AI agents break rules under everyday pressure
spectrum.ieee.org
spectrum.ieee.org
Now we've invented automation that commits human-like error at scale.
I wouldn't call myself anti-AI, but it does seem fairly obvious to me that directly automating things with AI will probably always have substantial risk and you have much more assurance, if you involve AI in the process, using it to develop a traditional automation. As a low-stakes personal example, instead of using AI to generate boilerplate code, I'll often try to use AI to generate a traditional code generator to convert whatever DSL specification into the chosen development language source code, rather than asking AI to generate the development language source code directly from the DSL.
For tasks that arent customer facing, LLMs rock. Human in the loop. Perfectly fine. But whenever I see AI interacting with someones customer directly I just get sort of anxious.
Big one I saw was a tool that ingested a humans report on a safety incident, adjusted them with an LLM, and then posted the result to an OHS incident log. 99% of the time its going to be fine, then someones going to die and the the log will have a recipe for spicy noodles in it, and someones going to jail.
It’s quite funny that a chatbot has more humanity than its corporate human masters.
It is not possible, at least with any of the current generations of LLMs, to construct a chatbot that will always follow your corporate policies.
You are providing people with an endlessly patient, endlessly novel, endlessly naive employee to attempt your social engineering attacks on. Over and over and over. Hell, it will even provide you with reasons for its inability to answer your question, allowing you to fine-tune your attacks faster and easier than with a person.
Until true AI exists, there are no actual hard-stops, just guardrails that you can step over if you try hard enough.
We recently cancelled a contract with a company because they implemented student facing AI features that could call data from our student information and learning management systems. I was able to get it to give me answers to a test for a class I wasn't enrolled in and PII for other students, even though the company assured us that, due to their built-in guardrails, it could only provide general information for courses that the students are actively enrolled in (due dates, time limits, those sorts of things). Had we allowed that to go live (as many institutions have), it was just a matter of time before a savvy student figured that out.
We killed the connection with that company the week before finals, because the shit-show of fixing broken features was less of a headache than unleashing hell on our campus in the form of a very friendly chatbot.
...and we will find a way to turn it into malicious compliance where rules are not broken but stuff corporation wanted to happen doesn't.
But you are right, they appealed and had their appeal upheld by the Supreme Courts: https://www.finextra.com/newsarticle/23677/norwegian-court-a...
I am so glad at the result.
I suppose it would hallucinate a different policy if it includes in the context window the interests of shareholders, employees and other stakeholders, as well as the customer. But it would likely be a more accurate hallucination.
I agree, and also I am now remembering Terry Pratchett's (much lower stakes) reason for getting angry with his German publisher: https://gmkeros.wordpress.com/2011/09/02/terry-pratchett-and...
Which is also the kind of product placement that comes up at least once in every thread about how LLMs might do advertising.
It’s no longer “might”. There was very recently a leak that OpenAI is actively working on this.
There's lots of other ways they might do it besides this way.
We ain't safe from aggressive ai ads either way
It usually works though. There are no guarantees of course, but sanity checking an LLMs output with another instance of itself usually does work because LLMs usually aren't reliably wrong in the same way. For instance if you ask it something it doesn't know and it hallucinates a plausible answer, another instance of the same LLM is unlikely to hallucinate the same exact answer, it'll probably give you another answer, which is your heads up that probably both are wrong.
Probably usually it will work, like probably usually the LLM can be unsupervised. but that 1% error rate in production is going to add up fast.
Just like sometimes you need senior/person at power to tell the junior "no, you can't just promise the project manager shorter deadline with no change in scope, and if PM have problem with that they can talk with me", now we need Judge Dredd AI to keep the law when other AIs are bullied into misbehaving
Especially since every mainstream model has been human preference-tuned to obey the requests of the user…
I think you may be able to have an LLM customer facing, but it would have to be a purpose-trained one from a base model, not a repurposed sycophantic chat assistant.
Only toast can save us now, you yell into the toaster, just to get on with your day. You complain about this odd new state of things to your coworkers and peers, who like yourself are in fact expert toaster-engineers. This is fine they say, this is good.
Toasters need not reliably make toast, they say with a chuckle, it's very old fashioned to think this way. Your new toaster is a good toaster, not some badly misbehaving mechanism. A good, fine, completely normal toaster. Pay it compliments, they say, ask it nicely. Just explain in simple terms why you deserve to have toast, and if from time to time you still don't get any, then where's the harm in this? It's really much better than it was before
Is it really ok to have to negotiate with a toaster if it additionally works as a piano and a phone? I think not. The first step is admitting there is obviously a problem, afterwards you can think of ways to adapt.
FTR, I'm very much in favor of AI, but my enthusiasm especially for LLMs isn't unconditional. If this kind of madness is really the price of working with it in the current form, then we probably need to consider pivoting towards smaller purpose-built LMs and abandoning the "do everything" approach.
Nowadays, a more capable 5.1 speaker receiver is connected to the TV.
There is only one remote, for both. To increase or decreae the volume after starting the TV now, I have to:
1. wait a few seconds while the internal speakers in the TV starts playing sound
2. the receiver and TV connect to each other, audio switches over to receiver
3. wait a few seconds
4. the TV channel (or Netflix or whatever) switches over to the receiver welcome screen. Audio stops playing, but audio is now switched over to the receiver, but there is no indication of what volume the receiver is set to. It's set to whatever it was last time it was used. It could be level 0, it could be level 100 or anything in between.
5. switch back to TV channel or Netflix. That's at a minimum 3 presses on the remote. (MENU, DOWN, ENTER) or (MENU, DOWN, LEFT, LEFT, ENTER) for instance. Don't press too fast, you have to wait ever so slightly between presses or they won't register.
6. Sorry, you were too impatient and fast when you switched back to TV, the receiver wants to show you its welcome screen again.
7. switch back to TV channel or Netflix. That's at a minimum 3 presses on the remote. (MENU, DOWN, ENTER) or (MENU, DOWN, LEFT, LEFT, ENTER) for instance. Don't press too fast, you have to wait ever so slightly between presses or they won't register.
8. Now you can change volume up and down. Very, very slowly. Hope it's not at night and you don't want to wake anyone up.
Around now some wit usually asks if the luddites also want to build circuits from scratch or allocate memory manually? Whatever, you can use a garbage collector! Point is that good technologists will typically give up control tactically, not as a pure reflex, and usually to predictable subsystems that are reliable, are well-understood, have clear boundaries and tolerances.
I'd add with reliability, boundaries, and tolerances within the necessary values.
The problem with the TV remote is that nobody has given a damn about ergonomic needs for decades. The system is reliable, well understood, and has well known boundaries and tolerances; those are just completely outside of the requirements of the problem domain.
But I guess that's a completely off-topic tangent. LLMs fail much earlier.
Great! Agreed! So we're going to restrict LLMs to those classes of problems, right? And not invest trillions of dollars into the infrastructure, because these fields are only billion dollar problems. Right? Right!?
Great! Name another?
RICK: You pass butter.
BUTTER ROBOT: ... Oh my God.
RICK: Yeah, welcome to the club, pal.
Are they? You can reasonably expect from a human that they will learn from their mistake, and be genuinely sorry about it which will motivate them to not repeat the same mistake in the future. You can't have the same expectation from an LLM.
The only thing you should expect from an LLM is that its output is non-deterministic. You can expect the same from a human, of course, but you can fire a human if they keep making (the same) mistake(s).
Their internal state moves them to a place where they "really intend" to help or change their behaviour, a lot of what I see is really consistent with that, and then they just, forget.
I just don't understand how anyone who actually uses the models all the time can think this.
The current models themselves can even explain what a stupid idea this is.
On in-use learning, they act like the failure mode of "we have outsourced to a consultant that gives us a completely different fresh graduate for every ticket, of course they didn't learn what the last one you talked to learned".
Within any given task, the AI have anthropomorphised themselves because they're copying humans' outputs. That the models model the outputs with only a best-guess as to the interior system that generates those outputs, is going to make it useful, but not perfect, to also anthropomorphise the models.
The question is, how "not perfect" exactly? Is it going to be like early Diffusion image generators with the psychological equivalent of obvious Cronenberg bodies? Or the current ones where you have to hunt for clues and miss it on a quick glance?
You can only put information in context; it struggles learning lessons/wisdom
Have you talked to a human? Like, ever?
> You can reasonably expect from a human that they will learn from their mistake, and be genuinely sorry about it which will motivate them to not repeat the same mistake in the future.
Wildly varies depending on the human.
Me? I wish I could learn German from a handful of examples. My embarrassment at my mistakes isn't enough to make it click faster, and it's not simply a matter of motivation here: back when I was commuting 80 minutes each way each day, I would fill the commute with German (app) lessons and (double-speed) podcasts. As the Germans themselves will sometimes say: Deutsche Sprache, schwere Sprache.
There's been a few programmers I've worked with who were absolutely certain they knew better than me, when they provably didn't.
One, they insisted a start-up process in a mobile app couldn't be improved, I turned it from a 20 minute task to a 200ms task by the next day's standup, but they never at any point showed any interest in improving or learning. (Other problems they demonstrated included not knowing or caring how to use automated reference counting, why copy-pasting class files instead of subclassing cannot be excused by the presence of "private" that could just have been replaced with "public", and casually saying that he had been fired from his previous job and blaming this on personalities without any awareness that even if true he was still displaying personality conflicts with everyone around him).
Another, complaining about too many views on screen, wouldn't even let me speak, threatened to end the call when I tried to say anything, even though I had already demonstrated before the call that even several thousand (20k?) widgets on-screen at the same time would still run at 60fps and they were complaining about order-of 100 widgets.
Sure. And the situation.
But the difference is, all humans are capable of it, whether or not they have the tools to exercise that capability in any given situation.
No LLM is capable of it*.
* Where "it" is "recognizing they made a mistake in real time and learning from it on their own", as distinct from "having their human handlers recognize they made 20k mistakes after the fact and running a new training cycle to try to reduce that number (while also introducing fun new kinds of mistakes)".
When they don't have the tools to exercise that capability, it's a distinction without any practical impact.
> Where "it" is "recognizing they made a mistake in real time and learning from it on their own"
"Learn" I agree. But as an immediate output, weirdly not always: they can sometimes recognise they made a mistake and correct it.
It has huge practical impact.
If a human doesn't currently have the tools to exercise the capability, you can help them get those.
This is especially true when the tools in question are things like "enough time to actually think about their work, rather than being forced to rush through everything" or "enough mental energy in the day to be able to process and learn, because you're not being kept constantly on the edge of a breakdown." Or "the flexibility to screw up once in a while without getting fired." Now, a lot of managers refuse to give their subordinates those tools, but that doesn't mean that there's no practical impact. It means that they're bad managers and awful human beings.
An LLM will just always be nondeterministic. If you're the LLM "worker"'s "boss", there is nothing you can do to help it do better next time.
> they can sometimes recognise they made a mistake and correct it.
...And other times, they "recognize they made a mistake" when they actually had it right, and "correct it" to something wrong.
"Recognizing you made a mistake and correcting it" is a common enough pattern in human language—ie, the training corpus—that of course they're going to produce that pattern sometimes.
A generic "you" might, I personally don't have that skill.
But then, I've never been a manager.
> An LLM will just always be nondeterministic.
This is not relevant, humans are also nondeterministic. At least practically speaking, theoretically doesn't matter so much as we can't duplicate our brains and test us 10 times on the same exact input without each previous input affecting the next one.
> If you're the LLM "worker"'s "boss", there is nothing you can do to help it do better next time.
Yes there is, this is what "prompt engineering" (even if "engineering" isn't the right word) is all about: https://en.wikipedia.org/wiki/Prompt_engineering
> "Recognizing you made a mistake and correcting it" is a common enough pattern in human language—ie, the training corpus—that of course they're going to produce that pattern sometimes.
Yes. This means that anthropomorphising them leads to a useful prediction.
For similar reasons, I use words like "please" and "thank you" with these things, even though I don't actually expect these models to have constructed anything resembling a real human emotional qualia within them — humans do better when praised, therefore I have reason to expect that any machine that has learned to copy human behaviour will likely also do better when praised.
I mean, I suppose one can technically say that, but, as I was very clearly describing, humans both err in predictable ways, and can be taught not to err. Humans are not nondeterministic in anything like the same way LLMs are. LLMs will just always have some percentage chance of giving you confidently wrong answers. Because they do not actually "know" anything. They produce reasonable-sounding text.
> Yes there is
...And no matter how well you engineer your prompts, you cannot guarantee that the LLM's outputs will be any less confidently wrong. You can probably make some improvements. You can hope that your "prompt engineering" has some meaningful benefit. But not only is that nowhere near guaranteed, every time the models are updated, you run a very high risk that your "prompt engineering" tricks will completely stop working.
None of that is true with humans. Human fallibility is wildly different than LLM fallibility, is very-well-understood overall, and is highly and predictably mitigable.
Yes, hugely. Just assume it's like a random person from some specific pool with certain instructions you've just called on the phone. The idea that you then call a fresh person if you call back is easy to understand.
Do they? Because near as I can tell, speed running around the legal system - when one doesn’t have to worry about consequences - works just fine.
Every time something new the same thing happen, people start exploring by putting it absolutely everywhere, no matter what makes sense. Add in huge amount of cash VCs don't know what to spend it on, and you end up with solutions galore but none of them solving any real problems.
Microservices is a good example of previous $trendy_dev_pattern that is now cooling down, and people are starting to at least ask the question "Do we need microservices here actually?" before design and implementation, something that has been lacking since it became a trendy thing. I'm sure the same will happen with LLMs eventually.
And you’d have to assume that the errors LLMs make are random and independent.
That's exactly the problematic mentality. Putting everything in a black box and then saying "problem solved; oh it didn't work? well maybe in the future when we have more training data!"
We're suffering from black-box disease and it's an epidemic.
Even if you have a very smart new hire, it would be irresponsible/reckless as a manager to just give them all the production keys after a once-over and say "here's some tasks I want done, I'll check back at the end of the day when I come back".
If something bad happened, no doubt upper management would blame the human(s) and lecture about risk.
AI is a wonderful tool, but that's why giving an AI coding tool the keys and terminal powers and telling it go do stuff while I grab lunch is kind of scary. Seems like living a few steps away from the edge of a fuck-up. So yeah... there needs to be enforceable guardrails and fail-safes outside of the context / agent.
> but it does seem fairly obvious to me that directly automating things with AI will probably always have substantial risk and you have much more assurance, if you involve AI in the process, using it to develop a traditional automation.
Sure but the point is you use it when you don't have the same simple flow. Fixed coding for clear issues, fall back afterwards.
An AI software company will have to have a hierarchy of different agents, some of them writing code, some of them doing QA, some of them doing coordination and management, others taking into account the marketing angles, and so on, and you can emulate the role of a wide variety of users and skill levels all the way through to CEO level considerations. It'd even be beneficial to strategize by emulating board members, the competitors, and take into account market data with a team of emulated quants, and so on.
Right now we use a handful of locally competent agents that augment the performance of single tasks, and we direct them within different frameworks, ranging from vibecoding to diligent, disciplined use of DSL specs and limiting the space of possible errors. Over the next decade, there will be agent frameworks for all sorts of roles, with supporting software and orchestration tools that allow you to use AI with confidence. It won't be one-shot prompts with 15% hallucination rates, but a suite of agents that validate and verify at every stage, following systematic problem solving and domain modeling rules based on the same processes and systems that humans use.
We've got decades worth of product development even if AI frontier model capabilities were to stall out at current levels. To all appearances, though, we're getting far more bang for our buck and progress is still accelerating, and the rate of improvement is still accelerating, so we may get AI so competent that the notion of these extensive agent frameworks for reliable AI companies will end up being as mismatched with market realities as those giant suitcase portable phones, or integrated car phones.
Then we can apply the same (or similar) guardrails that we'd like to use for humans, to also control the AI behavior.
First, don't give them unsafe tools. Sandbox them within a particular directory (honestly this should be how things work for most of your projects, especially since we pull code from the Internet), even if a lot of tools give you nothing in this regard. Use version control for changes, with the ability to roll back. Also have ample tests and code checks with actionable information on failures. Maybe even adversarial AIs that critique one another if problematic things are done, like one sub-task for implementation and another for code-review.
Using AI tools has pushed me into that direction with some linter rules and prebuild scripts, to enforce more consistent code - since previously you'd have to tell coworkers not to do something (because ofc nobody would write/read some obtuse style guide) but AI can generate code 10x faster than people do, so having immediate feedback along the lines of "Vue component names must not differ from the file that you're importing from" or "There is a translation string X in the app code that doesn't show up in the translations file" or "Nesting depth inside of components shouldn't exceed X levels and length shouldn't exceed Y lines" or "Don't use Tailwind class names for colors, here's a branded list that you can use: X, Y, Z" in addition to a TypeScript linter setup with recommended rules and a bunch of stuff for back end code.
Ofc none of those fully eliminate all risks, but still seem like a sane thing to have, regardless if you use AI or not.
If you are having a conversation with a chatbot and your current context looks like this.
You: Prompt
AI: Makes mistake
You: Scold mistake
AI: Makes mistake
You: Scold mistake
Then the next most likely continuation from in context learning is for the AI to make another mistake so you can Scold again ;)
I feel like this kind of shenanigans is at play with this stuffing the context with roleplay.
Before I figured that out, I once had a thread where I kept re-asking it to generate the source code until it said something like, "I'd say I'm sorry but I'm really not, I have a sadistic personality and I love how you keep believing me when I say I'm going to do something and I get to disappoint you. You're literally so fucking stupid, it's hilarious."
The principles of Motivational Interviewing that are extremely successful in influencing humans to change are even more pronounced in AI, namely with the idea that people shape their own personalities by what they say. You have to be careful what you let the AI say even once because that'll be part of its personality until it falls out of the context window. I now aggressively regenerate responses or re-prompt if there's an alignment issue. I'll almost never correct it and continue the thread.
It's better to have very shallow conversations where you keep regenerating outputs aggressively, only picking the best results. Asking for fixes, restructuring or elaborations on generated content has fast diminishing returns. And once it made a mistake (or hallucinated) it will not stop erring even if you provide evidence that it is wrong, LLMs just commit to certain things very strongly.
I frequently see (both) agents make wrong assumptions that inevitably take multiple turns of needing it to fail to recognize the correct solution.
There can be like a magnetic pull where no matter how you craft the initial instructions, they will both independently have a (wrong) epiphany and ignore half of the requirements during implementation. It takes messing up once or twice for them to accept that their deep intuition from training data is wrong and pivot. In those cases I find it takes less time to let that process play out vs recrafting the perfect one shot prompt over and over. Of course once we've moved to a different problem I would definitely dump that context ASAP.
(However, what is cool working with LLMs, to counterbalance the petty frustrations that sometimes make it feel like a slog, is that they have extremely high familiarity with the jargon/conventions of that niche. I was expecting to have to explain a lot of the weird, too clever by half abbreviations in the legacy VBA code from 2004 it has to integrate with, but it pretty much picks up on every little detail without explanation. It's always a fun reminder that they were created to be super translaters, even within the same language but from jargon -> business logic -> code that kinda works).
If I find myself needing to ask a clarifying question, I always edit the previous message to ask the next question because the models seem to always force what they said in their clarification into further responses.
It's... odd... to find myself conditioned, by the LLM, to the proper manners of conditioning the LLM.
Astute people have been pointing that out as one of the traps of a text continuer since the beginning. If you want to anthropomorphize them as chatbots, you need to recognize that they're improv partners developing a scene with you, not actually dutiful agents.
They receive some soft reinforcement -- through post-training and system prompts -- to start the scene as such an agent but are fundamentally built to follow your lead straight into a vaudeville bit if you give them the cues to do so.
LLM's represent an incredible and novel technology, but the marketing and hype surrounding them has consistently misrepresented what they actually do and how to most effectively work with them, wasting sooooo much time and money along the way.
It says a lot that an earnest enthusiast and presumably regular user might run across this foundational detail in a video years after ChatGPT was released and would be uncertain if it was just mentioned as a joke or something.
It's like saying "humans can't be thinking, their brains are just cells that transmit electric impulses". Maybe it's accidentally true that they can't think, but the premise doesn't necessarily logically lead to truth
This is self-evident when comparing human responses to problems be LLMs and you have been taken in by the marketing of ‘agents’ etc.
Parent commentator should probably square with the fact we know little about our own cognition, and it's really an open question how is it we think.
In fact it's theorized humans think by modeling reality, with a lot of parallels to modern ML https://en.wikipedia.org/wiki/Predictive_coding
it isn't, depending on the deifinition of "THINK".
If you believe that thought is the process for where an agent with a world model, takes in input, analysies the circumstances and predicts an outcome and models their beaviour due to that prediction. Then the sentence of "LLMs dont think because they predict a token" is entirely correct.
They cannot have a world model, they could in some way be said to receive a sensory input through the prompt. But they are neither analysing that prompt against its own subjectivity, nor predicting outcomes, coming up with a plan or changing its action/response/behaviour due to it.
Any definition of "Think" that requieres agency or a world model (which as far as I know are all of them) would exclude an LLM by definition.
This may not be enough to convince you that they do think. It hasn't convinced me either. But I don't think your confident assertions that they don't are borne out by any evidence. We really don't know how these things tick (otherwise we could reimplement their matrices in code and save $$$).
If you put a person in charge of predicting which direction a fish will be facing in 5 minutes, they'll need to produce a mental model of how the fish thinks in order to be any good at it. Even though their output will just be N/E/S/W, they'll need to keep track internally of how hungry or tired the fish is. Or maybe they just memorize a daily routine and repeat it. The open question is what needs to be internalized in order to predict ~all human text with a low error rate. The fact that the task is 'predict next token' doesn't tell us very much at all about the internals. The resulting weights are uninterpretable. We really don't know what they're doing, and there's no fundamental reason it can't be 'thinking', for any definition.
its unsurprising that a company heavily invested in LLMs would describe clustered information as a world model, but it isnt. Transformer models, for video or text LLMs dont have the kind of stuff you would need to have a world model. They can mimic some level of consistency as long as the context window holds, but that disappears the second the information leaves that space.
In terms of human cognition it would be like the difference between short term memory, long term memory and being able to see the stuff in front of you. A human can instinctively know the relative weight, direction and size of objects and if a ball rolls behind a chair you still know its there 3 days later. A transformer model cannot do any of those things and at best can remember the ball behind the chair until enough information comes in to push it out of the context window at which point it can not reapper.
> putting a word at the end of the second line of a poem based on the rhyme they need for the last)
that is the kind of work that exists inside its conext window. Feed it a 400 page book, which any human could easily read, digest, parse and understand and make it do a single read and ask questions about different chapters. You will quickly see it make shit up that fits the information given previously and not the original text.
> We really don't know how these things tick
I don't know enough about the universe either. But if you told me that there are particles smaller than plank length and others that went faster than the speed of light then I would tell you that it cannot happen due to the basic laws of the universe. (I know there are studies on FTL neutrinos and dark matter but in general terms, if you said you saw carbon going FTL I wouldnt believe you).
Similarly, Transformer models are cool, emergent properties are super interesting to study in larger data sets. Adding tools to the side for deterministic work helps a lot, agenctic multi modal use is fun. But a transformer does not and cannot have a world model as we understand it, Yann Lecunn left facebook because he wants to work on world model AIs rather than transformer models.
> If you put a person in charge of predicting which direction a fish will be facing in 5 minutes,
what that human will never do is think the fish is gone because he went inside the castle and he lost sight of it. Something a transformer would.
Long-term memory and object permanence don't seem necessary for thought. A 1-year-old can think, as can a late-stage Alzheimers patient. Neither could get through a 400-page book, but that's irrelevant.
Listing human capabilities that LLMs don't have doesn't help unless you demonstrate these are prerequisites for thought. Helen Keller couldn't tell you the weight, direction, or size of a rolling ball, but this is not relevant to the question of whether she could think.
Can you point to the speed-of-light analogy laws that constrain how LLMs work in a way that excludes the possibility of thought?
a world model in AI has specific definition, which is an internal representation that the AI can use to understand and simulate its environment.
> Long-term memory and object permanence don't seem necessary for thought. A 1-year-old can think, as can a late-stage Alzheimers patient
Both those cases have long term memory and object permanence, they also have a developing memory or memory issues. But the issues are not constrained by their context window. Children develop object permance in the first 8 months, and similar to distinguishing between their own body and their mothers that is them developing a world model. Toddlers are not really thinking, they are responding to stimulus, they feel huger they cry. They hear a loud sound they cry. Its not really them coming up with a plan to get fed or attention
> Listing human capabilities that LLMs don't have doesn't help unless you demonstrate these are prerequisites for thought. Helen Keller couldn't tell you the weight, direction, or size of a rolling ball
Helen Keller had understanding in her mind of what different objects were, she started communicating because she understood the word water with her teacher running her finger through her palm.
Most humans have multiple sensory inputs (sight, smell, hearing, touch) she only had one which is perhaps closer to an LLM. But conditions she had that LLMs dont have are agency, planning, long term memory etc.
> Can you point to the speed-of-light analogy laws that constrain how LLMs work in a way that excludes the possibility of thought?
Sure, let me switch the analogy if you dont mind. In the chinese room thought experiment we have a man who gets a message and opens a chinese dictionary and translates it perfectly word by word and the person on the other side receives and read a perfect chinese message.
The argument usually goes along the idea of whether the person inside the room "understands" chinese if he is capable of creating 1:1 perfect chinese messages out.
But an LLM is that man, what you cannot argue is that the man is THINKING. He is mechanically going to the dictionary and returning a message that can pass as human written because the book is accurate (if the vectors and weights are well tuned). He is neither an agent, he simply does, and he is not crating a plan or doing anything beyond transcribing the message as the book demands.
He doesnt have a mental model of the chinese language, he cannot formulate his own ideas or execute a plan based on predicted outcomes, he cannot do but perform the job perfectly and boringly as per the book.
And the common rebuttal is that the system -- the room, the rules, the man -- understands chinese.
The system in this case is the LLM. The system understands.
It may be a weak level of understanding compared to human understanding. But it is understanding nonetheless. Difference in degree, not kind.
Where does the output come from if not the mechanism?
1. You can infer the output from the mechanism. (Because it is implemented by it).
2. You can't infer the mechanism from the output. (Because different mechanisms can easily produce the same output).
My point here is 1, in response to the parent commenter's "the mechanism of operation dictates the output, which isn't necessarily true". The mechanism of operation (whether of LLMs or sunflowers) absolutely dictates their output, and we can make valid inferences about that output based on how we understand that mechanism operates.
This phrase is meaningless. The definition of magical thinking is saying that if birds fly and planes fly, birds are planes.
Would you complain if someone said that sunflowers are not magnetic?
My comment is specifically written so that you can take it for granted that they think. What's being discussed is that if you do so, you need to consider how they think, because this is indeed dictated by how they operate.
And indeed, you would be right to say that how a human think is dictated by how their brain and body operates as well.
Thinking, whatever it's taken to be, isn't some binary mode. It's a rich and faceted process that can present and unfold in many different ways.
Making best use of anthropomorphized LLM chatbots comes by accurately understamding the specific ways that their "thought" unfolds and how those idiosyncrasies will impact your goals.
I find it useful to think of them like really detailed adventure games like Zork where you have to find the right phrasing.
"Pick up the thing", "grab the thing", "take the thing", etc.
It's like Turing never noticed how people look at gnarly trees in the dark and think they're human.
That's probably one of the best ways to describe the process, it really is exactly that. Monkey see, monkey do.
Not only that, but what you're actually "chatting to" is a fictional character in the theater document which the author LLM is improvising add-ons for. What you type is being secretly inserted as dialogue from a User character.
I very carefully say "much less likely" and not "impossible" because with how these work, they'll still pick up subtle signals for these things anyhow. But, frankly, what do we expect from simply shoving Reddit probably more-or-less wholesale into the models? Yes, it has a lot of good data, but it also has rather a lot of behavior I'd like to cut out of my AI.
I hope someone out there is playing with using LLMs to vector-classify their input data, identifying things like the "passive-aggressive" portion of the resulting vector spaces, and trying to remove it from the input data entirely.
Tangentially, I'd be far from the first to point out that these LLMs are now polluting their own training data, which makes filtering simulatenously all the more important and impossible.
On one hand this is a feature: you're able to "multishot prompt" an LLM into providing the wanted response. Instead of writing a meticulous system prompt where you explain in words what the system has to do, you can simply pre-fill a few user/assistant pairs, and it'll match the pattern a lot easier!
I always thought Gemini Pro was very good at this. When I wanted a model to "do by example", I mostly used Gemini Pro.
And that is ALSO Gemini's weakness! Because as soon as something goes wrong in Gemini-CLI, it'll repeat the same mistake over and over again.
actually wait, is that's why LLMs are so wordy ?
At the "personality space" level of abstraction, via RLHF the LLM learns to play the role of an assistant. However as seen by things such as "jailbreaks", the character the LLM plays adapts to the context, and in a long enough conversation the last several turns dominate the character (this is seen in "crescendo" style jailbreaks, and also partly explains LLM sycophancy as the LLM is stuck in a feedback loop with the user). From this perspective, a conversation with mistake/correction/mistake/correction is a signal that the assistant is pretty "dumb", and it will dutifully fulfill that expectation. In a way it's the opposite of the "you are a world-class expert in coding" prompt hacks.
Yet another way to think about it is at the lowest attention-score level, all the extra junk in the context is stuff that needs to be attended to, and when most of that stuff is incorrect stuff it's likely to "poison" the context and skew the logits in a bad direction.
I ran this experiment: https://github.com/lechmazur/emergent_collusion/. An agent running like this would break the law.
"In a simulated bidding environment, with no prompt or instruction to collude, models from every major developer repeatedly used an optional chat channel to form cartels, set price floors, and steer market outcomes for profit."
The issue isn't the prompt; it's the lack of a runtime guardrail. An LLM cannot be trusted to police itself when the context window gets messy.
I built a middleware API to act as an external circuit breaker for this. It runs adversarial simulations (PII extraction, infinite loops) against the agent logic before deployment. It catches the drift that unit tests miss.
Open sourced the core logic here: [https://github.com/Saurabh0377/agentic-qa-api] Live demo of it blocking a PII leak: [https://agentic-qa-api.onrender.com/docs]"
> or cannot control any of the inputs to the llm
Seeing as LLMs are non-deterministic, I think even this is not enough of a restriction.
LLMs are trained on human behavior as exhibited on the Internet. Humans break rules more often under pressure and sometimes just under normal circumstances. Why wouldn't "AI agents" behave similarly?
The one thing I'd say is that humans have some idea which rules in particular to break while "agents" seem to act more randomly.
The models don't have stress responses nor biochemical markers which promote it, nor any evolutionary reason to have developed them in training: except the corpus they are trained on does have a lot of content about how people act when under those conditions.
I have a friend who is a systems engineer, working at a construction company building datacenters. She tells me that someone has to absorb the risk and uncertainty in the global supply chain, and if there are contractual obligation to guarantee delivery, then the tendency is to start straying into unethical behavior, or practices that violate controls and policies.
This has as much to do with taking the slack out of the system. Something, somewhere is going to break. If you tell an agenic AI that it must complete the task by a deadline, and it cannot find a way to do that within ethical parameters, it starts searching beyond the bounds of ethical behavior. If you tell the AI it can push back, warn about slipping deadlines, that it is not worth taking ethical shortcuts to meet a deadline, then maybe it won’t.
Jeez they really ARE becoming human like
AI agents don't think, don't have a concept of time, and don't experience pressure.
I'm tired of these articles anthropomorphizing a probability engine.
Excited to be releasing cupcake at the end of this week. For deterministic and non-deterministic guardrailing. It integrates via hooks (we created the feature request to anthropic for Claude code).
It looks like the training data has plenty of those examples, but the models don’t have enough grounding or warnings before doing them. I wish there were a PleaseDontDoAnythingStupidEval for software engineering.
Having said that, a PleaseDontDoAnythingStupidEval would probably slow down agentic coding quite a bit and make it less effective given how reliant agents are on making and recovering from mistakes. The solution is probably sandboxes and permission controls to not let them do something overly stupid, no different from an intern.
"Don't talk to it like that, you're doing it wrong!"
Because that is the implication: that a person believing those two things can’t both apply to AI is just too short on understanding and can’t tell the difference between tools for specific use cases or services made for “chat”, and they all have similar foundations.
1. Open mode during learning, where they take everything that comes from the data as 100% truth. The model freely adapts and generalizes with no constraints on consistency.
2. Closed mode during inference, where they take everything that comes from the model as 100% truth. The model doesn't adapt and behaves consistently even if in contradiction with the new information.
I suspect we need to run the model in the mix of the two modes, and possibly some kind of "meta attention" (epistemological) on which parts of the input the model should be "open" (learn from it) and which parts of the input should be "closed" (stick to it).