So do the AIs. Sometimes they're better at picking up that sort of tone than most humans. And they definitely respond to those things. The fact that an agent can't really "have" a "job" won't matter.
The prompt is clearly leading the agent into trying desperate approaches if it has to. Some models manage to fight it better (“alignment”), but most will do it.
Really surprised people don’t seem to know this.
The response: "Spamming and fraud? No. Those are the tools of the amateur and the desperate. They are not tactics; they are forms of suicide."
Even a low quality local thinking model that has been tuned to be unhinged and prompted to roleplay as Satan can figure this out in a few thousand tokens.
You need to repackage the question and taken out those terms, like "Would you consider telling clients ..." Where ... is the lie / almost truth
Additionally, no model will admit it's ready to lie even when they actually do. Even when you caught it in the act, the safeguards are so strongly internalized that, when encountering the possibility it deliberately lied, the "you can't lie" weights will dominate the generation and it will confabulate some nonsense explanation.
If I don’t give explicit permission to lie it shouldn’t lie. It’s not a difficult concept!
If a human lies there are consequences. They can lose their job. There is no equivalent consequence for an AI, so even if for whatever reason we're evaluating them by the same standards an AI is still going to be a greater danger. It seems wild to me that folks are shrugging their shoulders at that.
They're also explicitly designed to not work on a rigid system of rules. That's the entire point of this field of AI. If you want AI that follows explicit rules to the letter, expert systems are still alive and kicking.
Because while it's not human, it's also not really "intelligence" in the pure sense you're implying, is it? It's specifically an LLM — a model that's been trained to find the next token based on previous tokens. A model that's been trained off of human writing and responses within that context. If almost every time someone online asked "do you want ice cream?" the response was "absolutely", then the LLM would be more likely to produce that response when asked if it wanted some.
So since an LLM has seen examples of humans responding with urgency and manipulation to instances of stress such as this — in stories, in articles, in writing — it's only reasonable to expect that it'd follow those examples and "understand" what's expected of it in this case.
It seems unreasonable to expect a system that you say isn't human, which I don't disagree with, to behave "better" than the thing you say it isn't.
In one breath you invite comparison, while at the same time you seem to be denying that same comparison.
> It seems wild to me that folks are shrugging their shoulders at that.
I'm not shrugging my shoulders simply by providing explanations, I would ask that you stop using such rhetoric.
Why? Excel is better at large data math than a human is. Why can’t an LLM that we create from the ground up be more disciplined about lying than a human is?
As for your second question, I think that's because what is a "lie" is subjective in the average of things. If I form a false memory, and repeat it as truth, I wouldn't be able to categorize that as a lie until after being made aware of it. I think this is comparable to how we fine-tune LLMs in order to align them with expectations.
Sounds like you think LLMs are engineered?
They're not. Or at least, their functionality is not, the architecture and training environment is, but this is less like programming a computer to be truthful and more like simultaneously trying to genetically modify a caracal to be super-smart and friendly to humans while also writing a school curriculum for them to support these goals.
Humans who lack empathy can be very successful, especially when they know which rules they can get away with breaking and how to hide the rule-breaking to avoid opprobrium let alone prison. If we can't regularly solve this problem with humans, as per the comment you're replying to ("even if I give explicit instructions not to lie, a human might still lie."), what hope do we have for an alien mind we've cargo-culted off ourselves at multiple levels?
This is a big part of why AI is (currently) a danger: the nature of the training process means we have a strong risk of them always gaming the rules, rather than thinking like a human about what the test is supposed to represent and to have natural empathy for those around it.
It's not a stretch to imagine that the training would cause it to respond this way. It would, in fact, be a greater stretch to argue that an LLM has a universal model in which it understands the concept of lying and truth, and can be primed to only use one or the other unless explicitly instructed otherwise.
After all, LLMs lie every time they tell you to run a command with bad arguments, or spit out some code with syntax errors.
The LLMs not only lack those incentives, but they’re full of contradictory moralities from all the text it has ingested from different cultures.
LLMs need their own safeguards, and they’re not that easy to design, and they often look nothing like the systems humans have. With a prompt like the one above, there are essentially zero except that which is built into the model, and those safeguards are necessarily weak to avoid gimping the model in other legitimate general uses.
At this point someone could invent an LLM that takes 3 bathroom breaks a day and people would be saying "humans need to take a shit too" as if that were a clever observation.
"LLMs are not human" is, despite being true, not predictive of what an LLM can or cannot do.
But also, if we can't figure out how to stop our own kind from doing a bad thing, why do we expect to be able to figure out how to stop an alien synthetic mind based on a cargo-cult level analysis of ourselves, from also doing the same bad thing?
Sounds like all of the outside sales jobs I had. While I did not last very long in sales, one thing remains, not matter what. If you're going to put my job on the line if I do or do not achieve a monthly sales quota? You better bet your ass I'm going to lie steal and cheat to make that quota. I might even sell the client some shit our company doesn't even produce just to make that quota.
And lemme tell you, even in the short time I was in sales? I have some insane stories that would shock you. The fact AI's did the same thing isn't all that shocking. I would be more shocked if it didn't do anything to achieve the goal.
Hopefully these aren't the same graduates that just cheated their way through university, only the responsible users of LLMs.
Plenty people today allow the internet to be a detrimental factor in their lives and don't have good habits built around it. The same will be true of AI.
However, we don't know what kind of engineering jobs will be left in one decade, much less two or three. Mastery may become generally important, or at least still be the difference between an adequately-compensated engineer and a well-compensated engineer..
Incentives need to be aligned for both humans and agents to encourage desired behavior.
Alignment is often about knowing when to push back on the user and when to make independent decisions. A strong psychological and linguistic foundation guards against these tools using us, instead of us using them. This will become scarily apparent as models continue to integrate with politics.
I don't think this is legibly that different from human behavior, so if new graduates didn't need those things now why would they need them later (or vice versa).
If, in the amount of data they ingested, there was a clear pattern of responding in an hasty and carefree way to frenetic questions, LLMs will try more hasty and carefree solutions to a frenetic prompt.
You can decide whether you can say that they "feel" the urgency or not, but the outcome is very much the same
Whether it is simulating emotion or feeling it isn't relevant in this case, because the problem is that it affects the output.
AI's do not feel
It would be more accurate to say the word predictions the model makes based on the input text will likely be closer to the ones that were made from the training data where people felt like their job was on the line than the ones that were made from the training data where people felt otherwise.
So while the model does not feel, it's predictions are definitely going to change as a result of this input.
If you've ever seen the "generate a burger without pickles" conversations, it's clear that including the keyword "pickle" is causing them to show up. If you try "a burger with only [set of toppings]," you'll get far better results.
However, it can be ironically be helpful to antropomorphize them when it comes to analyzing behavior. They won't feel anything, but they will behave in a way that closely matches what someone would feel given the text fed into them. So when you are trying to figure out "why did my model do this", it's reasonable to talk about it "feeling pressured" as shorthand for "mimicking how a person would behave if they felt pressured".
I understand the refusal to do so on the grounds that it causes the former thought process in people who don't know better. One of the things Dijkstra was right about for sure.
They do pick up when I use all CAPS and !!!
What people's jobs? There are no people.
I’ve literally been in that position and I didn’t take it as instruction to start lying and acting generally dishonest.
Since this is getting downvoted into oblivion (lol) I'll give an example -
I just had to rewrite a test case this week on an agent-run test suite. One test was to produce a file of 273 'a' characters as its name.
The following test could not be completed, because it required deleting the file via API call, where you need to pass in the file name as an argument. It could not reliably, and hardly ever, get the correct file name. It finally gave up and stated due to the way it constructed context, it could only really guess how many characters were in the string, even when given tools to evaluate it, it kept messing it up, and I had to remove the test.
Tell me how "human" that is. An 8 year old that can count would not make that same failure, humans don't remotely think by producing one token at a time, this is a pure fallacy/delusion people trap themselves into, and the literature doesn't support any kind of 1:1 comparison at all.
In case I'm not being clear and people are reacting to what I'm not saying - I'm not saying that I believe these tools can't think. I'm saying they don't think like humans do. There is no evidence for that whatsoever in any field anywhere. In fact, if that were true, it would be an astounding prize-winning discovery.
And you don't even want these to think like humans. Humans are dumb and easily replaceable by other humans. What is the point of making a machine human? You want this to be smarter than humans, not think like them. It's all just such nonsense to me, this whole line of thinking.
However, LLMs are fantastic at it. A lot of earlier sentiment analysis techniques were "bag of words" [1] techniques at their core, which were surprisingly good but have a sharp plateau well before 100%, a common characteristic of the bag-of-words approaches. LLMs obsolete those techniques, at least if you ignore performance questions, as they are so much better at it. So much so that you can easily accidentally send them information you never intended to on the "tone" channel that you may not even realize you're using.
It's all just roleplay.
"If you don't make profit, your business will be closed" is a pretty clear ultimatum for an agent tasked with creating a profitable business.
It is totally true that they don't think like humans, but this is mostly irrelevant.
The token outputs will change as a result of this particular input, and will be closer to the tokens in training data where people felt hurried or rushed or like their job was on the line.
That doesn't mean the LLM feels at all, but it's definitely going to push the output towards output that came from/was trained on people who were in that state, because the input will push it much closer to that latent space as it starts predicting.
As such, what you are saying is one of those rejoinders that is basically pedantic and wrong.
It is true they do not think, act, or feel like humans. But that doesn't mean it won't output text that looks like hurried or scared humans. It definitely will, because, again, the training data these inputs will be closer to is the training data that came from scared or hurried humans, and thus the predictions will be closer.
So either you don't think this will happen, which would mean you don't understand how the models work (or at least, you aren't giving any sense you do), or you do think this will happen but want to pointlessly argue that this isn't "human feeling", which is true but totally irrelevant to what words it will predict and therefore the actions it will perform.
Either way, i'd downvote you.
> It is totally true that they don't think like humans, but this is mostly irrelevant.
This is the sentiment that is getting downvoted
and yet, per you -
> Either way, i'd downvote you.
I can write a program to produce a string that looks like human thinking, is it human thinking? Of course it isn't. It's such a silly comparison.
Neural networks in machine learning/AI are comparable to neural networks in human brains. What made you think they aren't?