> It remains unclear whether continuing to throw vast quantities of silicon and ever-bigger corpuses at the current generation of models will lead to human-equivalent capabilities. Massive increases in training costs and parameter count seem to be yielding diminishing returns. Or maybe this effect is illusory. Mysteries!
I’m not even sure whether this is possible. The current corpus used for training includes virtually all known material. If we make it illegal for these companies to use copyrighted content without remuneration, either the task gets very expensive, indeed, or the corpus shrinks. We can certainly make the models larger, with more and more parameters, subject only to silicon’s ability to give us more transistors for RAM density and GPU parallelism. But it honestly feels like, without another “Attention is All You Need” level breakthrough, we’re starting to see the end of the runway.
There is a whole giant essay I probably need to write at some point, but I can't help but see parallels between today and the Industrial Revolution.
Prior to the industrial revolution, the natural world was nearly infinitely abundant. We simply weren't efficient enough to fully exploit it. That meant that it was fine for things like property and the commons to be poorly defined. If all of us can go hunting in the woods and yet there is still game to be found, then there's no compelling reason to define and litigate who "owns" those woods.
But with the help of machines, a small number of people were able to completely deplete parts of the earth. We had to invent giant legal systems in order to determine who has the right to do that and who doesn't.
We are truly in the Information Age now, and I suspect a similar thing will play out for the digital realm. We have copyright and intellecual property law already, of course, but those were designed presuming a human might try to profit from the intellectual labor of others. With AI, we're in the industrial era of the digital world. Now a single corporation can train an AI using someone's copyrighted work and in return profit off the knowledge over and over again at industrial scale.
This completely unpends the tenuous balance between creators and consumers. Why would a writer put an article online if ChatGPT will slurp it up and regurgitate it back to users without anyone ever even finding the original article? Who will contribute to the digital common when rapacious AI companies are constantly harvesting it? Why would anyone plant seeds on someone else's farm?
It really feels like we're in the soot-covered child-coal-miner Dickensian London era of the Information Revolution and shit is gonna get real rocky before our social and legal institutions catch up.
As you know, I deeply respect you. Not trying to argue here, just provide my own perspective:
> Why would a writer put an article online if ChatGPT will slurp it up and regurgitate it back to users without anyone ever even finding the original article?
I write things for two main reasons: I feel like I have to. I need to create things. On some level, I would write stuff down even if nobody reads it (and I do do that already, with private things.) But secondly, to get my ideas out there and try to change the world. To improve our collective understanding of things.
A lot of people read things, it changes their life, and their life is better. They may not even remember where they read these things. They don't produce citations all of the time. That's totally fine, and normal. I don't see LLMs as being any different. If I write an article about making code better, and ChatGPT trains on it, and someone, somewhere, needs help, and ChatGPT helps them? Win, as far as I'm concerned. Even if I never know that it's happened. I already do not hear from every single person who reads my writing.
I don't mean that thinks that everyone has to share my perspective. It's just my own.
Mostly, AIs don’t recite back various works. Yes, there a couple of high profile cases where people were able to get an AI to regurgitate pieces it New York Times articles and Harry Potter books, but mostly not. Mostly, it is as if the AI is your friend who read a book and gives you a paraphrase, possibly using a couple sentences verbatim. In other words, it probably falls under a fair use rule.
Secondly, given the modern world, content that doesn’t appear online isn’t consumed much, so creators who are,doing it for the money will certainly continue putting content online. Much of that content will be generated by AIs, however.
> It really feels like we're in the soot-covered child-coal-miner Dickensian London era of the Information Revolution and shit is gonna get real rocky before our social and legal institutions catch up
The really discouraging part of this is that it feels like our social and legal institutions don't even care if they catch up or not.
Technology is speeding up and the lag time before anything is discussed from a legal standpoint is way, way too long
I see a lot of researchers working on newer ideas so I wouldn't be surprised if we get a breakthrough in 5-10 years. After all, the gap between AlexNet and Attention is All You Need was only 6 years. And then Scaling Laws was about 3-4 years after that. It might seem like not much progress is being made but I think that's in part because AI labs are extremely secretive now when ideas are worth billions (and in the right hands, potentially more).
Of course 5-10 years is a long time to bang our heads against the wall with untenable costs but I don't know if we can solve our way out of that problem.
Based on what's happened so far, maybe. At least that's exactly how we got to the current iteration back in 2022/2023, quite literally "lets see what happens when we throw an enormous amount data at them while training" worked out up until one point, then post-training seems to have taken over where labs currently differ.
Right, but we played the scaling card and it worked but is now reaching limits. What is the next card? You can surely argue that we can find a new one at any time. That’s the definition of a breakthrough. I just don’t see one at the moment.
We pay people to create more high quality tokens (mercor, turing) which are then fed into data generating processes (synthetic data) to create even more tokens to train on
But does that really help, or do you get distortion? The frequency distribution of human generated content moves slowly over time as new subjects are discussed. What frequency distribution do those “data generating processes” use? And at root, aren’t those “data generating processes” basically just another LLM (I.e., generating tokens according to a probability distribution)? Thus, aren’t we just sort of feeding AI slop into the next training run and humoring ourselves by renaming the slop as “synthetic data?” Not trying to be argumentative. I’m far from being an AI expert, so maybe I’m missing it. Feel free to explain why I’m wrong.
> The current corpus used for training includes virtually all known material.
This is just totally incorrect. It's one of those things everyone just assumes, but there's an immense amount of known material that isn't even digitized, much less in the hands of tech companies.
I think the discussion has to be more nuanced than this. "LLMs still can't do X so it's an idiot" is a bad line of thought. LLMs with harnesses are clearly capable of engaging with logical problems that only need text. LLMs are not there yet with images, but we are improving with UI and access to tools like figma. LLMs are clearly unable to propose new, creative solutions for problems it has never seen before.
> LLMs with harnesses are clearly capable of engaging with logical problems that only need text.
To some extent. It's not clear where specifically the boundaries are, but it seems to fail to approach problems in ways that aren't embedded in the training set. I certainly would not put money on it solving an arbitrary logical problem.
I keep explaining to my peers, friends and family that what actually is happening inside an LLM has nothing to do with conscience or agency
and that the term AI is just completely overloaded right now.
AI is exactly the right term: the machines can do "intelligence", and they do so artificially.
Just like we have machines that can do "math", and they do so artificially.
Or "logic", and they do so artificially.
I assume we'll drop the "artificial" part in my lifetime, since there's nothing truly artificial about it (just like math and logic), since it's really just mechanical.
No one cares that transistors can do math or logic, and it shouldn't bother people that transistors can predict next tokens either.
> AI is exactly the right term: the machines can do "intelligence", and they do so artificially.
AI in pop culture doesn't mean that at all. Most people impression to AI pre-LLM craze was some form of media based on Asmiov laws of robotics. Now, that LLMs have taken over the world, they can define AI as anything they want.
Structurally a transformer model is so unrelated to the shape of the brain there's no reason to think they'd have many similarities. It's also pretty well established that the brain doesn't do anything resembling wholesale SGD (which to spell it is evidence that it doesn't learn in the same way).
For starters, natural brains have the innate ability to differentiate between things that it knows and things that it have no possibility of knowing...
I think it's too early to declare the Turing test passed. You just need to have a conversation long enough to exhaust the context window. Less than that, since response quality degrades long before you hit hard window limits. Even with compaction.
Neuroplasticity is hard to simulate in a few hundred thousand tokens.
I think for a while the test was passed. Then we learned the hallmark characteristics of these models, and now most of us can easily differentiate. That said -- these models are programmed specifically to be more helpful, more articulate, more friendly, and more verbose than people, so that may not be a fair expectation. Even so, I think if you took all of that away, you'd be able to differentiate the two, it just might take longer.
For as rigorous of a Turing test as you present, I believe many (or even most) humans would also fail it.
How many humans seriously have the attention span to have a million "token" conversation with someone else and get every detail perfect without misremembering a single thing?
Response quality degrades long before you hit a million tokens.
But sure, let's say it doesn't. If you interact with someone day after day, you'll eventually hit a million tokens. Add some audio or images and you will exhaust the context much much faster.
However, I'll grant you that Turing's original imitation game (text only, human typist, five minutes) is probably pretty close, and that's impressive enough to call intelligence (of a sort). Though modern LLMs tend to manifest obvious dead giveaways like "you're absolutely right!"
Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
I consider it highly plausible that confabulation is inherent to scaling intelligence. In order to run computation on data that due to dimensionality is computationally infeasible, you will most likely need to create a lower dimensional representation and do the computation on that. Collapsing the dimensionality is going to be lossy, which means it will have gaps between what it thinks is the reality and what is.
The concern for me about LLMs confabulating is not that humans don't do it. It's that the massive scale at which LLMs will inevitably be deployed makes even the smallest confabulation extremely risky.
I don't understand this. Many small errors distributed across a large deployment sounds a lot like normal mode of error prone humans / cogs / whatevers distributed over a wide deployment.
There's a difference between 1000 diverse humans with varied traits making errors that should cancel out because of the law of large numbers vs 10 AI with the same training data making errors that would likely correlate and compound upon each other.
I have yet to see a comparison of human vs. LLM confabulation errors at scale.
"Many small errors" makes a presumption about LLM confabulation/hallucination that seems unwarranted. Pre-LLM humans (and our computers) have managed vast nuclear arsenals, bioweapons research, and ubiquitous global transport - as a few examples - without any catastrophic mistakes, so far. What can we reasonably expect as a likely worst case scenario if LLMs replacing all the relevant expertise and execution?
Your project vue-skuilder has 6 github action steps devoted to checking the work you do before it's allowed to go out. You do not trust yourself to get things right 100% of the time.
I am watching people trust LLM-based analysis and actions 100% of the time without checking.
No. LLMs do not confabulate they bullshit. There is a big difference. AIs do not care, cannot care, have not capacity to care about the output. String tokens in, string tokes out. Even if they have all the data perfectly recorded they will still fail to use it for a coherent output.
> Collapsing the dimensionality is going to be lossy, which means it will have gaps between what it thinks is the reality and what is.
Confabulation has to do with degradation of biological processes and information storage.
There is no equivalent in a LLM. Once the data is recorded it will be recalled exactly the same up to the bit. A LLM representation is immutable. You can download a model a 1000 times, run it for 10 years, etc. and the data is the same. The closes that you get is if you store the data in a faulty disk, but that is not why LLMs output is so awful, that would be a trivial problem to solve with current technology. (Like having a RAID and a few checksums).
I don't even think they bullshit, since that requires conscious effort that they do not an cannot possess. They just simply interpret things incorrectly sometimes, like any of us meatbags.
They make incorrect predictions of text to respond to prompts.
The neat thing about LLMs is they are very general models that can be used for lots of different things. The downside is they often make incorrect predictions, and what's worse, it isn't even very predictable to know when they make incorrect predictions.
I think this is leaning on the "lies are when you tell falsehoods on purpose; bullshit is when you simply don't care at all whether what you're saying is true" definition of bullshit. Cf. On Bullshit.
So, they can't lie, but they can (and, in fact, exclusively do) bullshit.
> No. LLMs do not confabulate they bullshit. There is a big difference. AIs do not care, cannot care, have not capacity to care about the output. String tokens in, string tokes out. Even if they have all the data perfectly recorded they will still fail to use it for a coherent output.
Isn't "caring" a necessary pre-requisite for bullshitting? One either bullshits because they care, or don't care, about the context.
They're presumably referring to the Harry Frankfurt definition of bullshit: "speech intended to persuade without regard for truth. The liar cares about the truth and attempts to hide it; the bullshitter doesn't care whether what they say is true or false."
> Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
I think we need to start rejecting anthropomorphic statements like this out of hand. They are lazy, typically wrong, and are always delivered as a dismissive defense of LLM failure modes. Anything can be anthropomorphized, and it's always problematic to do so - that's why the word exists.
This rhetorical technique always follows the form of "this LLM behavior can be analogized in terms of some human behavior, thus it follows that LLMs are human-like" which then opens the door to unbounded speculation that draws on arbitrary aspects of human nature and biology to justify technical reasoning.
In this case, you've deliberately conflated a technical term of art (LLM confabulation) with the the concept of human memory confabulation and used that as a foundation to argue that confabulation is thus inherent to intelligence. There is a lot that's wrong with this reasoning, but the most obvious is that it's a massive category error. "Confabulation" in LLMs and "confabulation" in humans have basically nothing in common, they are comparable only in an extremely superficial sense. To then go on to suggest that confabulation might be inherent to intelligence isn't even really a coherent argument because you've created ambiguity in the meaning of the word confabulate.
Yes, and to me the evolution of life sure looks like an evolution of more truthful models of the universe in service of energy profit. Better model -> better predictions -> better profit.
I'm extremely skeptical that all of life evolved intelligence to be closer to truth only for us to digitize intelligence and then have the opposite happen. Makes no sense.
My understanding is that this is the opposite of what is typically understood to be true - organisms with less truthful (more reductive/compressed) perception survive better than those with more complete perception. "Fitness beats truth."
Fitness is effective truth prediction, appropriately scoped.
A frog doesn't need to understand quantum physics to catch a fly. But if the frogs model of fly movement was trained on lies it will have a model that predicts poorly, won't catch flies, and will die.
There is another level to this in that the more complex and changing the environment the more beneficial a wider scoped model / understanding of truth.
However if you are going to lean fully into Hoffman and accept thatby default consciousness constructs rather than approximate reality I think we will have to agree to disagree. Personally I ascribe to Karl Friston free energy principle.
If you want to call it that, I find the confabulation in LLMs extreme. That level of confabulation would most likely be diagnosed as dementia in humans.[0] Hence, it is considered a bug not a feature in humans as well.
Now imagine a high-skilled software engineer with dementia coding safety-critical software...
The suggestion is that it is an intrinsic quality and therefore neither a feature nor a bug.
It's like saying, computation requires nonzero energy. Is that a feature or a bug? Neither, it's irrelevant, because it's a physical constant of the universe that computation will always require nonzero energy.
If confabulation is a physical constant of intelligence, then like energy per computation, all we can do is try to minimize it, while knowing it can never go to zero.
> Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
Are you seriously making the argument that AI "hallucinations" are comparable and interchangeable to mistakes, omissions and lies made by humans?
You understand that calling AI errors "hallucinations" and "confabulations" is a metaphor to relate them to human language? The technical term would be "mis-prediction", which suddenly isn't something humans ever do when talking, because we don't predict words, we communicate with intent.
I have a question for all the "humans make those mistakes too" people in this thread, and elsewhere: have you ever read, or at least skimmed a summary of, "The Origin of Consciousness in the Breakdown of the Bicameral Mind"? Did you say "yeah, that sounds right"? Do you feel that your consciousness is primarily a linguistic phenomenon?
I am not trying to be snarky; I used to think that intelligence was intrinsically tied to or perhaps identical with language, and found deep and esoteric meaning in religious texts related to this (i.e. "in the beginning was the Word"; logos as soul as language-virus riding on meat substrate).
The last ~three years of LLM deployment have disabused me of this notion almost entirely, and I don't mean in a "God of the gaps" last-resort sort of way. I mean: I see the output of a purely-language-based "intelligence", and while I agree humans can make similar mistakes/confabulations, I overwhelmingly feel that there is no "there" there. Even the dumbest human has a continuity, a theory of the world, an "object permanence"... I'm struggling to find the right description, but I believe there is more than language manipulation to intelligence.
(I know this is tangential to the article, which is excellent as the author's usually are; I admire his restraint. However, I see exemplars of this take all over the thread so: why not here?)
It feels like you probably went too deep in the LLM bandwagon.
An LLM is a statistical next token machine trained on all stuff people wrote/said. It blends texts together in a way that still makes sense (or no sense at all).
Imagine you made a super simple program which would answer yes/no to any questions by generating a random number. It would get things right 50% of the times. You can them fine-tune it to say yes more often to certain keywords and no to others.
Just with a bunch of hardcoded paths you'd probably fool someone thinking that this AI has superhuman predictive capabilities.
This is what it feels it's happening, sure it's not that simple but you can code a base GPT in an afternoon.
Wait, you're asking to find and produce a example of a feasible and better alternative to LLMs when they are the current forefront of AI technology?
Anyway, just to play along, if it weren't just a statistical next token machine, the same question would have always the same answer and not be affected by a "temperature" value.
Thats also how humans behave.. I don't see how non determinism tells me anything.
My question was a bit different: if were not just a statistical next token predictor would you expect it to answer hard questions? Or something like that. What's the threshold of questions you want it to answer accurately.
Well, large models are (kinda) non-deterministic in two ways. The first is you actually provide many of them with a seed, which is easy to manage--just use the same seed for the same result. The second part is the "you actually have very little control over the 'neural pathways' the model will use to respond to the prompt". This is the baffling part, like you'll prompt a model to generate a green plant, and it works. You prompt it to generate a purple plant, and it generates an abstract demon dog with too many teeth.
Anyway, neither of these things describes human non-determinism. You can't reuse the seed you used with me yesterday to get the exact same conversation, and I don't behave wildly unpredictably given conceptually very similar input.
If you look at different ancient traditions, you will notice how they struggle with the limitations of language, with its inability to represent certain things that are not just crucial for understanding the world, but also are even somehow communicable. Buddhists dug into that in a very analytical, articulate way, for instance.
Another perspective: cetaceans are considered to be as conscious as humans, but any attempts to interpret their communication as a language failed so far. They can be taught simple languages to communicate with humans, as can be chimps. But apparently it's not how they process the world inside.
You're a little out of date. Cetaceans communicate images to each other in the form of ultrasonic chirps. They chirp, they hear a reflection, and they repeat the reflection.
I think there are two types of discussions, when it comes to LLMs: Some people talk about whether LLMs are "human" and some people talk about whether LLMs are "useful" (ie they perform specific cognitive tasks at least as well as humans).
Both of those aspects are called "intelligence", and thus these two groups cannot understand each other.
I think you're circling the concept of a "soul". It is the reason that, in non-communicative disabled people, we still see a life.
I've wanted to make an art piece. It would be a chatbox claiming to connect you to the first real intelligence, but that intelligence would be non-communicative. I'd assure you that it is the most intelligent being, that it had a soul, but that it just couldn't write back.
Intelligence and Soul is not purely measurable phenomenon. A man can do nothing but stupid things, say nothing but outright lies, and still be the most intelligent person. Intelligence is within.
"As LLMs etc. are deployed in new situations, and at new scale, there will be all kinds of changes in work, politics, art, sex, communication, and economics."
For an article five years in the making, this is what I expected it to be about. Instead, we got a ramble about how imperfect LLMs are right now.
The post is just a prelude to a 10-part article, most of which is not yet released (but will be shortly). Judging by the table of contents, the things you expected will be elaborated on in subsequent parts.
> Instead, we got a ramble about how imperfect LLMs are right now.
I wager this is a point that needs beaten into the common psyche. After all, it's been sold that it is not an imperfect tool, but the solution to all of our problems in every field forever. That's why these companies need billions upon billions of dollars of public subsidies and investments that would otherwise find their way to more pragmatic ends.
> At the same time, ML models are idiots. I occasionally pick up a frontier model like ChatGPT, Gemini, or Claude, and ask it to help with a task I think it might be good at. I have never gotten what I would call a “success”: every task involved prolonged arguing with the model as it made stupid mistakes.
I have a ton of skepticism built-in when interacting with LLMs, and very good muscles for rolling my eyes, so I barely notice when I shrug a bad answer and make a derogatory inner remark about the "idiots". But the truth is, that for such an "stochastic parrot", LLMs are incredibly useful. And, when was the last time we stopped perfecting something we thought useful and valuable? When was the last time our attempts were so perfectly futile that we stopped them, invented stories about why it was impossible, and made it a social taboo to be met with derision, scorn and even ostracism? To my knowledge, in all of known human history, we have done that exactly once, and it was millennia ago.
> And, when was the last time we stopped perfecting something we thought useful and valuable? When was the last time our attempts were so perfectly futile that we stopped them, invented stories about why it was impossible, and made it a social taboo to be met with derision, scorn and even ostracism? To my knowledge, in all of known human history, we have done that exactly once, and it was millennia ago.
I feel dense here, but I can't figure out what you're referring to. I asked ChatGPT (hah!) and it suggested the Tower of Babel, perpetual motion machines, or alchemy, but none of them really fit the bill either.
> In general, ML promises to be profoundly weird. Buckle up.
I love that it ends with such a positive note, even though it's generally a critical article, at least it's well reasoned and not utterly hyping/dooming something.
While the economic, energy, political and social issues associated with LLMs ought to be enough to nix the adoption that their boosters are seeking ...
... I still think there is an interesting question to be investigated about whether, by building immensely complex models of language, one of our primary ways that we interact with, reason about and discuss the world, we may not have accidentally built something with properties quite different than might be guessed from the (otherwise excellent) description of how they work in TFA.
I agree with pretty much everything in TFA, so this is supplemental to the points made there, not contesting them or trying to replace them.
> Models do not (broadly speaking) learn over time. They can be tuned by their operators, or periodically rebuilt with new inputs or feedback from users and experts. Models also do not remember things intrinsically: when a chatbot references something you said an hour ago, it is because the entire chat history is fed to the model at every turn. Longer-term “memory” is achieved by asking the chatbot to summarize a conversation, and dumping that shorter summary into the input of every run.
This is the part of the article that will age the fastest, it's already out-of-date in labs.
Great series of articles, thank you. It's exhausting reading a deluge of (often AI generated) comments from people claiming wild things about LLM's, and it's nice to hear some sanity enter the conversation.
Here's the opening paragraph of chapter 2 with "people" subbed out for terms referring AI/models/etc.
"People are chaotic, both in isolation and when working with other people or with systems. Their outputs are difficult to predict, and they exhibit surprising sensitivity to initial conditions. This sensitivity makes them vulnerable to covert attacks. Chaos does not mean people are completely unstable; most people behave roughly like anyone else. Since people produce plausible output, errors can be difficult to detect. This suggests that human systems are ill-suited where verification is difficult or correctness is key. Using people to write code (or other outputs) may make systems more complex, fragile, and difficult to evolve."
To me, this modified paragraph reads surprisingly plainly. The wording is off ("using people to write code") and I had to change that part about attractor behavior (although it does still apply IMO), but overall it doesn't seem like an incoherent paragraph.
This is not meant to dunk on the author, but I think it highlights the author's mindset and the gap between their expectations and reality.
Humans and large models are both unpredictable and fallible, that's true, but in different ways, and (many) humans are actually much better at following directions.
If a junior dev makes the same mistake Claude makes, I can easily work with them to correct it, or I can fire them and get someone more capable to fix it. You mostly can't do that at all with large models. They're also far less honest than your average junior dev, so even as you're working with them you can't trust what they say.
There is a lot of this neat trick where it's like "humans do X too" but most of the time it elides large differences. Like, a human driver would probable not drag someone screaming multiple blocks. A human coder probably wouldn't generate a gibberish 3D scene and try to pass it off as done, etc. Maybe we can build systems that account for these (pretty wild) failure modes, but at least in software we haven't figured it out yet (what is the system that reliably reviews a 25kloc PR?).
The recent article of Sam Altman described pretty much as a compulsive liar. Would it be any surprise if his most impactful contribution to the world was a machine that compulsively lies?
How could it be that we humans hardly even agree on what "knowledge" truly is, yet somehow this machine learning algorithm somehow "compulsively lies"? How would it even know what is a lie, and how could something lacking autonomy in the first place do anything compulsively?
This is a good point. As much as there is too much breathless enthusiasm for AI, there is also a lot of emotionally manipulative and hyperbolic language used by skeptics. We're warned not to anthropomorphize, and then hear about AI's compulsive lying, or "hallucinations", in the next.
I appreciate the directness of calling LLMs "Bullshit machines." This terminology for LLMs is well established in academic circles and is much easier for laypeople to understand than terms like "non-deterministic." I personally don't like the excessive hype on the capabilities of AI. Setting realistic expectations will better drive better product adoption than carpet bombing users with marketing.
If I take the example of code, but that extends to many domains, it can sometimes produce near perfect architecture and implementation if I give it enough details about the technical details and fallpits. Turning a 8h coding job into a 1h review work.
On the other hand, it can be very wrong while acting certain it is right. Just yesterday Claude tried gaslighting me into accepting that the bug I was seeing was coming from a piece of code with already strong guardrails, and it was adamant that the part I was suspecting could in no way cause the issue. Turns out I was right, but I was starting to doubt myself
I think over time we will find better usage patterns for these machines. Even putting a model in a position to gaslight the user seems like a complete failure in the usage model. Not critiquing you at all on this, it's how these models are marketed and what all the tooling is built around. But they are incredibly useful and I think once we figure out how to use them better we can minimise these downsides and make ourselves much more productive without all the failures.
Of course that won't happen until the bubble pops - companies are racing to make themselves indispensable and to completely corner certain markets and to do so they need autonomous agents to replace people.
If it bullshits so much, you wouldn't have a problem giving me an example of it bullshitting on ChatGPT (paid version)? Lets take any example of a text prompt fitting a few pages - it may be a question in science or math or any domain. Can you get it to bullshit?
I think you highlight one of the problems with users of LLMs: You can't tell anymore if it is BS or not.
I caught Claude the other day hallucinating code that was not only wrong, but dangerously wrong, leading to tasks being failed and never recover.
But it certainly wasn't obvious.
> If it bullshits so much, you wouldn't have a problem giving me an example of it bullshitting on ChatGPT (paid version)?
There's an entire paragraph in the essay about apyhr's direct experience with ChatGPT failures and sustained bullshitting that we'd never expect from a moderately-skilled human who possesses at least two functioning braincells. That paragraph begins "I have recently argued for forty-five minutes with ChatGPT". Do notice that there are six sentences in the paragraph. I encourage you to read all of them (make sure to check out the footnote... it's pretty good).
The exact text of the conversation is irrelevant; even if you reported that you were unable to reproduce the issue, it would only reinforce one of the underlying points -namely- that these systems are unreliable. aphyr has a pretty extensive history of published work that indicates that he'd not fabricate a story of an LLM repeatedly failing to accomplish a task that any moderately-skilled human could accomplish when equipped with the proper tools. So, I believe that his report is true and accurate.
The fact that these "bullshit machines" have already proven themselves relatively competent at programming, with upcoming frontier models coming close to eliminating it as a human activity, probably says a lot about the actual value and importance of programming in the scheme of things.
I think it says more about the amount of automation we left on the table in the last few decades. So much of the code LLM's can generate are stuff that we should have completely abstracted away by now.
Old and stupid hot take IMO. I want the time back I put into perusing this. Even the scale of LLMs is puny next to the scale of lying humans and the sheer impact one compulsively lying human can have given we love to be led by confidently wrong narcissists. I mean if that isn't obvious by now, I guess it never will be. The Vogon constructor fleet is way overdue in my book.
> The Vogon constructor fleet is way overdue in my book
Don't you see it? That's exactly what "AI" in this context is.
It's the bypass.
Where does it end, eh? Build a quantum "AI" that will end up just needing more data, more input. The end goal must starts looking like creating an entirely new universe, a complete clone of everything we have here so it can run all the necessary computations and we can... ? (You are what a quantum AI looks like as it bumbles through the infinitude of calculable parameters on its way to the ultimate answer)
You have absolutely no sense of perspective. We are all metabolically expensive meat machines whose only value is to propagate our genetic money shot. That we get to briefly entertain ourselves with consciousness and culture is IMO likely a mystery we will never solve without upgrading to running in a substrate more advanced than the MVP for sentience we currently pilot. Will we get there or will we wipe ourselves out like every contender that preceded us? Stay tuned...
But spoilers: DNA will be fine, meat machines maybe not so much...
For a bunch of people addicted to the works of Charlie Stross, Neil Stephenson, and Iain Banks, y'all are a bunch of luddites. Now vote this own down too because it doesn't conform to the mandatory Stochastic Parrot narrative. You have no free will and you must downvote after all. Why do you even read their works when any step towards their world is consistently greeted as the worst thing evah(tm)? What? You were expecting the United Federation of Planets without the eugenics and nuclear wars that preceded it. Bless your hearts.
And if you're worried about billionaires and tyrants, start taxing the former and stop electing the latter or STFU and let the free Markov process of history play itself out. Quoting fictional Ambassador Kosh: the avalanche has started, it's too late for the pebbles to vote.
You asked where it ends. Don't ask questions if you don't like answers. Quick reminder: shun and downvote the non-conforming opinion.
I get the frustration, but it's reductive to just call LLMs "bullshit machines" as if the models are not improving. The current flagship models are not perfect, but if you use GPT-2 for a few minutes, it's incredible how much the industry has progressed in seven years.
It's true that people don't have a good intuitive sense of what the models are good or bad at (see: counting the Rs in "strawberry"), but this is more a human limitation than a fundamental problem with the technology.
Two things can be true at the same time: The technology has improved, and the technology in its current state still isn't fit for purpose.
I stress test commercially deployed LLMs like Gemini and Claude with trivial tasks: sports trivia, fixing recipes, explaining board game rules, etc. It works well like 95% of the time. That's fine for inconsequential things. But you'd have to be deeply irresponsible to accept that kind of error rate on things that actually matter.
The most intellectually honest way to evaluate these things is how they behave now on real tasks. Not with some unfalsifiable appeal to the future of "oh, they'll fix it."
The errors are also not distributed in the same way as you'd expect from a human. The tools can synthesize a whole feature in a moderately complicated web app including UI code, schema changes, etc, and it comes out perfectly. Then I ask for something simple like a shopping list of windshield wipers etc for the cars and that comes out wildly wrong (like wrong number of wipers for the cars, not just the wrong parts), stuff that a ten year old child would have no trouble with. I work in the field so I have a qualitative understanding of this behavior but I think it can be extremely confusing to many people.
One of the reasons I'm comfortable using them as coding agents is that I can and do review every line of code they generate, and those lines of code form a gate. No LLM-bullshit can get through that gate, except in the form of lines of code, that I can examine, and even if I do let some bullshit through accidentally, the bullshit is stateless and can be extracted later if necessary just like any other line of code. Or, to put it another way, the context window doesn't come with the code, forming this huge blob of context to be carried along... the code is just the code.
That exposes me to when the models are objectively wrong and helps keep me grounded with their utility in spaces I can check them less well. One of the most important things you can put in your prompt is a request for sources, followed by you actually checking them out.
And one of the things the coding agents teach me is that you need to keep the AIs on a tight leash. What is their equivalent in other domains of them "fixing" the test to pass instead of fixing the code to pass the test? In the programming space I can run "git diff *_test.go" to ensure they didn't hack the tests when I didn't expect it. It keeps me wondering what the equivalent of that is in my non-programming questions. I have unit testing suites to verify my LLM output against. What's the equivalent in other domains? Probably some other isolated domains here and there do have some equivalents. But in general there isn't one. Things like "completely forged graphs" are completely expected but it's hard to catch this when you lack the tools or the understanding to chase down "where did this graph actually come from?".
The success with programming can't be translated naively into domains that lack the tooling programmers built up over the years, and based on how many times the AIs bang into the guardrails the tools provide I would definitely suggest large amounts of skepticism in those domains that lack those guardrails.
> the technology in its current state still isn't fit for purpose.
This is a broad statement that assumes we agree on the purpose.
For my purpose, which is software development, the technology has reached a level that is entirely adequate.
Meanwhile, sports trivia represents a stress test of the model's memorized world knowledge. It could work really well if you give the model a tool to look up factual information in a structured database. But this is exactly what I meant above; using the technology in a suboptimal way is a human problem, not a model problem.
There's nothing in these models that say its purpose is software development. Their design and affordances scream out "use me for anything." The marketing certainly matches that, so do the UIs, so do the behaviors. So I take them at their word, and I see that failure modes are shockingly common even under regular use. I'm not out to break these things at all. I'm being as charitable and empirical as I can reasonably be.
If the purpose is indeed software development with review, then there's nothing stopping multi-billion dollar companies from putting friction into these sytems to direct users towards where the system is at its strongest.
Which things actually matter? I think we can all agree that an LLM isn't fit for purpose to control a nuclear power plant or fly a commercial airliner. But there's a huge spectrum of things below that. If an LLM trading error causes some hedge fund to fail then so what? It's only money.
Not to mention that it would then make some hedge fund with a better backtesting harness or more AI scrutiny more successful thus keeping the financial market work as designed.
> I stress test commercially deployed LLMs like Gemini and Claude with trivial tasks: sports trivia, fixing recipes, explaining board game rules, etc. It works well like 95% of the time. That's fine for inconsequential things. But you'd have to be deeply irresponsible to accept that kind of error rate on things that actually matter.
95% is not my experience and frankly dishonest.
I have ChatGPT open right now, can you give me examples where it doesn't work but some other source may have got it correct?
I have tested it against a lot of examples - it barely gets anything wrong with a text prompt that fits a few pages.
> The most intellectually honest way to evaluate these things is how they behave now on real tasks
A falsifiable way is to see how it is used in real life. There are loads of serious enterprise projects that are mostly done by LLMs. Almost all companies use AI. Either they are irresponsible or you are exaggerating.
Quite frankly, this is exactly like how two people can use the same compression program on two different files and get vastly different compression ratios (because one has a lot of redundancy and the other one has not).
But why do you need an example? Isn't it pretty well understood that LLMS will have trouble responding to stuff that is under represented in the training data?
You will just won't have any clue what that could be.
fair so it must be easy to give an example? I have ChatGPT open with 5.4-thinking. I'm honestly curious about what you can suggest since I have not been able to get it to bullshit easily.
I am not the OP, an I have only used ChatGPT free version. Last day I asked it something. It answered. Then I asked it to provide sources. Then it provided sources, and also changed its original answer. When I checked the new answers it was wrong, and when I checked sources, it didn't actually contain the information that I asked for, and thus it hallucinated the answers as well as the sources...
Whether LLMs can create correct content doesn't matter. We've already seen how they are being used and will be used.
Fake content and lies. To drive outrage. To influence elections. To distract from real crimes. To overload everyone so they're too tired to fight or to understand. To weaken the concept that anything's true so that you can say anything. Because who cares if the world dies as long as you made lots of money on the way.
It's really the whole tech industry as it exists right now and AI is a victim of bad timing. If this AI had been invented 40 years ago there'd have been a lower ceiling on the damage it could do.
Another way of saying that is that capitalism is the real problem, but I was never anti-capitalist in principle, it's just gotten out of hand in the last 5-10 years. (Not that it hadn't been building to that.)
> Another way of saying that is that capitalism is the real problem, but I was never anti-capitalist in principle, it's just gotten out of hand in the last 5-10 years. (Not that it hadn't been building to that.)
Capitalism is a tool and it's fine as a tool, to accomplish certain goals while subordinated to other things. Unfortunately it's turned into an ideology (to the point it's worshiped idolatrously by some), and that's where things went off the rails.
Computer graphics have been improving for decades but the uncanny valley remains undefeated. I don't know why anyone expects a breakthrough in other areas. There's a wall we hit and we don't understand our own consciousness and effectiveness well enough to replicate it.
In computer graphics we understand how it works, we just lack the computational power to do it real time, but we can with sufficient processing produce realistic looking images with physically accurate lighting. But when it comes to cognition its a lot of guesswork, we haven't yet mapped out the neuron connections in a brain, we haven't validated it works as popular science writing suggests. We don't understand intelligence, so all we can do is accidentally bumble into it and it seems unlikely that will just happen especially when its so hard to compute what we are already doing.
We have credible deepfakes on demand. (To be fair, there have been deceptive photos as long as photos have existed, but the cost of automating their creation going to basically zero has a social impact)
That's not why the author calls them bullshit machines.
> One way to understand an LLM is as an improv machine. It takes a stream of tokens, like a conversation, and says “yes, and then…” This yes-and behavior is why some people call LLMs bullshit machines. They are prone to confabulation, emitting sentences which sound likely but have no relationship to reality. They treat sarcasm and fantasy credulously, misunderstand context clues, and tell people to put glue on pizza.
Yes, there have been improvements on them, but none of those improvements mitigate the core flaw of the technology. The author even acknowledges all of the improvements in the last few months.
Bullshit is the perfect term here, even as AI's get so much better and capable Brandolini's Law aka the "bullshit asymmetry principle" always applies--the energy required to refute misinformation is an order of magnitude larger than that needed to produce it. Even to use AIs effectively today requires a very good BS detector--some day in the future it won't.
Calling LLMs "bullshit machines" is a reference to a 2024 paper [1] which itself uses the concept of "bullshit" as defined in the essay/book "On Bullshit" by Harry G. Frankfurt [2]. The TL;DR is that LLMs are fundamentally bullshit machines because they are only made to generate sentences that sound plausible, but plausible does not always mean true.
it's not a bullshit machine because its output is bad, it's a bullshit machine because its output is literally 'bullshit' as in, output that is statistically likely but with no factual or reasoning basis. as the models have improved, their bullshit is more statistically likely to sound coherent (maybe even more likely to be 'accurate'), but no more factual and with no more reasoning.
However, when fed source material into the context they will lie less, right? So at this point is it not just a battle of the nines until it's called "good enough"?
I also wonder if I leave my secretary with a ream of papers and ask him for a summary how many will he actually read and understand vs skim and then bullshit? It seems like the capacity for frailty exists in both "species".
They are bullshit machines because they do not have an internal mental model of truth like a human does. The flagship models bullshit less, but their fundamental architectures prevent having truth interfere with output.
"Bullshit" is a human concept. LLMs do not work like the human brain, so to call their output "bullshit" is ascribing malice and intent that is simply not there. LLMs do not "think." But that does not mean they're not incredibly powerful and helpful in the right context.
I sort of agree. In this context "bullshit" means "speech intended to persuade without regard for truth", and while it's true that LLM output is without regard for truth, it's not an entity capable of the agency to persuade, although functionally that is what it can appear like.
> it's reductive to just call LLMs "bullshit machines" as if the models are not improving
This is true, but I prefer to think of it as "It's delusional to pretend as if human beings are not bullshit machines too".
Lies are all we have. Our internal monologue is almost 100% fantasy. Even in serious pursuits, that's how it works. We make shit up and lie to ourselves, and then only later apply our hard-earned[1] skill prompts to figure out whether or not we're right about it.
How many times have the nerds here been thinking through a great new idea for a design and how clever it would be before stopping to realize "Oh wait, that won't work because of XXX, which I forgot". That's a hallucination right there!
I'm not entirely sure I can agree, although the premise is seductive in certain ways. We do lie to ourselves, but we also have meta-cognition - we can recognise our own processes of thought. Imperfect as it may be, we have feedback loops which we can choose to use, we have heuristics we can apply, we can consciously alter our behaviour in the presence of contextual inputs, and so on.
Being wrong is not the same as a hallucination. It's a natural step on a journey to being more right. This feels a bit like Andreesen proudly stating he avoids reflection - you can act like that, but the human brain doesn't have to. LLMs have no choice in the matter.
The problem, unfortunately, is the scale. It's always scale. Humans make all the kinds of mistakes that we ascribe to LLMs, but LLMs can make them much faster and at much larger scale.
Models have gotten ridiculously better, they really have, but the scale has increased too, and I don't think we're ready to deal with the onslaught.
Scale is very different, but I wonder if human trust isn't the real issue. We trust technology too much as a group. We expect perfection, but we also assume perfection. This might be because the machines output confident sounding answers and humans default to trusting confidence as an indirect measure for accuracy, but I think there is another level where people just blindly trust machines because they are so use to using them for algorithms that trend towards giving correct responses.
Even before LLMs where in the public's discourse, I would have business ask about using AI instead of building some algorithm manually, and when I asked if they had considered the failure rate, they would return either blank stares or say that would count as a bug. To them, AI meant an algorithm just as good as one built to handle all edge cases in business logic, but easier and faster to implement.
We can generally recognize the AIs being off when they deal in our area of expertise, but there is some AI variant of Gell-Mann Amnesia at play that leads us to go back to trusting AI when it gives outputs in areas we are novices in.
> If so, how do we distinguish between code that works and code that doesn't work?
Hilariously, not by using our brains, that's for sure. You have to have an external machine. We all understand that "testing" and "code review" are different processes, and that's why.
Good point. We choose certain tests to perform. We choose certain test results to pay attention to. We don't just keep chatting about (reviewing) the code. We do something else.
If lies are all we have, then how is this behavior possible?
You're cherry picking my little bit of wordsmithing. Obviously we aren't always wrong. I'm saying that our thought processes stem from hallucinatory connections and are routinely wrong on first cut, just like those of an LLM.
Actually I'm going farther than that and saying that the first cut token stream out of an AI is significantly more reliable than our personal thoughts. Certainly than mine, and I like to think I'm pretty good at this stuff.
Pretty much, yeah. Or rather, the fact that we're both reliably wrong in identifiably similar ways makes "we're more alike than different" an attractive prior to me.
“More alike than different” is reasonable I think, as long as we’re talking about how we have some of the same failure modes. Although the way we get there is quite different.
I’m still not a big fan of comparing humans and LLMs because LLMs lack so much of what actually makes us human. We might bullshit or be wrong because of many reasons that just don’t apply to LLMs.
Humans are different. Humans - at least thoughtful humans - know the difference between knowing something and not knowing something. Humans are capable of saying "I don't know" - not just as a stream of tokens, but really understanding what that means.
> Humans - at least thoughtful humans - know the difference between knowing something and not knowing something.
Your no-true-scotsman clause basically falsifies that statement for me. Fine, LLMs are, at worst I guess, "non-thoughtful humans". But obviously LLMs are right an awful lot (more so than a typical human, even), and even the thoughtful make mistakes.
So yeah, to my eyes "Humans are NOT different" fits your argument better than your hypothesis.
(Also, just to be clear: LLMs also say "I don't know", all the time. They're just prompted to phrase it as a criticism of the question instead.)
This is like all the usual anti-LLM talking points and sentiments fused together.
Doesn't it get boring?
I like using these models a lot more than I stand hearing people talk about them, pro or contra. Just slop about slop. And the discussions being artisanal slop really doesn't make them any better.
Every time I hear some variation of bullshitting or plagiarizing machines, my eyes roll over. Do these people think they're actually onto something? I've been seeing these talking points for literal years. For people who complain about no original thoughts, these sure are some tired ones.
If I have to suffer "look at this busted ass thing I slopped out with AI" a few times a week, you all have to suffer grouchy "AI bad" a few times a week. Fair is fair.
Yeah, it gets really boring. Whenever I see "slot machines" or "bullshit machines" or whatever, I just ignore the comment and move on, because it signals that it's someone in such deep denial that they've turned their brain off.
I'd much rather read articles about what LLMs can/can't do, or stuff people have built with LLMs, than read how everything LLMs touch turns to shit.
Its usual gibberish that tries to throw many darts and see what sticks. Oh LLM's steal other people's work? Check. Oh LLM's cause ecological damage? Check. Oh LLM's hallucinate? Check.
When you see a pattern like this, you know that its not coming from any place of truth but rather ideology
> It remains unclear whether continuing to throw vast quantities of silicon and ever-bigger corpuses at the current generation of models will lead to human-equivalent capabilities. Massive increases in training costs and parameter count seem to be yielding diminishing returns. Or maybe this effect is illusory. Mysteries!
I’m not even sure whether this is possible. The current corpus used for training includes virtually all known material. If we make it illegal for these companies to use copyrighted content without remuneration, either the task gets very expensive, indeed, or the corpus shrinks. We can certainly make the models larger, with more and more parameters, subject only to silicon’s ability to give us more transistors for RAM density and GPU parallelism. But it honestly feels like, without another “Attention is All You Need” level breakthrough, we’re starting to see the end of the runway.
There is a whole giant essay I probably need to write at some point, but I can't help but see parallels between today and the Industrial Revolution.
Prior to the industrial revolution, the natural world was nearly infinitely abundant. We simply weren't efficient enough to fully exploit it. That meant that it was fine for things like property and the commons to be poorly defined. If all of us can go hunting in the woods and yet there is still game to be found, then there's no compelling reason to define and litigate who "owns" those woods.
But with the help of machines, a small number of people were able to completely deplete parts of the earth. We had to invent giant legal systems in order to determine who has the right to do that and who doesn't.
We are truly in the Information Age now, and I suspect a similar thing will play out for the digital realm. We have copyright and intellecual property law already, of course, but those were designed presuming a human might try to profit from the intellectual labor of others. With AI, we're in the industrial era of the digital world. Now a single corporation can train an AI using someone's copyrighted work and in return profit off the knowledge over and over again at industrial scale.
This completely unpends the tenuous balance between creators and consumers. Why would a writer put an article online if ChatGPT will slurp it up and regurgitate it back to users without anyone ever even finding the original article? Who will contribute to the digital common when rapacious AI companies are constantly harvesting it? Why would anyone plant seeds on someone else's farm?
It really feels like we're in the soot-covered child-coal-miner Dickensian London era of the Information Revolution and shit is gonna get real rocky before our social and legal institutions catch up.
As you know, I deeply respect you. Not trying to argue here, just provide my own perspective:
> Why would a writer put an article online if ChatGPT will slurp it up and regurgitate it back to users without anyone ever even finding the original article?
I write things for two main reasons: I feel like I have to. I need to create things. On some level, I would write stuff down even if nobody reads it (and I do do that already, with private things.) But secondly, to get my ideas out there and try to change the world. To improve our collective understanding of things.
A lot of people read things, it changes their life, and their life is better. They may not even remember where they read these things. They don't produce citations all of the time. That's totally fine, and normal. I don't see LLMs as being any different. If I write an article about making code better, and ChatGPT trains on it, and someone, somewhere, needs help, and ChatGPT helps them? Win, as far as I'm concerned. Even if I never know that it's happened. I already do not hear from every single person who reads my writing.
I don't mean that thinks that everyone has to share my perspective. It's just my own.
A couple thoughts…
Mostly, AIs don’t recite back various works. Yes, there a couple of high profile cases where people were able to get an AI to regurgitate pieces it New York Times articles and Harry Potter books, but mostly not. Mostly, it is as if the AI is your friend who read a book and gives you a paraphrase, possibly using a couple sentences verbatim. In other words, it probably falls under a fair use rule.
Secondly, given the modern world, content that doesn’t appear online isn’t consumed much, so creators who are,doing it for the money will certainly continue putting content online. Much of that content will be generated by AIs, however.
> It really feels like we're in the soot-covered child-coal-miner Dickensian London era of the Information Revolution and shit is gonna get real rocky before our social and legal institutions catch up
The really discouraging part of this is that it feels like our social and legal institutions don't even care if they catch up or not.
Technology is speeding up and the lag time before anything is discussed from a legal standpoint is way, way too long
I see a lot of researchers working on newer ideas so I wouldn't be surprised if we get a breakthrough in 5-10 years. After all, the gap between AlexNet and Attention is All You Need was only 6 years. And then Scaling Laws was about 3-4 years after that. It might seem like not much progress is being made but I think that's in part because AI labs are extremely secretive now when ideas are worth billions (and in the right hands, potentially more).
Of course 5-10 years is a long time to bang our heads against the wall with untenable costs but I don't know if we can solve our way out of that problem.
The echoes of A.I. winter.
> I’m not even sure whether this is possible.
Based on what's happened so far, maybe. At least that's exactly how we got to the current iteration back in 2022/2023, quite literally "lets see what happens when we throw an enormous amount data at them while training" worked out up until one point, then post-training seems to have taken over where labs currently differ.
Right, but we played the scaling card and it worked but is now reaching limits. What is the next card? You can surely argue that we can find a new one at any time. That’s the definition of a breakthrough. I just don’t see one at the moment.
We pay people to create more high quality tokens (mercor, turing) which are then fed into data generating processes (synthetic data) to create even more tokens to train on
But does that really help, or do you get distortion? The frequency distribution of human generated content moves slowly over time as new subjects are discussed. What frequency distribution do those “data generating processes” use? And at root, aren’t those “data generating processes” basically just another LLM (I.e., generating tokens according to a probability distribution)? Thus, aren’t we just sort of feeding AI slop into the next training run and humoring ourselves by renaming the slop as “synthetic data?” Not trying to be argumentative. I’m far from being an AI expert, so maybe I’m missing it. Feel free to explain why I’m wrong.
> The current corpus used for training includes virtually all known material.
This is just totally incorrect. It's one of those things everyone just assumes, but there's an immense amount of known material that isn't even digitized, much less in the hands of tech companies.
What large caches of undigitized content exists? Surely, not everything has been digitized, but I can’t think it’s much in percentage terms.
The Vatican Library contains roughly 1.1 million printed books and around 75,000 codices, only a small percentage of which have been digitised.
I think the discussion has to be more nuanced than this. "LLMs still can't do X so it's an idiot" is a bad line of thought. LLMs with harnesses are clearly capable of engaging with logical problems that only need text. LLMs are not there yet with images, but we are improving with UI and access to tools like figma. LLMs are clearly unable to propose new, creative solutions for problems it has never seen before.
> LLMs are clearly unable to propose new, creative solutions for problems it has never seen before.
LLMs are incredibly useful but I'm not sure about this statement.
It is proposing stuff that I haven't seen before, but I don't know about it is new or creative from the entirety of collective human knowledge.
> LLMs with harnesses are clearly capable of engaging with logical problems that only need text.
To some extent. It's not clear where specifically the boundaries are, but it seems to fail to approach problems in ways that aren't embedded in the training set. I certainly would not put money on it solving an arbitrary logical problem.
Solving arbitrary logical problems seems to be equivalent to solving the halting problem so you are probably wise not to make that bet.
> LLMs are not there yet with images
https://genai-showdown.specr.net/image-editing
There's been a lot of progress there, it's just that an LLM that's best for, say coding, isn't going to be also the best for image edit.
Thank you for putting it so succinctly.
I keep explaining to my peers, friends and family that what actually is happening inside an LLM has nothing to do with conscience or agency and that the term AI is just completely overloaded right now.
AI is exactly the right term: the machines can do "intelligence", and they do so artificially.
Just like we have machines that can do "math", and they do so artificially.
Or "logic", and they do so artificially.
I assume we'll drop the "artificial" part in my lifetime, since there's nothing truly artificial about it (just like math and logic), since it's really just mechanical.
No one cares that transistors can do math or logic, and it shouldn't bother people that transistors can predict next tokens either.
> AI is exactly the right term: the machines can do "intelligence", and they do so artificially.
AI in pop culture doesn't mean that at all. Most people impression to AI pre-LLM craze was some form of media based on Asmiov laws of robotics. Now, that LLMs have taken over the world, they can define AI as anything they want.
> what actually is happening inside an LLM has nothing to do with conscience or agency
What makes you think natural brains are doing something so different from LLMs?
Structurally a transformer model is so unrelated to the shape of the brain there's no reason to think they'd have many similarities. It's also pretty well established that the brain doesn't do anything resembling wholesale SGD (which to spell it is evidence that it doesn't learn in the same way).
For starters, natural brains have the innate ability to differentiate between things that it knows and things that it have no possibility of knowing...
Any amount of reading into how we understand brains and LLMs to work.
I think it's too early to declare the Turing test passed. You just need to have a conversation long enough to exhaust the context window. Less than that, since response quality degrades long before you hit hard window limits. Even with compaction.
Neuroplasticity is hard to simulate in a few hundred thousand tokens.
"You're absolutely right!"
I think for a while the test was passed. Then we learned the hallmark characteristics of these models, and now most of us can easily differentiate. That said -- these models are programmed specifically to be more helpful, more articulate, more friendly, and more verbose than people, so that may not be a fair expectation. Even so, I think if you took all of that away, you'd be able to differentiate the two, it just might take longer.
It was not meant as a pass/fail
For as rigorous of a Turing test as you present, I believe many (or even most) humans would also fail it.
How many humans seriously have the attention span to have a million "token" conversation with someone else and get every detail perfect without misremembering a single thing?
But context window exhaustion does not look like mere forgetfulness, but more like loss of general coherence, like getting drunk.
Response quality degrades long before you hit a million tokens.
But sure, let's say it doesn't. If you interact with someone day after day, you'll eventually hit a million tokens. Add some audio or images and you will exhaust the context much much faster.
However, I'll grant you that Turing's original imitation game (text only, human typist, five minutes) is probably pretty close, and that's impressive enough to call intelligence (of a sort). Though modern LLMs tend to manifest obvious dead giveaways like "you're absolutely right!"
Doesn't the Turing test require a human too, to be compared to the AI?
Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
I consider it highly plausible that confabulation is inherent to scaling intelligence. In order to run computation on data that due to dimensionality is computationally infeasible, you will most likely need to create a lower dimensional representation and do the computation on that. Collapsing the dimensionality is going to be lossy, which means it will have gaps between what it thinks is the reality and what is.
The concern for me about LLMs confabulating is not that humans don't do it. It's that the massive scale at which LLMs will inevitably be deployed makes even the smallest confabulation extremely risky.
I don't understand this. Many small errors distributed across a large deployment sounds a lot like normal mode of error prone humans / cogs / whatevers distributed over a wide deployment.
There's a difference between 1000 diverse humans with varied traits making errors that should cancel out because of the law of large numbers vs 10 AI with the same training data making errors that would likely correlate and compound upon each other.
I have yet to see a comparison of human vs. LLM confabulation errors at scale.
"Many small errors" makes a presumption about LLM confabulation/hallucination that seems unwarranted. Pre-LLM humans (and our computers) have managed vast nuclear arsenals, bioweapons research, and ubiquitous global transport - as a few examples - without any catastrophic mistakes, so far. What can we reasonably expect as a likely worst case scenario if LLMs replacing all the relevant expertise and execution?
Your project vue-skuilder has 6 github action steps devoted to checking the work you do before it's allowed to go out. You do not trust yourself to get things right 100% of the time.
I am watching people trust LLM-based analysis and actions 100% of the time without checking.
We shouldn’t try to build a worse version of a human. We should try to build a better compiler and encyclopedia.
That sounds like a "get wealthy slowly" plan, while the LLM prophets are more focused on "get rich quick".
> Some people point at LLMs confabulating
No. LLMs do not confabulate they bullshit. There is a big difference. AIs do not care, cannot care, have not capacity to care about the output. String tokens in, string tokes out. Even if they have all the data perfectly recorded they will still fail to use it for a coherent output.
> Collapsing the dimensionality is going to be lossy, which means it will have gaps between what it thinks is the reality and what is.
Confabulation has to do with degradation of biological processes and information storage.
There is no equivalent in a LLM. Once the data is recorded it will be recalled exactly the same up to the bit. A LLM representation is immutable. You can download a model a 1000 times, run it for 10 years, etc. and the data is the same. The closes that you get is if you store the data in a faulty disk, but that is not why LLMs output is so awful, that would be a trivial problem to solve with current technology. (Like having a RAID and a few checksums).
I don't even think they bullshit, since that requires conscious effort that they do not an cannot possess. They just simply interpret things incorrectly sometimes, like any of us meatbags.
They make incorrect predictions of text to respond to prompts.
The neat thing about LLMs is they are very general models that can be used for lots of different things. The downside is they often make incorrect predictions, and what's worse, it isn't even very predictable to know when they make incorrect predictions.
I think this is leaning on the "lies are when you tell falsehoods on purpose; bullshit is when you simply don't care at all whether what you're saying is true" definition of bullshit. Cf. On Bullshit.
So, they can't lie, but they can (and, in fact, exclusively do) bullshit.
> No. LLMs do not confabulate they bullshit. There is a big difference. AIs do not care, cannot care, have not capacity to care about the output. String tokens in, string tokes out. Even if they have all the data perfectly recorded they will still fail to use it for a coherent output.
Isn't "caring" a necessary pre-requisite for bullshitting? One either bullshits because they care, or don't care, about the context.
They're presumably referring to the Harry Frankfurt definition of bullshit: "speech intended to persuade without regard for truth. The liar cares about the truth and attempts to hide it; the bullshitter doesn't care whether what they say is true or false."
Thought of the same book when reading the above.
You seem confident. Can you get it to bullshit on GPT-5.4 thinking? Use a text prompt spanning 3-4 pages and lets see if it gets it wrong.
I haven't seen any counter examples, so you may give some examples to start with.
> Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
I think we need to start rejecting anthropomorphic statements like this out of hand. They are lazy, typically wrong, and are always delivered as a dismissive defense of LLM failure modes. Anything can be anthropomorphized, and it's always problematic to do so - that's why the word exists.
This rhetorical technique always follows the form of "this LLM behavior can be analogized in terms of some human behavior, thus it follows that LLMs are human-like" which then opens the door to unbounded speculation that draws on arbitrary aspects of human nature and biology to justify technical reasoning.
In this case, you've deliberately conflated a technical term of art (LLM confabulation) with the the concept of human memory confabulation and used that as a foundation to argue that confabulation is thus inherent to intelligence. There is a lot that's wrong with this reasoning, but the most obvious is that it's a massive category error. "Confabulation" in LLMs and "confabulation" in humans have basically nothing in common, they are comparable only in an extremely superficial sense. To then go on to suggest that confabulation might be inherent to intelligence isn't even really a coherent argument because you've created ambiguity in the meaning of the word confabulate.
Yes, and to me the evolution of life sure looks like an evolution of more truthful models of the universe in service of energy profit. Better model -> better predictions -> better profit.
I'm extremely skeptical that all of life evolved intelligence to be closer to truth only for us to digitize intelligence and then have the opposite happen. Makes no sense.
My understanding is that this is the opposite of what is typically understood to be true - organisms with less truthful (more reductive/compressed) perception survive better than those with more complete perception. "Fitness beats truth."
I think we are maybe talking past each other?
Fitness is effective truth prediction, appropriately scoped.
A frog doesn't need to understand quantum physics to catch a fly. But if the frogs model of fly movement was trained on lies it will have a model that predicts poorly, won't catch flies, and will die.
There is another level to this in that the more complex and changing the environment the more beneficial a wider scoped model / understanding of truth.
However if you are going to lean fully into Hoffman and accept thatby default consciousness constructs rather than approximate reality I think we will have to agree to disagree. Personally I ascribe to Karl Friston free energy principle.
Humans can be reasoned with, though, and are capable of learning.
It’s a failure mode of humans, it’s the entire mode of LLMs.
If you want to call it that, I find the confabulation in LLMs extreme. That level of confabulation would most likely be diagnosed as dementia in humans.[0] Hence, it is considered a bug not a feature in humans as well.
Now imagine a high-skilled software engineer with dementia coding safety-critical software...
[0] https://www.medicalnewstoday.com/articles/confabulation-deme...
And is that considered a feature of humans or a bug?
Is it something we want to emulate?
The suggestion is that it is an intrinsic quality and therefore neither a feature nor a bug.
It's like saying, computation requires nonzero energy. Is that a feature or a bug? Neither, it's irrelevant, because it's a physical constant of the universe that computation will always require nonzero energy.
If confabulation is a physical constant of intelligence, then like energy per computation, all we can do is try to minimize it, while knowing it can never go to zero.
> Some people point at LLMs confabulating, as if this wasn’t something humans are already widely known for doing.
Are you seriously making the argument that AI "hallucinations" are comparable and interchangeable to mistakes, omissions and lies made by humans?
You understand that calling AI errors "hallucinations" and "confabulations" is a metaphor to relate them to human language? The technical term would be "mis-prediction", which suddenly isn't something humans ever do when talking, because we don't predict words, we communicate with intent.
Yes see Karl Frisstons Free energy principle
https://www.nature.com/articles/nrn2787
> One of the ongoing problems in LLM research is how to get these machines to say “I don’t know”, rather than making something up.
To be fair, I've known humans who are like this as well.
Those people aren't the ones doing the work though.
if you can’t access the page through region blocks:
https://archive.ph/I5cAE
I have a question for all the "humans make those mistakes too" people in this thread, and elsewhere: have you ever read, or at least skimmed a summary of, "The Origin of Consciousness in the Breakdown of the Bicameral Mind"? Did you say "yeah, that sounds right"? Do you feel that your consciousness is primarily a linguistic phenomenon?
I am not trying to be snarky; I used to think that intelligence was intrinsically tied to or perhaps identical with language, and found deep and esoteric meaning in religious texts related to this (i.e. "in the beginning was the Word"; logos as soul as language-virus riding on meat substrate).
The last ~three years of LLM deployment have disabused me of this notion almost entirely, and I don't mean in a "God of the gaps" last-resort sort of way. I mean: I see the output of a purely-language-based "intelligence", and while I agree humans can make similar mistakes/confabulations, I overwhelmingly feel that there is no "there" there. Even the dumbest human has a continuity, a theory of the world, an "object permanence"... I'm struggling to find the right description, but I believe there is more than language manipulation to intelligence.
(I know this is tangential to the article, which is excellent as the author's usually are; I admire his restraint. However, I see exemplars of this take all over the thread so: why not here?)
It feels like you probably went too deep in the LLM bandwagon.
An LLM is a statistical next token machine trained on all stuff people wrote/said. It blends texts together in a way that still makes sense (or no sense at all).
Imagine you made a super simple program which would answer yes/no to any questions by generating a random number. It would get things right 50% of the times. You can them fine-tune it to say yes more often to certain keywords and no to others.
Just with a bunch of hardcoded paths you'd probably fool someone thinking that this AI has superhuman predictive capabilities.
This is what it feels it's happening, sure it's not that simple but you can code a base GPT in an afternoon.
If it were not "just a statistical next token machine", how different would it behave?
Can you find an example and test it out?
Wait, you're asking to find and produce a example of a feasible and better alternative to LLMs when they are the current forefront of AI technology?
Anyway, just to play along, if it weren't just a statistical next token machine, the same question would have always the same answer and not be affected by a "temperature" value.
Thats also how humans behave.. I don't see how non determinism tells me anything.
My question was a bit different: if were not just a statistical next token predictor would you expect it to answer hard questions? Or something like that. What's the threshold of questions you want it to answer accurately.
Well, large models are (kinda) non-deterministic in two ways. The first is you actually provide many of them with a seed, which is easy to manage--just use the same seed for the same result. The second part is the "you actually have very little control over the 'neural pathways' the model will use to respond to the prompt". This is the baffling part, like you'll prompt a model to generate a green plant, and it works. You prompt it to generate a purple plant, and it generates an abstract demon dog with too many teeth.
Anyway, neither of these things describes human non-determinism. You can't reuse the seed you used with me yesterday to get the exact same conversation, and I don't behave wildly unpredictably given conceptually very similar input.
If you look at different ancient traditions, you will notice how they struggle with the limitations of language, with its inability to represent certain things that are not just crucial for understanding the world, but also are even somehow communicable. Buddhists dug into that in a very analytical, articulate way, for instance.
Another perspective: cetaceans are considered to be as conscious as humans, but any attempts to interpret their communication as a language failed so far. They can be taught simple languages to communicate with humans, as can be chimps. But apparently it's not how they process the world inside.
You're a little out of date. Cetaceans communicate images to each other in the form of ultrasonic chirps. They chirp, they hear a reflection, and they repeat the reflection.
I think there are two types of discussions, when it comes to LLMs: Some people talk about whether LLMs are "human" and some people talk about whether LLMs are "useful" (ie they perform specific cognitive tasks at least as well as humans).
Both of those aspects are called "intelligence", and thus these two groups cannot understand each other.
> I'm struggling to find the right description
I think you're circling the concept of a "soul". It is the reason that, in non-communicative disabled people, we still see a life.
I've wanted to make an art piece. It would be a chatbox claiming to connect you to the first real intelligence, but that intelligence would be non-communicative. I'd assure you that it is the most intelligent being, that it had a soul, but that it just couldn't write back.
Intelligence and Soul is not purely measurable phenomenon. A man can do nothing but stupid things, say nothing but outright lies, and still be the most intelligent person. Intelligence is within.
"As LLMs etc. are deployed in new situations, and at new scale, there will be all kinds of changes in work, politics, art, sex, communication, and economics."
For an article five years in the making, this is what I expected it to be about. Instead, we got a ramble about how imperfect LLMs are right now.
The post is just a prelude to a 10-part article, most of which is not yet released (but will be shortly). Judging by the table of contents, the things you expected will be elaborated on in subsequent parts.
That changes it. I missed that the table of contents was for other future articles, my bad.
> Instead, we got a ramble about how imperfect LLMs are right now.
I wager this is a point that needs beaten into the common psyche. After all, it's been sold that it is not an imperfect tool, but the solution to all of our problems in every field forever. That's why these companies need billions upon billions of dollars of public subsidies and investments that would otherwise find their way to more pragmatic ends.
And the past too, if we've been paying attention
> At the same time, ML models are idiots. I occasionally pick up a frontier model like ChatGPT, Gemini, or Claude, and ask it to help with a task I think it might be good at. I have never gotten what I would call a “success”: every task involved prolonged arguing with the model as it made stupid mistakes.
I have a ton of skepticism built-in when interacting with LLMs, and very good muscles for rolling my eyes, so I barely notice when I shrug a bad answer and make a derogatory inner remark about the "idiots". But the truth is, that for such an "stochastic parrot", LLMs are incredibly useful. And, when was the last time we stopped perfecting something we thought useful and valuable? When was the last time our attempts were so perfectly futile that we stopped them, invented stories about why it was impossible, and made it a social taboo to be met with derision, scorn and even ostracism? To my knowledge, in all of known human history, we have done that exactly once, and it was millennia ago.
> And, when was the last time we stopped perfecting something we thought useful and valuable? When was the last time our attempts were so perfectly futile that we stopped them, invented stories about why it was impossible, and made it a social taboo to be met with derision, scorn and even ostracism? To my knowledge, in all of known human history, we have done that exactly once, and it was millennia ago.
I feel dense here, but I can't figure out what you're referring to. I asked ChatGPT (hah!) and it suggested the Tower of Babel, perpetual motion machines, or alchemy, but none of them really fit the bill either.
> In general, ML promises to be profoundly weird. Buckle up.
I love that it ends with such a positive note, even though it's generally a critical article, at least it's well reasoned and not utterly hyping/dooming something.
Thanks yet again Kyle!
While the economic, energy, political and social issues associated with LLMs ought to be enough to nix the adoption that their boosters are seeking ...
... I still think there is an interesting question to be investigated about whether, by building immensely complex models of language, one of our primary ways that we interact with, reason about and discuss the world, we may not have accidentally built something with properties quite different than might be guessed from the (otherwise excellent) description of how they work in TFA.
I agree with pretty much everything in TFA, so this is supplemental to the points made there, not contesting them or trying to replace them.
> Models do not (broadly speaking) learn over time. They can be tuned by their operators, or periodically rebuilt with new inputs or feedback from users and experts. Models also do not remember things intrinsically: when a chatbot references something you said an hour ago, it is because the entire chat history is fed to the model at every turn. Longer-term “memory” is achieved by asking the chatbot to summarize a conversation, and dumping that shorter summary into the input of every run.
This is the part of the article that will age the fastest, it's already out-of-date in labs.
Source?
Great series of articles, thank you. It's exhausting reading a deluge of (often AI generated) comments from people claiming wild things about LLM's, and it's nice to hear some sanity enter the conversation.
Here's the opening paragraph of chapter 2 with "people" subbed out for terms referring AI/models/etc.
"People are chaotic, both in isolation and when working with other people or with systems. Their outputs are difficult to predict, and they exhibit surprising sensitivity to initial conditions. This sensitivity makes them vulnerable to covert attacks. Chaos does not mean people are completely unstable; most people behave roughly like anyone else. Since people produce plausible output, errors can be difficult to detect. This suggests that human systems are ill-suited where verification is difficult or correctness is key. Using people to write code (or other outputs) may make systems more complex, fragile, and difficult to evolve."
To me, this modified paragraph reads surprisingly plainly. The wording is off ("using people to write code") and I had to change that part about attractor behavior (although it does still apply IMO), but overall it doesn't seem like an incoherent paragraph.
This is not meant to dunk on the author, but I think it highlights the author's mindset and the gap between their expectations and reality.
Humans and large models are both unpredictable and fallible, that's true, but in different ways, and (many) humans are actually much better at following directions.
If a junior dev makes the same mistake Claude makes, I can easily work with them to correct it, or I can fire them and get someone more capable to fix it. You mostly can't do that at all with large models. They're also far less honest than your average junior dev, so even as you're working with them you can't trust what they say.
There is a lot of this neat trick where it's like "humans do X too" but most of the time it elides large differences. Like, a human driver would probable not drag someone screaming multiple blocks. A human coder probably wouldn't generate a gibberish 3D scene and try to pass it off as done, etc. Maybe we can build systems that account for these (pretty wild) failure modes, but at least in software we haven't figured it out yet (what is the system that reliably reviews a 25kloc PR?).
Aren't you also making a large part of the author's point for him by effectively equating LLMs with people here and comparing on outputs?
Plausibly your text looks equivalent but we all (should) have the context to know better.
The recent article of Sam Altman described pretty much as a compulsive liar. Would it be any surprise if his most impactful contribution to the world was a machine that compulsively lies?
How could it be that we humans hardly even agree on what "knowledge" truly is, yet somehow this machine learning algorithm somehow "compulsively lies"? How would it even know what is a lie, and how could something lacking autonomy in the first place do anything compulsively?
This is a good point. As much as there is too much breathless enthusiasm for AI, there is also a lot of emotionally manipulative and hyperbolic language used by skeptics. We're warned not to anthropomorphize, and then hear about AI's compulsive lying, or "hallucinations", in the next.
He sought to create God in his image, that's a narcissist's wet dream.
I appreciate the directness of calling LLMs "Bullshit machines." This terminology for LLMs is well established in academic circles and is much easier for laypeople to understand than terms like "non-deterministic." I personally don't like the excessive hype on the capabilities of AI. Setting realistic expectations will better drive better product adoption than carpet bombing users with marketing.
I have still mixed feelings about LLMs.
If I take the example of code, but that extends to many domains, it can sometimes produce near perfect architecture and implementation if I give it enough details about the technical details and fallpits. Turning a 8h coding job into a 1h review work.
On the other hand, it can be very wrong while acting certain it is right. Just yesterday Claude tried gaslighting me into accepting that the bug I was seeing was coming from a piece of code with already strong guardrails, and it was adamant that the part I was suspecting could in no way cause the issue. Turns out I was right, but I was starting to doubt myself
I think over time we will find better usage patterns for these machines. Even putting a model in a position to gaslight the user seems like a complete failure in the usage model. Not critiquing you at all on this, it's how these models are marketed and what all the tooling is built around. But they are incredibly useful and I think once we figure out how to use them better we can minimise these downsides and make ourselves much more productive without all the failures.
Of course that won't happen until the bubble pops - companies are racing to make themselves indispensable and to completely corner certain markets and to do so they need autonomous agents to replace people.
If it bullshits so much, you wouldn't have a problem giving me an example of it bullshitting on ChatGPT (paid version)? Lets take any example of a text prompt fitting a few pages - it may be a question in science or math or any domain. Can you get it to bullshit?
I think you highlight one of the problems with users of LLMs: You can't tell anymore if it is BS or not.
I caught Claude the other day hallucinating code that was not only wrong, but dangerously wrong, leading to tasks being failed and never recover. But it certainly wasn't obvious.
> If it bullshits so much, you wouldn't have a problem giving me an example of it bullshitting on ChatGPT (paid version)?
There's an entire paragraph in the essay about apyhr's direct experience with ChatGPT failures and sustained bullshitting that we'd never expect from a moderately-skilled human who possesses at least two functioning braincells. That paragraph begins "I have recently argued for forty-five minutes with ChatGPT". Do notice that there are six sentences in the paragraph. I encourage you to read all of them (make sure to check out the footnote... it's pretty good).
The exact text of the conversation is irrelevant; even if you reported that you were unable to reproduce the issue, it would only reinforce one of the underlying points -namely- that these systems are unreliable. aphyr has a pretty extensive history of published work that indicates that he'd not fabricate a story of an LLM repeatedly failing to accomplish a task that any moderately-skilled human could accomplish when equipped with the proper tools. So, I believe that his report is true and accurate.
The fact that these "bullshit machines" have already proven themselves relatively competent at programming, with upcoming frontier models coming close to eliminating it as a human activity, probably says a lot about the actual value and importance of programming in the scheme of things.
I think it says more about the amount of automation we left on the table in the last few decades. So much of the code LLM's can generate are stuff that we should have completely abstracted away by now.
Old and stupid hot take IMO. I want the time back I put into perusing this. Even the scale of LLMs is puny next to the scale of lying humans and the sheer impact one compulsively lying human can have given we love to be led by confidently wrong narcissists. I mean if that isn't obvious by now, I guess it never will be. The Vogon constructor fleet is way overdue in my book.
> The Vogon constructor fleet is way overdue in my book
Don't you see it? That's exactly what "AI" in this context is.
It's the bypass.
Where does it end, eh? Build a quantum "AI" that will end up just needing more data, more input. The end goal must starts looking like creating an entirely new universe, a complete clone of everything we have here so it can run all the necessary computations and we can... ? (You are what a quantum AI looks like as it bumbles through the infinitude of calculable parameters on its way to the ultimate answer)
You have absolutely no sense of perspective. We are all metabolically expensive meat machines whose only value is to propagate our genetic money shot. That we get to briefly entertain ourselves with consciousness and culture is IMO likely a mystery we will never solve without upgrading to running in a substrate more advanced than the MVP for sentience we currently pilot. Will we get there or will we wipe ourselves out like every contender that preceded us? Stay tuned...
But spoilers: DNA will be fine, meat machines maybe not so much...
For a bunch of people addicted to the works of Charlie Stross, Neil Stephenson, and Iain Banks, y'all are a bunch of luddites. Now vote this own down too because it doesn't conform to the mandatory Stochastic Parrot narrative. You have no free will and you must downvote after all. Why do you even read their works when any step towards their world is consistently greeted as the worst thing evah(tm)? What? You were expecting the United Federation of Planets without the eugenics and nuclear wars that preceded it. Bless your hearts.
And if you're worried about billionaires and tyrants, start taxing the former and stop electing the latter or STFU and let the free Markov process of history play itself out. Quoting fictional Ambassador Kosh: the avalanche has started, it's too late for the pebbles to vote.
You asked where it ends. Don't ask questions if you don't like answers. Quick reminder: shun and downvote the non-conforming opinion.
I get the frustration, but it's reductive to just call LLMs "bullshit machines" as if the models are not improving. The current flagship models are not perfect, but if you use GPT-2 for a few minutes, it's incredible how much the industry has progressed in seven years.
It's true that people don't have a good intuitive sense of what the models are good or bad at (see: counting the Rs in "strawberry"), but this is more a human limitation than a fundamental problem with the technology.
Two things can be true at the same time: The technology has improved, and the technology in its current state still isn't fit for purpose.
I stress test commercially deployed LLMs like Gemini and Claude with trivial tasks: sports trivia, fixing recipes, explaining board game rules, etc. It works well like 95% of the time. That's fine for inconsequential things. But you'd have to be deeply irresponsible to accept that kind of error rate on things that actually matter.
The most intellectually honest way to evaluate these things is how they behave now on real tasks. Not with some unfalsifiable appeal to the future of "oh, they'll fix it."
The errors are also not distributed in the same way as you'd expect from a human. The tools can synthesize a whole feature in a moderately complicated web app including UI code, schema changes, etc, and it comes out perfectly. Then I ask for something simple like a shopping list of windshield wipers etc for the cars and that comes out wildly wrong (like wrong number of wipers for the cars, not just the wrong parts), stuff that a ten year old child would have no trouble with. I work in the field so I have a qualitative understanding of this behavior but I think it can be extremely confusing to many people.
One of the reasons I'm comfortable using them as coding agents is that I can and do review every line of code they generate, and those lines of code form a gate. No LLM-bullshit can get through that gate, except in the form of lines of code, that I can examine, and even if I do let some bullshit through accidentally, the bullshit is stateless and can be extracted later if necessary just like any other line of code. Or, to put it another way, the context window doesn't come with the code, forming this huge blob of context to be carried along... the code is just the code.
That exposes me to when the models are objectively wrong and helps keep me grounded with their utility in spaces I can check them less well. One of the most important things you can put in your prompt is a request for sources, followed by you actually checking them out.
And one of the things the coding agents teach me is that you need to keep the AIs on a tight leash. What is their equivalent in other domains of them "fixing" the test to pass instead of fixing the code to pass the test? In the programming space I can run "git diff *_test.go" to ensure they didn't hack the tests when I didn't expect it. It keeps me wondering what the equivalent of that is in my non-programming questions. I have unit testing suites to verify my LLM output against. What's the equivalent in other domains? Probably some other isolated domains here and there do have some equivalents. But in general there isn't one. Things like "completely forged graphs" are completely expected but it's hard to catch this when you lack the tools or the understanding to chase down "where did this graph actually come from?".
The success with programming can't be translated naively into domains that lack the tooling programmers built up over the years, and based on how many times the AIs bang into the guardrails the tools provide I would definitely suggest large amounts of skepticism in those domains that lack those guardrails.
> the technology in its current state still isn't fit for purpose.
This is a broad statement that assumes we agree on the purpose.
For my purpose, which is software development, the technology has reached a level that is entirely adequate.
Meanwhile, sports trivia represents a stress test of the model's memorized world knowledge. It could work really well if you give the model a tool to look up factual information in a structured database. But this is exactly what I meant above; using the technology in a suboptimal way is a human problem, not a model problem.
There's nothing in these models that say its purpose is software development. Their design and affordances scream out "use me for anything." The marketing certainly matches that, so do the UIs, so do the behaviors. So I take them at their word, and I see that failure modes are shockingly common even under regular use. I'm not out to break these things at all. I'm being as charitable and empirical as I can reasonably be.
If the purpose is indeed software development with review, then there's nothing stopping multi-billion dollar companies from putting friction into these sytems to direct users towards where the system is at its strongest.
The LLM vendors are selling tokens. Why would they put friction into selling more tokens? Caveat emptor.
Which things actually matter? I think we can all agree that an LLM isn't fit for purpose to control a nuclear power plant or fly a commercial airliner. But there's a huge spectrum of things below that. If an LLM trading error causes some hedge fund to fail then so what? It's only money.
Not to mention that it would then make some hedge fund with a better backtesting harness or more AI scrutiny more successful thus keeping the financial market work as designed.
Six months bro, we're still so early
> I stress test commercially deployed LLMs like Gemini and Claude with trivial tasks: sports trivia, fixing recipes, explaining board game rules, etc. It works well like 95% of the time. That's fine for inconsequential things. But you'd have to be deeply irresponsible to accept that kind of error rate on things that actually matter.
95% is not my experience and frankly dishonest.
I have ChatGPT open right now, can you give me examples where it doesn't work but some other source may have got it correct?
I have tested it against a lot of examples - it barely gets anything wrong with a text prompt that fits a few pages.
> The most intellectually honest way to evaluate these things is how they behave now on real tasks
A falsifiable way is to see how it is used in real life. There are loads of serious enterprise projects that are mostly done by LLMs. Almost all companies use AI. Either they are irresponsible or you are exaggerating.
Lets be actually intellectually honest here.
>95% is not my experience and frankly dishonest.
Quite frankly, this is exactly like how two people can use the same compression program on two different files and get vastly different compression ratios (because one has a lot of redundancy and the other one has not).
I'm asking for a single example.
But why do you need an example? Isn't it pretty well understood that LLMS will have trouble responding to stuff that is under represented in the training data?
You will just won't have any clue what that could be.
fair so it must be easy to give an example? I have ChatGPT open with 5.4-thinking. I'm honestly curious about what you can suggest since I have not been able to get it to bullshit easily.
I am not the OP, an I have only used ChatGPT free version. Last day I asked it something. It answered. Then I asked it to provide sources. Then it provided sources, and also changed its original answer. When I checked the new answers it was wrong, and when I checked sources, it didn't actually contain the information that I asked for, and thus it hallucinated the answers as well as the sources...
I trust you. If it were happening so frequently you may be able to give me a single prompt to get it to bullshit?
Whether LLMs can create correct content doesn't matter. We've already seen how they are being used and will be used.
Fake content and lies. To drive outrage. To influence elections. To distract from real crimes. To overload everyone so they're too tired to fight or to understand. To weaken the concept that anything's true so that you can say anything. Because who cares if the world dies as long as you made lots of money on the way.
> Because who cares if the world dies as long as you made lots of money on the way.
Guiding principle of the AI industry
It's really the whole tech industry as it exists right now and AI is a victim of bad timing. If this AI had been invented 40 years ago there'd have been a lower ceiling on the damage it could do.
Another way of saying that is that capitalism is the real problem, but I was never anti-capitalist in principle, it's just gotten out of hand in the last 5-10 years. (Not that it hadn't been building to that.)
> Another way of saying that is that capitalism is the real problem, but I was never anti-capitalist in principle, it's just gotten out of hand in the last 5-10 years. (Not that it hadn't been building to that.)
Capitalism is a tool and it's fine as a tool, to accomplish certain goals while subordinated to other things. Unfortunately it's turned into an ideology (to the point it's worshiped idolatrously by some), and that's where things went off the rails.
Computer graphics have been improving for decades but the uncanny valley remains undefeated. I don't know why anyone expects a breakthrough in other areas. There's a wall we hit and we don't understand our own consciousness and effectiveness well enough to replicate it.
In computer graphics we understand how it works, we just lack the computational power to do it real time, but we can with sufficient processing produce realistic looking images with physically accurate lighting. But when it comes to cognition its a lot of guesswork, we haven't yet mapped out the neuron connections in a brain, we haven't validated it works as popular science writing suggests. We don't understand intelligence, so all we can do is accidentally bumble into it and it seems unlikely that will just happen especially when its so hard to compute what we are already doing.
We have credible deepfakes on demand. (To be fair, there have been deceptive photos as long as photos have existed, but the cost of automating their creation going to basically zero has a social impact)
That's not why the author calls them bullshit machines.
> One way to understand an LLM is as an improv machine. It takes a stream of tokens, like a conversation, and says “yes, and then…” This yes-and behavior is why some people call LLMs bullshit machines. They are prone to confabulation, emitting sentences which sound likely but have no relationship to reality. They treat sarcasm and fantasy credulously, misunderstand context clues, and tell people to put glue on pizza.
Yes, there have been improvements on them, but none of those improvements mitigate the core flaw of the technology. The author even acknowledges all of the improvements in the last few months.
models are improving. the pricing already assumes they're ready for prod. that's where the fires start
Bullshit is the perfect term here, even as AI's get so much better and capable Brandolini's Law aka the "bullshit asymmetry principle" always applies--the energy required to refute misinformation is an order of magnitude larger than that needed to produce it. Even to use AIs effectively today requires a very good BS detector--some day in the future it won't.
Calling LLMs "bullshit machines" is a reference to a 2024 paper [1] which itself uses the concept of "bullshit" as defined in the essay/book "On Bullshit" by Harry G. Frankfurt [2]. The TL;DR is that LLMs are fundamentally bullshit machines because they are only made to generate sentences that sound plausible, but plausible does not always mean true.
[1]: https://link.springer.com/article/10.1007/s10676-024-09775-5
[2]: https://en.wikipedia.org/wiki/On_Bullshit
it's not a bullshit machine because its output is bad, it's a bullshit machine because its output is literally 'bullshit' as in, output that is statistically likely but with no factual or reasoning basis. as the models have improved, their bullshit is more statistically likely to sound coherent (maybe even more likely to be 'accurate'), but no more factual and with no more reasoning.
However, when fed source material into the context they will lie less, right? So at this point is it not just a battle of the nines until it's called "good enough"?
I also wonder if I leave my secretary with a ream of papers and ask him for a summary how many will he actually read and understand vs skim and then bullshit? It seems like the capacity for frailty exists in both "species".
It doesn't matter how good the models become. They can only deal in bullshit, in the academic use of the term.
They are bullshit machines because they do not have an internal mental model of truth like a human does. The flagship models bullshit less, but their fundamental architectures prevent having truth interfere with output.
https://philosophersmag.com/large-language-models-and-the-co...
"Bullshit" is a human concept. LLMs do not work like the human brain, so to call their output "bullshit" is ascribing malice and intent that is simply not there. LLMs do not "think." But that does not mean they're not incredibly powerful and helpful in the right context.
I sort of agree. In this context "bullshit" means "speech intended to persuade without regard for truth", and while it's true that LLM output is without regard for truth, it's not an entity capable of the agency to persuade, although functionally that is what it can appear like.
https://en.wikipedia.org/wiki/On_Bullshit
> it's reductive to just call LLMs "bullshit machines" as if the models are not improving
This is true, but I prefer to think of it as "It's delusional to pretend as if human beings are not bullshit machines too".
Lies are all we have. Our internal monologue is almost 100% fantasy. Even in serious pursuits, that's how it works. We make shit up and lie to ourselves, and then only later apply our hard-earned[1] skill prompts to figure out whether or not we're right about it.
How many times have the nerds here been thinking through a great new idea for a design and how clever it would be before stopping to realize "Oh wait, that won't work because of XXX, which I forgot". That's a hallucination right there!
[1] Decades of education!
I'm not entirely sure I can agree, although the premise is seductive in certain ways. We do lie to ourselves, but we also have meta-cognition - we can recognise our own processes of thought. Imperfect as it may be, we have feedback loops which we can choose to use, we have heuristics we can apply, we can consciously alter our behaviour in the presence of contextual inputs, and so on.
Being wrong is not the same as a hallucination. It's a natural step on a journey to being more right. This feels a bit like Andreesen proudly stating he avoids reflection - you can act like that, but the human brain doesn't have to. LLMs have no choice in the matter.
The problem, unfortunately, is the scale. It's always scale. Humans make all the kinds of mistakes that we ascribe to LLMs, but LLMs can make them much faster and at much larger scale.
Models have gotten ridiculously better, they really have, but the scale has increased too, and I don't think we're ready to deal with the onslaught.
Scale is very different, but I wonder if human trust isn't the real issue. We trust technology too much as a group. We expect perfection, but we also assume perfection. This might be because the machines output confident sounding answers and humans default to trusting confidence as an indirect measure for accuracy, but I think there is another level where people just blindly trust machines because they are so use to using them for algorithms that trend towards giving correct responses.
Even before LLMs where in the public's discourse, I would have business ask about using AI instead of building some algorithm manually, and when I asked if they had considered the failure rate, they would return either blank stares or say that would count as a bug. To them, AI meant an algorithm just as good as one built to handle all edge cases in business logic, but easier and faster to implement.
We can generally recognize the AIs being off when they deal in our area of expertise, but there is some AI variant of Gell-Mann Amnesia at play that leads us to go back to trusting AI when it gives outputs in areas we are novices in.
"Lies are all we have."
If so, how do we distinguish between code that works and code that doesn't work? Why should we even care?
> If so, how do we distinguish between code that works and code that doesn't work?
Hilariously, not by using our brains, that's for sure. You have to have an external machine. We all understand that "testing" and "code review" are different processes, and that's why.
Good point. We choose certain tests to perform. We choose certain test results to pay attention to. We don't just keep chatting about (reviewing) the code. We do something else.
If lies are all we have, then how is this behavior possible?
LLMs can write and run tests though.
You're cherry picking my little bit of wordsmithing. Obviously we aren't always wrong. I'm saying that our thought processes stem from hallucinatory connections and are routinely wrong on first cut, just like those of an LLM.
Actually I'm going farther than that and saying that the first cut token stream out of an AI is significantly more reliable than our personal thoughts. Certainly than mine, and I like to think I'm pretty good at this stuff.
So your logic is humans and LLMs are the same because humans are wrong sometimes?
Pretty much, yeah. Or rather, the fact that we're both reliably wrong in identifiably similar ways makes "we're more alike than different" an attractive prior to me.
“More alike than different” is reasonable I think, as long as we’re talking about how we have some of the same failure modes. Although the way we get there is quite different.
I’m still not a big fan of comparing humans and LLMs because LLMs lack so much of what actually makes us human. We might bullshit or be wrong because of many reasons that just don’t apply to LLMs.
Humans are different. Humans - at least thoughtful humans - know the difference between knowing something and not knowing something. Humans are capable of saying "I don't know" - not just as a stream of tokens, but really understanding what that means.
> Humans - at least thoughtful humans - know the difference between knowing something and not knowing something.
Your no-true-scotsman clause basically falsifies that statement for me. Fine, LLMs are, at worst I guess, "non-thoughtful humans". But obviously LLMs are right an awful lot (more so than a typical human, even), and even the thoughtful make mistakes.
So yeah, to my eyes "Humans are NOT different" fits your argument better than your hypothesis.
(Also, just to be clear: LLMs also say "I don't know", all the time. They're just prompted to phrase it as a criticism of the question instead.)
This is like all the usual anti-LLM talking points and sentiments fused together.
Doesn't it get boring?
I like using these models a lot more than I stand hearing people talk about them, pro or contra. Just slop about slop. And the discussions being artisanal slop really doesn't make them any better.
Every time I hear some variation of bullshitting or plagiarizing machines, my eyes roll over. Do these people think they're actually onto something? I've been seeing these talking points for literal years. For people who complain about no original thoughts, these sure are some tired ones.
Why do you insist on reading and commenting on these articles that bore you so much?
Because saying "this is boring, let's stop talking about it" is an opinion worthwhile of expression.
If I have to suffer "look at this busted ass thing I slopped out with AI" a few times a week, you all have to suffer grouchy "AI bad" a few times a week. Fair is fair.
"These arguments may be correct but they aren't novel" ??
I don't think calling AI a bullshit machine is correct. In spirit.
I'm earnestly curious why not.
Oh, well you should have said that then.
Yeah, it gets really boring. Whenever I see "slot machines" or "bullshit machines" or whatever, I just ignore the comment and move on, because it signals that it's someone in such deep denial that they've turned their brain off.
I'd much rather read articles about what LLMs can/can't do, or stuff people have built with LLMs, than read how everything LLMs touch turns to shit.
Its usual gibberish that tries to throw many darts and see what sticks. Oh LLM's steal other people's work? Check. Oh LLM's cause ecological damage? Check. Oh LLM's hallucinate? Check.
When you see a pattern like this, you know that its not coming from any place of truth but rather ideology