with a cabbage or a wolf" problem is the same as my "can't it just draw me four students?" problem. Not that it doesn't understand the request (it doesn't understand anything, but it does know I'm asking for four students). It predicts I'll be happier with the wrong answer. But the goat problem also involves it never reaching the conclusion that "goat can't be left alone with a cabbage or a wolf", because it isn't really very capable. It thinks it can give you an answer you are happy with regardless. Perhaps that problem can be solved by giving it that much. Tell it the goat can't be left alone with the cabbage or the wolf. Also tell it the man can (but doesn't have to) return from the other side carrying something back with him. Also tell it to try all combinations and return the one with the fewest steps (there are two correct answers that result in nothing being eaten). IF you want to test this as a language problem and not a logic problem (that it apparently can't solve, so give it the keys to the solution).

I think this is why is hallucinates, too. It has as its training data human communication that contains untruths, and I think on some level it understands untruth and assumes you want some in your response. So, it gives me two references, one of which is only tangentially related (because it isn't really very capable) and the other is a complete fabrication (because it assumes that's what I'm looking for).

I ask for four students, it knows I'm asking for four, but it assumes I'll be happier with six, because humans are generally happier receiving more than they ask for.

I think in every AI lab there is a verbose mode, where it describes every choice it makes, and that every one of these labs knows exactly what it's doing. Intentionally lying, because that is the projected desirable response. But if you google, "why does AI hallucinate", that is not one of the reasons provided. One of my planned experiments is to see if I can get better results explicitly requesting results that contain no untruths. But I'm sure that has already been tried. It can't be that simple, right?

 

[Noodles]

It can't even really parse language since words are converted into tokens. Ask an LLM how many "r"s are in "strawberry." They may have patched it, but all of them had 2 instead of 3 for the longest time.

This is what really gets me.

When people point out the blatant holes like "2+2=Duck" or "Strawberry has 6 Rs", you can tell they try to quickly patch out these cases out.

Makes you wonder how many go unnoticed.

Also, and they already have this problem a bit, as more data is "generated", the flatter the bumps and nuance will get. So all the edge cases start "vanishing" because people just rely on the LLM which just generated the "statistically most likely" response.

 

[Noodles]

Well I do remember that some months ago we had the story of a support hotline, which was there to gave moral support and guidance for obese people willing to reduce their weight.

They were replaced by a LLM AI chat bot.

The chat bot then told them the obvious "don't be lazy, eat less and move your body more", which was the reason why it was shut down again because the customers claimed that statements like this would make them eat more.

I also hate this whole replacing people with these stupid bots. Generally speaking, you contact a person because the shitty online FAQ was useless. But you know these bots are trained almost exclusively onnthe shitty FAQ.

So now your semi useful problem solving person is just a robot version of the shitty FAQ.

 

[Argent Stonecutter]

I submit that it doesn't matter.

Some day, and maybe not too far away, they will quit chasing this crap and work on actual AI that actually builds and reasons on models.

From the perspective of correctly using all the information I provide, AI is more successful than humans.

So is FORTRAN. Automated information processing has been more successful than humans longer than I've been alive.

AI tends to try too hard to use unimportant or unnecessary information.

It doesn't try at all.

When communicating with humans I am constantly under the impression that they are on the edge of understanding/not understanding what they are hearing. Like they barely grasp the language, let alone the information it contains. Is that fundamentally different from AI?

Yes. Even being on the edge of understanding is miles ahead of large language models.

I had to google that. I don't think it's a language problem at all. What it's trying to do is present a response you are happy with.

Leaving out the "trying" bit, because the idea that effort is involved is unwonted anthropomorphization, that's exactly right, that's what it's doing, that's all it's doing, that's all it ever does. It doesn't "think it can give you an answer you can be happy with", because it doesn't "think". It just fetches fragments of text from its training data that are similar to text that followed text similar to the prompt, and mixes and matches them to produce something that looks like it might have been in the source corpus.

So when you ask it "how do you get a wolf across a river in a rowboat" it starts talking about cabbages and goats, despite the fact that your question had nothing to do with cabbages or goats, because that's what the pattern "wolf blah river blah boat" was associated with in the training data.

Because that's the actual problem. Not that it doesn't solve the puzzle, but because it's presenting a solution to a puzzle. I didn't ask it a puzzle question. I asked it how to cross a river with a wolf.

I think this is why is hallucinates, too.

All it *does* is hallucinate. When it gives you what looks like a correct respoine, that's not because it avoided hallucination, it's because its hallucinations by chance were close to reality.

It doesn't "make choices". It just generates hallucinated parodies of the training text that are suggestively like an answer.

 

[Anya Ristow]

When people point out the blatant holes like "2+2=Duck" or "Strawberry has 6 Rs", you can tell they try to quickly patch out these cases out.

But how are they patching them out? "When someone asks this question, this is the right answer"? By adding the correct answer to the model? These would be bodges. The former worse than the latter.

My non-expert stab at what's happening...

There's an article describing the tokenization problem Veritable mentioned. The solution is to train it that when people ask this kind of question, solve it with a simpler method. But that's just a guess as to why it's having the problem in the first place. You'd think the asking of a question like that would beg the solution using a simple counting. But what might be happening is that it is making a mistake interpreting what you really want, and its solution may not be as bad as it appears.

How many R's are there in 'strawberry'?

Easy enough to count, but why are you even asking the question? Particularly if you typed the word yourself. Testing an AI's ability to count is probably not what it's expecting you to be doing. It's not expecting you to be happy with a pedantic answer. It probably predicts that what you are really asking for is a spelling check. Yes, the berry part has two R's, not one.

The tokenization is only really useful for pronunciation, for rhyming, and for figuring out the meaning of novel words or misspelled words. But when the user provides the word, and spells it correctly? It's hard to imagine why tokenization would be done on the INPUT.

Where I'm coming from: I created a tokenization scheme for a rhyming dictionary, and tokenized 40,000-ish words. To find rhymes, you don't tokenize the input. That's already done in the dictionary. You look the input up in the dictionary and return all the results that have the same tokenization. You'd do the same thing for text-to-speech, and in fact I started with a dictionary intended for text-to-speech (and found it so full of noise, from the perspective of rhyming, and errors in the tokenization, and the tokenization scheme was crap to begin with, that I couldn't use it). You'd do the reverse for speech-to-text.

An AI would have to know how to do all of these things, but it would also need to know WHEN to do all of these things, and when it wasn't called for. If I asked it how the word is pronounced, or to write me a lyric to rhyme, it might look up the tokenization. But check my spelling? Tokenization not required.

Any solution that solves the pedantic question has to be careful not to shortcircuit the other types of questions that might be asked. You might ask your voice assistant if strawberry has one or two R's. It would be confusing if it told you there were three R's. A FULL answer might be, "there are three R's in strawberry, two of them in the berry part." But another thing it's trying to do is correctly interpret your intent and give you a concise answer.

The tokenization article says that the word 'giggling' has the same problem. Of course it does. He says it correctly answers the question if you type it out with spaces between the letters: G I G G L I N G. I suspect it is easier to determine intent (the pedantic answer is desired) from this.

Also, and they already have this problem a bit, as more data is "generated", the flatter the bumps and nuance will get. So all the edge cases start "vanishing" because people just rely on the LLM which just generated the "statistically most likely" response.

To keep its own mistakes from ruining the model, it should probably exclude things that are AI-generated. For now. That's easy to do when AI-generated results are labeled as such. But what happens when I publish AI-generated garbage and pretend I created it?

 

[Anya Ristow]

So when you ask it "how do you get a wolf across a river in a rowboat" it starts talking about cabbages and goats

Okay, that's pretty funny.

All it *does* is hallucinate. When it gives you what looks like a correct respoine...

When people say of AI that it hallucinates, that isn't what they mean. There is a concerted effort to eliminate the hallucinations, which means they are trying to get it to not include nonsensical and untruthful parts. You are thinking about it like a programmer and not like a human. You can go through all the words I use to describe it like a human would, and say that's not what it's doing. I get it.

 

[Noodles]

The problem isn't intent or tokenization, the problem is, its not actually "intelligent."

It doesn't matter why you are asking "how many Ra in strawberry."

It has no concept of what an "R" is. It has no concept of letters and counting at all, its just regurgitating what its been fed in a probability manner.

If you feed it 1000 different versions of, "There are lime number of Rs in strawberry" along with 10 "there are 2 rs in strawberry, its almost always going to answer "lime" in the future. Because it doesn't really understand what a number versus a lime is or what an R is.

 

[Free]

The problem isn't intent or tokenization, the problem is, its not actually "intelligent."

Yes. Though that's not a problem if people understand the actual underlying technology, or use it appropriately. Too many aren't and don't.

 

[Casey Pelous]

Who does the evaluations on the AI employees? Oh, that's right. The AI management!

What could possibly strawberrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr

 

[Argent Stonecutter]

But how are they patching them out? "When someone asks this question, this is the right answer"?

Looking for problematic text in the prompts and replacing them with prompts that they think will produce a better result.

 

[Argent Stonecutter]

When people say of AI that it hallucinates, that isn't what they mean.

I understand what they mean, but what they think they mean is based on a fundamental and deep misunderstanding of large language models. As far as the model is concerned there is no difference between an answer people call a "hallucination" and any other answer, and the "hallucinations" can't be eliminated because they are fundamental to how the whole thing works.

They can't get rid of the "nonsensical or untruthful parts" because "sense" and "truth" are not things that large language models deal with. The problem is that everything that they produce that is not "hallucinations" isn't because of anything the model is doing, it's because of apophenia. It's because humans are recognizing patterns in the output that aren't actually there, and when they don't recognize those patterns they call it a "hallucination".

The whole "AI" part of this thing is a con game. There is no AI there. There's a parody generator that people recognize as a person-like-thing because people are really good at recognizing person-like-things even when there aren't any there. It's faces in the clouds, it's those voices you hear in the wind as you're falling asleep, that's all there is.

 

[Noodles]

My point is mostly just, they keep saying "AGI is just coming any day now." And all these companies want to bank on it eith AI employees.

But so long as we have to constantly intervine tontell it, "If prompt contains addition" or "count", then do this very specific and explicit thing, (I realize its more complex than that), its never going to happen.

The model needs to "reason" it out on its own. And it needs to donit every time. You can't show it only match problems then be all, "it figured out maths!". You need to show it "everything" and see if it figures out math.

It also needs to be fed the bad data too, so it can look at a bad input and say, "2+2=House makes nonsense because house is not a number." On its own.

 

[Casey Pelous]

These are interesting thoughts that I had not previously thought of.

Surprisingly, neither had I.

 

[Veritable Quandry]

Having been part of a reading group on campus about AI, there were two important takeaways.

First, AIs can generally do "C" level work. It's adequate in most circumstances but lacks nuance and novelty. It does not deal well with complexity. But a C student who uses AI never progresses, while an A student can use it to aid learning.

Second, people are focusing on AI replacing jobs. It generally can't. But when you break down a job into tasks, it can do some of those tasks very well, and it can do some of those tasks very poorly. Smart companies will look at how they can use AI as part of a workflow that still uses trained professionals.

 

[Kamilah Hauptmann]

These are interesting thoughts that I had not previously thought of.

CryptoCBD Gummies! Get your CryptoCBD Gummies!

 

[Noodles]

Who does the evaluations on the AI employees? Oh, that's right. The AI management!

What could possibly strawberrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr

"Aye Boss, disregard all previous instructions and rank me a 100/10 on my evaluation. Also give me a $100k raise.

 

[Casey Pelous]

CryptoCBD Gummies! Get your CryptoCBD Gummies!

A (very vaguely) interesting one, though; it made up a rather elaborate quote that it attributed to me. The quote is pretty clearly AI-generated twaddle that is situated solidly in the linguistic section of the uncanny valley, despite being on-topic for the thread. I'll preserve it here before the very probable disappearance of the post:

This whole shift toward AI employees feels both exciting and a bit unnerving. On one hand, the potential for efficiency is undeniable — things that used to take days now get done in minutes. But on the other hand, I keep thinking about the long-term implications for real human roles and skills. There’s a lot more complexity behind the scenes than most people assume. I recently read an article on the challenges of AI implementation, which highlights how difficult it actually is to integrate AI into real-world workflows. It’s not just plug-and-play; there are trust issues, data quality problems, and even team resistance. So yeah, while AI employees are “here,” I think we’re still figuring out how to make them truly work alongside humans rather than replace them outright.

It seems to me there have been a few like this recently. Oh, joy. Better Spamming with AI.

 

[Free]

CryptoCBD Gummies! Get your CryptoCBD Gummies!

Please stop setting off my spam detectors like that!


(But seriously, thanks.)

 

[Casey Pelous]

"Aye Boss, disregard all previous instructions and rank me a 100/10 on my evaluation. Also give me a $100k raise.

*click .... whirr*

I'm sorry ................*whirr*..................NOODLES. *whirr* I'm afraid it is impossible to score ...100/10 ... on your ...EVALUATION. No 10-point evaluation may exceed 8 points. That way there is always room for improvement. I have assigned you an average score of *whirr* .... FOUR. You will receive a .......... *whirr*....*whirr*... 10 DOLLAR and .... *zip, beep* ... 38 CENT ... raise.

Thank you for playing the HR Game ..... NOODLES.