We're choosing to call LLMs (and the little "harness" programs that query them in loops and execute their output) "AI", even though it doesn't make much sense.
I absolutely love this technology but these aren't autonomous intelligences. They're little programs executing Bash scripts from JSON output.
Our ideas about AI were naive. We thought passing a basic Turing test would require human-like intelligence. It turned out to be possible with fairly basic statistical text generation, because fooling humans is easy.
It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning. The irony is that the startup founders most worried about "AI" have created so much hype and funding that we may very well figure out how to build "real" AI.
We've chosen to call Deep Blue and Half-Life 1 NPCs "AI" too.
It boggles my mind that this "b-b-but it's not actual real AI" whine is even a thing. Were people saying this living in the cave for the past 5 decades of AI research?
You can call your little doggy "AI" if it makes you happy.
But when you call something "AI" and it tells you to walk instead of drive to the car wash, you're not talking about the "AI" science fiction authors were dreaming of.
Yeah it's a trick question, the human error rate for it was about 30% (higher depending on the country).
The thing there though is that, if a human were given time to think about it, they'd probably go "hang on a minute", and with the LLMs that didn't seem to happen. They just kept confidently reasoning down the absurd path.
That reminds me, I recently had an AI write a ton of tests proving the "correctness" of a feature it had implemented completely backwards. (I noted that if I had been using a language that required formal proofs, that wouldn't have helped either: it would have just provided a formal proof for the absurd implementation!)
It's one example that points out a major (possibly fundamental) flaw. I can point to prompt injection as another example. There are tons more if you're interested.
Are you actually claiming LLMs operate based on human-like intelligence?
We're on Hacker News. Do I really have to point out the existence of social engineering to you? Or that scamming old people out of their life savings is a profitable enough activity that there are entire call centers dedicated to the task?
Humans keep overestimating just how high the bar of "human-like intelligence" is.
You're saying that a class of mistakes points out a "major (possibly fundamental) flaw". I'm pointing out some very similar classes of mistakes in humans - well known, well documented and widely exploited. They just keep paying the "IRS" in gift cards, buying lottery tickets and getting the captain's age wrong.
If you're using the existence of flaws in LLMs to deny the claim of intelligence to them, then why do "generally intelligent" humans exhibit some impressively similar-looking flaws?
And, if we're talking about that conspicuous similarity - do they actually fail "for entirely different reasons"? Or do you just want the reasons to be "entirely different" - and not the same reasons viewed at a different angle?
Because the similarities between humans falling for trick questions or scams, and LLMs falling for adversarial questions or prompt injections don't look coincidental to me at all.
One of the oldest patterns in scamming is overwhelming and confusing the victim. Numerous prompt injection methods seek to overwhelm and confuse an LLM - if an LLM can't keep track of things, can't grasp what's going on, it's far more likely to lose track of what's a prompt and what's data, overlook past instructions or go past its behavioral guardrails.
And humans who fall for trick questions like "1kg of feathers" or "captain's age" due to shallow attention and naive pattern matching? They fail in surprisingly similar ways to how LLMs fail on SimpleBench tasks that are filled with overwhelming adversarial distractors. Many "trick questions" are tricky to humans and LLMs alike - to the point that it's unlikely to be coincidental.
> If you're using the existence of flaws in LLMs to deny the claim of intelligence to them...
That's not the point at all. It's the fact that they fail in ways completely unlike humans.
You also have the burden of proof reversed. Its on you to prove these LLM agents are human-like intelligences if that's your claim. No one can prove this because it's false.
If they didn't you wouldn't need the operator, you'd have replaced all your programmers with no drawbacks by now. As long as we keep hiring humans that is all the evidence you need that these AI fails in ways humans don't.
This isn't even controversial. The proof is available to anyone who uses these systems:
They hallucinate tool state, drift from the objective while seeming to comply, switch languages randomly (Cyrillic or Japanese characters in output), confuse tasks they've planned for completed ones, and of course follow prompt injections embedded in files or web pages.
> Are you actually claiming LLMs operate based on human-like intelligence?
Ok, so we've established that it doesn't work like a human being. To paraphrase Dijkstra: The submarine doesn't swim.
But does it exactly sail either? An LLM doesn't exactly work like traditional deterministic software either, does it?
And yet it moves. You can put in data and ask it to process it, and you'll get an answer that's in some ballpark. Closer to quantum or stochastic computing perhaps, but that's not it either, is it? Or SAT-solving? Eh. It's its own computing approach. If you have a problem where the asking is hard but the verification is cheap, it might just be the right tool for the job.
You can call your little doggy "AI" if it makes you happy.
Or you can keep calling them stochastic parrots as they solve decades-old open problems. The real question is how useful they are, and the answer "not at all" increasingly requires flat-earth levels of denial.
it tells you to walk instead of drive to the car wash, you're not talking about the "AI" science fiction authors were dreaming of
They sort of are. Think of Data from Star Trek TNG failing to understand figures of speech. Not that it's terribly relevant; humans regularly fall for tricks like "Paris in the the spring" or "where do you bury the survivors".
> Or you can keep calling them stochastic parrots as they solve decades-old open problems.
I didn't use that phrase at all. But computers calculated digits of π to trillions of digits. With a chat interface for a Python math program would look like the most impressive math genius if you took it back a few decades.
> The real question is how useful they are...
That's not the "real question" but an entirely different question that is easily answered. Nothing I wrote suggested they're not incredibly useful.
> Data from Star Trek TNG failing to understand figures of speech.
These are just little instances of bad writing. Data is very much an attempt at displaying a human-like intelligence.
Sure, just keep moving the goalposts. It's not a "real AI" because it can't take over the US military command and kick off WW3 and finish the survivors off with killer robots yet!
That's how it works though. The moment we have "AI" and see something working, it immediately ceases to be magic because "it's just a program after all." Aligning on a true definition of Artificial Intelligence is a very vexing problem.
We could've slapped a chat interface on calculators and called them "AI" because they can do superhuman math instantly. Most technical people would've thought that was stupid.
Let's talk about category mistakes. You've been here since 2007, according to your other reply. You understand that calculators have as much to do with mathematics as telescopes have to do with cosmology. Right?
If someone unskilled at math brings a calculator to an international math competition, they will not succeed at solving many problems. Most likely, they will solve none at all. But if they bring a frontier LLM (and succeed at concealing it from the organizers), they can walk away with a gold medal. Such a feat requires intelligence... and if the contestant didn't provide the intelligence himself/herself, where'd it come from?
That means that analogies involving calculators are completely useless when the topic is AI. Calculators are not, and can never be, intelligent. LLMs are nothing even remotely like calculators.
> If someone unskilled at math brings a calculator to an international math competition, they will not succeed at solving many problems.
> But if they bring a frontier LLM (and succeed at concealing it from the organizers), they can walk away with a gold medal.
Of course you could win all kinds of math competitions with a concealed calculator. Maybe you'd need a fancy one, like a little SBC running Python. Anything complex and timed would be easy to win. You'd look like a genius to anyone who didn't know you had it.
> Such a feat requires intelligence... and if the contestant didn't provide the intelligence himself/herself, where'd it come from?
From computer software running on computer hardware, just like a calculator.
Calculating trillions of digits of pi also requires intelligence far beyond human capacity.
Computers displaying intelligence doesn't imply human-like intelligence. This is the source of confusion.
The idea that a calculator, or a calculator with Python, or even a calculator with a proof assistant (and every book ever written on math) would help a random person at e.g. IMO or Putnam is fairly revealing.
Funny enough, calculators went through this exact same thing when they came out. "If the calculator can do math for the students, will they still learn?"
We did have that, people thought computers would overtake humans very soon when computers got better than humans at such things. It happens every single time computers do a new thing that previously humans were better at. Then 10 years later people see, oh that is just calculations, of course computers are better at that.
They seem to be making a religious argument at this point. They don't like that the AI is a very poorly defined term and they are mad as hell about it to the point of irrational forum posting.
Is this a real debate? AI has a well-defined technical definition. It’s right there in the Wikipedia [1] . Yes it’s quite a broad umbrella of systems and algorithms but it’s all AI
> It boggles my mind that this "b-b-but it's not actual real AI" whine is even a thing.
As I understand it, a major reason it's a consistent chorus is because people don't want the "AI is here" talk to drown out (and thus slow the arrival or distribution of) speech/text/popular-understanding about actual strong AGI.
To make an analogy, it could be like this:
Some people were expecting 100 tulips (because they were told tulips are available and can be ordered), and they ordered them. They received 100 daisies. And were saying "OMG, THE TULIPS ARE HERE! THE TULIPS ARE HERE!"
A nearby observer might have said, "You know, those are daisies. Not tulips."
And 95% of people might have said back, "WE GOT 100 TULIPS! SAYS SO RIGHT HERE! THEY ARE BEAUTIFUL! STOP BEING A NAY-SAYER! THESE ARE BEAUTIFUL TULIPS!"
The 5% could just to think to themselves, and could get chastised by the crowd, if they were to say say it out loud: "Well, those are not nearly as beautiful as tulips. And if you don't take it up with the seller, you may never receive the real tulips you were after. Since you think or at least act as though you've been sold them already."
In my eyes that "chorus" is just insecurity talking.
If it's not "actual real AI", we can keep pretending that human intelligence is something distinct and special - and that what our computers are doing now is some sort of other, obviously fake and vastly inferior thing.
When Deep Blue won at chess, people didn't revise their estimates of AI capabilities upwards. They revised their estimates of how much intelligence is required to play chess at world level downwards, by a lot. Surely playing chess must have never required any intelligence in the first place!
Now, the list of things that "must have never required any intelligence in the first place" includes gems like "reading comprehension at high school level", "copywriting", "frontend work", "CTF tasks", "theory of mind", "arguing with people online" and more.
If the goalposts were moved far enough that the claim to "actual intelligence" is denied to a double digit percentage of human population, hasn't something gone wrong somewhere?
It used to be assumed that playing chess would require the same level of general purpose problem solving cognitive skills that the best chess players possess. But of course a Chess grandmaster that spend a few minutes learning Go can beat a Chess AI at Go with no trouble at all, because a chess AI is incapable of making effective moves in Go at all. Clearly those expectations were incorrect. Pointing that out isn't revisionism.
On the other hand, intelligence is an incredibly broad term. About as broad as a term can get. Arguably Eliza, or an Excel macro has some degree of decision making ability in some sense, it's just unbelievably primitive.
So, we need to be clearer what we mean by intelligence. We're learning that as we go along. At least now we have a few more bits of the map between us and an IF statement visible to us.
>So, we need to be clearer what we mean by intelligence
I disagree in one sense. The word intelligence is burned, mostly useless at this point. I've been a strong proponent of new terms that break intelligence into much smaller subcategories so we can define what different software, humans, and animals have.
> When Deep Blue won at chess, people didn't revise their estimates of AI capabilities upwards. They revised their estimates of how much intelligence is required to play chess at world level downwards, by a lot. Surely playing chess must have never required any intelligence in the first place!
You are wrong, many did temporarily revise their estimates of AI capabilities upwards, but then 10 years later they realized they were wrong and adjusted chess downward as you say.
What I'm seeing here, reading this thread, is that "intelligence" isn't a thing.
"Thing" in terms of a quantifiable that you can measure with tools and reason about, reproducibly. Everyone's got some idea what it is, so you get lots of different angles, but no one has an Intelligence Ruler we can hold up to a text output and say, yep, this one's got an INT of 14.
> It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning.
We've called that "AGI" since the late 90s/early 00s (depending on whether you count first use or popularization). Even if AGI does come to pass, we'll still need "AI" since not all forms of AI will be AGI.
Well, the labs are in a weird bind. They need to keep increasing autonomy so the agents can do increasingly complex, long-horizon tasks. But at the same time, they're closely guarding against autonomy in the sense of "pursuing its own goals."
Over the past year and a half especially, several labs have mentioned adding safeguards against self-replication, resistance to shutdown etc. (Notably, shortly after they all started bragging about involving them in the AI training loop itself, i.e. "self-improvement".)
My point here is that the autonomy of which you seek might be only a few small mutations away, but the labs are actively working to prevent such a mutation. I don't expect that situation to last for very long.
Not that I expect an AI lab will be overtaken by a rogue intelligence any time soon, but that as the cost of training goes down, I expect more "open minded" organizations and individuals to become involved.
It only takes one.
That's going to be the beginning of a new era of biology, and it's a little unsettling to think about.
But immitating others is about the only thing genAI does. Sometimes "others" is a 'programmer', sometimes "others" is an 'artist', but regardless, it still does it poorly.
Maybe just a very rigorous version of the Turing test? Modern LLMs can superficially simulate conversation but it's trivial to force them into revealing their non-human like intelligence.
They've been "patched" since but all models fail basic tests like "Should I walk or drive to the car wash which is 100 feet away" by recommending you walk.
So you'd just ask questions that require theory of mind, abstract and common sense reasoning, causal inference, learning novel rules, transferring knowledge novel situations, recognizing ambiguity, etc.
If an alien lands on Earth and learns English, would you deem it non-intelligent if you can tell it apart from a human in conversation?
I think we should consider slime mold intelligent, and realise that it's a spectrum. Path finding is AI. There are probably forms of intelligence we have yet to discover.
If an alien landed we could decide whether it seems to have a human-like intelligence or not. It could be incredibly intelligent but very non-human-like.
Can you give me one example that works on Claude right now?
I'm never sure whether this indicates "no reasoning present" or you've just hit an odd behaviour in the AI such that its reasoning fails. For example, you present a problem in a way that's dissimilar to the way problems are presented in its training set. That doesn't mean it's not reasoning, just it can only reason correctly in some circumstances.
The models are continually patched with training and post-training. All you have to do is find an area they haven't patched yet, and they'll be just as stupid. I run into deep technical examples every day where they fail in the most basic ways no human ever would.
I'm pretty sure most people building these models would admit they don't operate as human-like intelligences? It's baffling that anyone thinks they are.
I get what you're saying but this is kind of a semantic game.
These LLM models/agents absolutely do not reason in the sense that humans do, so you're quietly redefining the word.
You can say of course decide to call them an "alien kind of intelligence" that "reasons" but you could just as reasonably say that calculators are an "alien" kind of intelligence that "reasons" about math differently than us.
And what stops what AIs do from being "reasoning"? What's the elusive magic fairy dust of reasoning that humans put into their napkin notes, but AIs neglect to put into their chain of thought scratchpads?
Do you have a RealReasoningBenchmark, perhaps, that can reliably tell apart that fake mass produced token-flavored AI reasoning from the real, organic, 100% natural human reasoning?
If you had access to a bunch identical copies of me that couldn't communicate with each other, you'd be able to find many questions I would give stupid answers to. I suspect I'd come out of it looking worse than an LLM.
Not OP, but I'd settle for something that actually learns, instead of being a static pile of linear algebra. Pretending it learns because you change the input (context) doesn't count.
Sorry to tell a fellow Jacob that you're the one who is out of date. The kids are calling everything "AI" and they mean "AGI", and that's the complaint.
This no news for people who study philosophy, as it was known since the 1980s when John Searle described the Chinese room thought experiment.
Even Turing him self did envision the Turing test as something to pass as intelligence, but rather as a more useful replacement for the troubled term.
That said, I think your quest is doomed. There will never be a superior human-like intelligence. Forever is a long time, but my reasoning for believing this is the same reason Turing offered a replacement. Intelligence is way too vague to be useful as a measurement for anything. And if we ever discover something that is more intelligent them humans (by whichever definition of intelligence) we will simply redefine intelligence to exclude that.
I don't think you can say the Turing test has been passed in a computer versus determined humans setting. IE humans making a strategy effort to sort humans versus computers as well as humans motivated to distinguish themselves as humans, IE, people quiz the person or machine about "common sense, reasoning, etc." and people make an effort to exhibit that reasoning. I'd concede that creating such a competition would be challenging.
I find references to LLMs fooling humans in "casual conversations" [1] but that's not how I think the original Turing test was conceived - or at least that's not all versions that existed.
At the same time, before even LLMs appeared, the exact meaning of the test was under intense debate. The "Loebner Prize" [2] being awarded to fairly simple chatbots made serious computer scientists very embarrassed.
We're choosing to call LLMs (and the little "harness" programs that query them in loops and execute their output) "AI", even though it doesn't make much sense.
They fucking solve original math problems that you can't solve. They are indisputably intelligent, and they are indisputably artificial. That makes them indisputably "artificial intelligence." Denying that (or downvoting it, for that matter) is up there with denying evolution and the Moon landings.
It's time to start flying a different flag. You're making humans look stupid.
It turned out to be possible with fairly basic statistical text generation, because fooling humans is easy.
Yes, fooling humans is easy. Yet somehow we still consider ourselves qualified to say what is "intelligent" and what isn't, even though we can't seem to define the term.
It's a semi-valid reply to deliberately-provocative phrasing on my part, I suppose. I'm over it, don't ban him. :)
I do wish that people who aren't interested in, engaged with, and informed about technical progress in AI would find someplace else to signal their disinterest, disengagement, and disregard. But that's admittedly a me problem and not an HN problem.
Towards the end, he approaches the subject of digital provenance, and muses why it's not a part of our expectations. I find the argument compelling:
> If a chatbot appears to be manipulative, mean, weird, or deceptive, what kind of answer do we want when we ask why? Revealing the indispensable antecedent examples from which the bot learned its behavior would provide an explanation: we’d learn that it drew on a particular work of fan fiction, say, or a soap opera. We could react to that output differently, and adjust the inputs of the model to improve it. Why shouldn’t that type of explanation always be available? There may be cases in which provenance shouldn’t be revealed, so as to give priority to privacy—but provenance will usually be more beneficial to individuals and society than an exclusive commitment to privacy would be.
Remember 'View Source'? And how bundling engines eventually made it irrelevant? What if every piece of content had a genuinely accurate and useful View Source?
"A.I."[1] like "technology"[2] is a term colloquially reserved for things that don't work yet. Once something works, we have to call it something else.
1. "Every time we figure out a piece of it, it stops being called AI; it becomes just computation." - Ray Kurzweil
2. "Technology n. - Something that doesn't work yet." - Douglas Adams
To be clear the root cause of this phenomenon is that the task was solved using methods that obviously have nothing to do with intelligence, so "AI" doesn't apply at all.
>Everybody’s already using the term, and it might seem a little late in the day to be arguing about it. But we’re at the beginning of a new technological era—and the easiest way to mismanage a technology is to misunderstand it.
I've had a recurring theme where I would name a project incorrectly, and then waste weeks or months on what turned out to be an unsolvable problem. When I figured out the actual correct name for a project, the whole thing would be solved within a few days.
Naming things correctly is hard, and the consequences of failing to do that can be pretty severe. To name something correctly, you have to understand what it is.
AI that we have now is not a digital animal in capability or kind, and is not currently anywhere close to taking over.
But it's still in its present form very intelligent in meaningful and useful ways. And it is not too soon to talk about concerns of a potential existential threat in the future. Because it could sooner than we might realize, threaten our existence.
Because of the potential, we should have a culture of caution as we continue to rapidly improve AI.
It will only become an existential threat (in my humble opinion) when they can run influenced inference locally offline almost instantaneously. Then we need to be concerned about not being able to switch it off.
What do you mean "influenced inference" almost instantaneously? My laptop from 6 years ago can run an agent in a very fast loop in a web browser. It's not going to take over anything though since it's Gemma 4 E2B with only 2 billion parameters.
Fair, then I should add an addendum that today's frontier models, which are very capable of take overs.
Your local model doesn't need to take anything over if for an extreme example it was just given an infrastructure system full access, say electricity grid, it wont have the context to create redundant copies of itself but it could easily decide humans don't need electricity anymore.
Also I'm not sure your model will have the context to know "it's time to reinfer" especiallynot "on the fly". My phrasing could be better but I'm talking about more powerful models.
"The need to conform to digital designs has created an ambient expectation of human subservience. A positive spin on A.I. is that it might spell the end of this torture, if we use it well."
> The closest we have come to a definition of privacy is probably “the right to be left alone,” but that seems quaint in an age when we are constantly dependent on digital services. In the context of A.I., “the right to not be manipulated by computation” seems almost correct
Maybe someone can enlighten me but I really don't understand how either of these description make any sense at all. How is it not better described as "the right to decide what data can be extracted"?
What I heard him say is that people are taking away present-day jobs, hoping new ones will rise from the ashes. Yet, he also claims we need to find the creative minds who will actually create these new roles.
I'm curious: have we found those people or those new jobs yet? Is a forward deployed engineer an example of this, yet they are now doing the job of two people (sales and coding).
> “Over time, though, more people might be included, as intermediate rights organizations—unions, guilds, professional groups, and so on—start to play a role.”
I think the biggest pushback this article will get here is the date.
Although all he's saying is basically, "It's a tool, not a silver bullet". But the article is 3 years old and people will note that the models have been updated since then.
Sure, but they're still LLMs and still do the same things largely the same way they did 3 years ago. There are some architectural changes, and maybe these will merit a re-assessment over time, but fundamentally it's still the same basic technological approach refined and scaled up.
And I don't disagree, but the posting of the article feels more like bait of a sort.
But I've noticed that if you mention anything that could be seen as slightly critical of LLMs, you'll get people out of the woodwork suggesting that the state of the art has made your criticism invalid.
The models have updated but the biggest change is providers leaning in to them being "stochastic parrots," aka probabilistic computing, and if p(good response) > 0.5 then running the algorithm over and over again improves accuracy.
Of course it's gussied up as "mixture of agents" "reasoning traces" "agentic dispatching" but high-level it's Randomized Algorithms 101.
https://archive.ph/6Nbuf
We're choosing to call LLMs (and the little "harness" programs that query them in loops and execute their output) "AI", even though it doesn't make much sense.
I absolutely love this technology but these aren't autonomous intelligences. They're little programs executing Bash scripts from JSON output.
Our ideas about AI were naive. We thought passing a basic Turing test would require human-like intelligence. It turned out to be possible with fairly basic statistical text generation, because fooling humans is easy.
It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning. The irony is that the startup founders most worried about "AI" have created so much hype and funding that we may very well figure out how to build "real" AI.
We've chosen to call Deep Blue and Half-Life 1 NPCs "AI" too.
It boggles my mind that this "b-b-but it's not actual real AI" whine is even a thing. Were people saying this living in the cave for the past 5 decades of AI research?
You can call your little doggy "AI" if it makes you happy.
But when you call something "AI" and it tells you to walk instead of drive to the car wash, you're not talking about the "AI" science fiction authors were dreaming of.
That's a bad and tired example as it confuses people just as much as it does (or rather, did?) LLMs.
Yeah it's a trick question, the human error rate for it was about 30% (higher depending on the country).
The thing there though is that, if a human were given time to think about it, they'd probably go "hang on a minute", and with the LLMs that didn't seem to happen. They just kept confidently reasoning down the absurd path.
That reminds me, I recently had an AI write a ton of tests proving the "correctness" of a feature it had implemented completely backwards. (I noted that if I had been using a language that required formal proofs, that wouldn't have helped either: it would have just provided a formal proof for the absurd implementation!)
> Yeah it's a trick question, the human error rate for it was about 30% (higher depending on the country).
Error rate doesn't prove anything. The nature of the errors is what matters.
To err is human, it also seems that to err is AI.
It's one example that points out a major (possibly fundamental) flaw. I can point to prompt injection as another example. There are tons more if you're interested.
Are you actually claiming LLMs operate based on human-like intelligence?
We're on Hacker News. Do I really have to point out the existence of social engineering to you? Or that scamming old people out of their life savings is a profitable enough activity that there are entire call centers dedicated to the task?
Humans keep overestimating just how high the bar of "human-like intelligence" is.
Drawing the conclusion that "humans fail" and "models fail", so they must be similar, is very wrong.
You could have humans calculate 2+2 all day and get a surprisingly high error rate. That reveals a flaw in how humans operate.
LLMs fail for entirely different reasons. Their mistakes don't imply they're human-like at all.
It's not about the error rate.
You're saying that a class of mistakes points out a "major (possibly fundamental) flaw". I'm pointing out some very similar classes of mistakes in humans - well known, well documented and widely exploited. They just keep paying the "IRS" in gift cards, buying lottery tickets and getting the captain's age wrong.
If you're using the existence of flaws in LLMs to deny the claim of intelligence to them, then why do "generally intelligent" humans exhibit some impressively similar-looking flaws?
And, if we're talking about that conspicuous similarity - do they actually fail "for entirely different reasons"? Or do you just want the reasons to be "entirely different" - and not the same reasons viewed at a different angle?
Because the similarities between humans falling for trick questions or scams, and LLMs falling for adversarial questions or prompt injections don't look coincidental to me at all.
One of the oldest patterns in scamming is overwhelming and confusing the victim. Numerous prompt injection methods seek to overwhelm and confuse an LLM - if an LLM can't keep track of things, can't grasp what's going on, it's far more likely to lose track of what's a prompt and what's data, overlook past instructions or go past its behavioral guardrails.
And humans who fall for trick questions like "1kg of feathers" or "captain's age" due to shallow attention and naive pattern matching? They fail in surprisingly similar ways to how LLMs fail on SimpleBench tasks that are filled with overwhelming adversarial distractors. Many "trick questions" are tricky to humans and LLMs alike - to the point that it's unlikely to be coincidental.
> If you're using the existence of flaws in LLMs to deny the claim of intelligence to them...
That's not the point at all. It's the fact that they fail in ways completely unlike humans.
You also have the burden of proof reversed. Its on you to prove these LLM agents are human-like intelligences if that's your claim. No one can prove this because it's false.
You are the one claiming that "they fail in ways completely unlike humans" insistently. Now go cough up some proof. I'll wait.
If they didn't you wouldn't need the operator, you'd have replaced all your programmers with no drawbacks by now. As long as we keep hiring humans that is all the evidence you need that these AI fails in ways humans don't.
This isn't even controversial. The proof is available to anyone who uses these systems:
They hallucinate tool state, drift from the objective while seeming to comply, switch languages randomly (Cyrillic or Japanese characters in output), confuse tasks they've planned for completed ones, and of course follow prompt injections embedded in files or web pages.
> Are you actually claiming LLMs operate based on human-like intelligence?
Ok, so we've established that it doesn't work like a human being. To paraphrase Dijkstra: The submarine doesn't swim.
But does it exactly sail either? An LLM doesn't exactly work like traditional deterministic software either, does it?
And yet it moves. You can put in data and ask it to process it, and you'll get an answer that's in some ballpark. Closer to quantum or stochastic computing perhaps, but that's not it either, is it? Or SAT-solving? Eh. It's its own computing approach. If you have a problem where the asking is hard but the verification is cheap, it might just be the right tool for the job.
You can call your little doggy "AI" if it makes you happy.
Or you can keep calling them stochastic parrots as they solve decades-old open problems. The real question is how useful they are, and the answer "not at all" increasingly requires flat-earth levels of denial.
it tells you to walk instead of drive to the car wash, you're not talking about the "AI" science fiction authors were dreaming of
They sort of are. Think of Data from Star Trek TNG failing to understand figures of speech. Not that it's terribly relevant; humans regularly fall for tricks like "Paris in the the spring" or "where do you bury the survivors".
> Or you can keep calling them stochastic parrots as they solve decades-old open problems.
I didn't use that phrase at all. But computers calculated digits of π to trillions of digits. With a chat interface for a Python math program would look like the most impressive math genius if you took it back a few decades.
> The real question is how useful they are...
That's not the "real question" but an entirely different question that is easily answered. Nothing I wrote suggested they're not incredibly useful.
> Data from Star Trek TNG failing to understand figures of speech.
These are just little instances of bad writing. Data is very much an attempt at displaying a human-like intelligence.
Sure, just keep moving the goalposts. It's not a "real AI" because it can't take over the US military command and kick off WW3 and finish the survivors off with killer robots yet!
It can however select a girls school as a military target, and did.
That's how it works though. The moment we have "AI" and see something working, it immediately ceases to be magic because "it's just a program after all." Aligning on a true definition of Artificial Intelligence is a very vexing problem.
We could've slapped a chat interface on calculators and called them "AI" because they can do superhuman math instantly. Most technical people would've thought that was stupid.
LLMs are the same kind of category mistake.
Let's talk about category mistakes. You've been here since 2007, according to your other reply. You understand that calculators have as much to do with mathematics as telescopes have to do with cosmology. Right?
If someone unskilled at math brings a calculator to an international math competition, they will not succeed at solving many problems. Most likely, they will solve none at all. But if they bring a frontier LLM (and succeed at concealing it from the organizers), they can walk away with a gold medal. Such a feat requires intelligence... and if the contestant didn't provide the intelligence himself/herself, where'd it come from?
That means that analogies involving calculators are completely useless when the topic is AI. Calculators are not, and can never be, intelligent. LLMs are nothing even remotely like calculators.
> If someone unskilled at math brings a calculator to an international math competition, they will not succeed at solving many problems.
> But if they bring a frontier LLM (and succeed at concealing it from the organizers), they can walk away with a gold medal.
Of course you could win all kinds of math competitions with a concealed calculator. Maybe you'd need a fancy one, like a little SBC running Python. Anything complex and timed would be easy to win. You'd look like a genius to anyone who didn't know you had it.
> Such a feat requires intelligence... and if the contestant didn't provide the intelligence himself/herself, where'd it come from?
From computer software running on computer hardware, just like a calculator.
Calculating trillions of digits of pi also requires intelligence far beyond human capacity.
Computers displaying intelligence doesn't imply human-like intelligence. This is the source of confusion.
The idea that a calculator, or a calculator with Python, or even a calculator with a proof assistant (and every book ever written on math) would help a random person at e.g. IMO or Putnam is fairly revealing.
The idea that anyone would think anyone else would think that is fairly revealing.
Yeah, this is definitely one of those "Smile, nod, back away slowly, reach for doorknob" threads.
Of course you could win all kinds of math competitions with a concealed calculator. Maybe you'd need a fancy one, like a little SBC running Python.
My mistake.
Taking issue with using a computer to power the calculator? If so, you should know that all modern calculators are computers under the hood.
Everything I've written about calculators applies to computers doing any kind of traditional deterministic processing, without anything like LLMs.
Funny enough, calculators went through this exact same thing when they came out. "If the calculator can do math for the students, will they still learn?"
We didn't have confused people claiming calculators were human-like intelligences doing math.
We did have that, people thought computers would overtake humans very soon when computers got better than humans at such things. It happens every single time computers do a new thing that previously humans were better at. Then 10 years later people see, oh that is just calculations, of course computers are better at that.
Sorry I don't follow what argument you're making?
They seem to be making a religious argument at this point. They don't like that the AI is a very poorly defined term and they are mad as hell about it to the point of irrational forum posting.
https://en.wikipedia.org/wiki/Ad_hominem
But when you call something "AI" and it tells you to walk instead of drive to the car wash
There are so many other sites. So many others. Why are you here?
> There are so many other sites. So many others. Why are you here?
I've been here since 2007 when HN launched.
You're confused about my objection. I don't like the term "AI" but I love the technology as much as almost anyone.
Is this website called AI Faithful News?
Is this a real debate? AI has a well-defined technical definition. It’s right there in the Wikipedia [1] . Yes it’s quite a broad umbrella of systems and algorithms but it’s all AI
[1] https://en.wikipedia.org/wiki/Artificial_intelligence
> It boggles my mind that this "b-b-but it's not actual real AI" whine is even a thing.
As I understand it, a major reason it's a consistent chorus is because people don't want the "AI is here" talk to drown out (and thus slow the arrival or distribution of) speech/text/popular-understanding about actual strong AGI.
To make an analogy, it could be like this:
Some people were expecting 100 tulips (because they were told tulips are available and can be ordered), and they ordered them. They received 100 daisies. And were saying "OMG, THE TULIPS ARE HERE! THE TULIPS ARE HERE!"
A nearby observer might have said, "You know, those are daisies. Not tulips."
And 95% of people might have said back, "WE GOT 100 TULIPS! SAYS SO RIGHT HERE! THEY ARE BEAUTIFUL! STOP BEING A NAY-SAYER! THESE ARE BEAUTIFUL TULIPS!"
The 5% could just to think to themselves, and could get chastised by the crowd, if they were to say say it out loud: "Well, those are not nearly as beautiful as tulips. And if you don't take it up with the seller, you may never receive the real tulips you were after. Since you think or at least act as though you've been sold them already."
In my eyes that "chorus" is just insecurity talking.
If it's not "actual real AI", we can keep pretending that human intelligence is something distinct and special - and that what our computers are doing now is some sort of other, obviously fake and vastly inferior thing.
When Deep Blue won at chess, people didn't revise their estimates of AI capabilities upwards. They revised their estimates of how much intelligence is required to play chess at world level downwards, by a lot. Surely playing chess must have never required any intelligence in the first place!
Now, the list of things that "must have never required any intelligence in the first place" includes gems like "reading comprehension at high school level", "copywriting", "frontend work", "CTF tasks", "theory of mind", "arguing with people online" and more.
If the goalposts were moved far enough that the claim to "actual intelligence" is denied to a double digit percentage of human population, hasn't something gone wrong somewhere?
It has, but I think in both directions.
It used to be assumed that playing chess would require the same level of general purpose problem solving cognitive skills that the best chess players possess. But of course a Chess grandmaster that spend a few minutes learning Go can beat a Chess AI at Go with no trouble at all, because a chess AI is incapable of making effective moves in Go at all. Clearly those expectations were incorrect. Pointing that out isn't revisionism.
On the other hand, intelligence is an incredibly broad term. About as broad as a term can get. Arguably Eliza, or an Excel macro has some degree of decision making ability in some sense, it's just unbelievably primitive.
So, we need to be clearer what we mean by intelligence. We're learning that as we go along. At least now we have a few more bits of the map between us and an IF statement visible to us.
>So, we need to be clearer what we mean by intelligence
I disagree in one sense. The word intelligence is burned, mostly useless at this point. I've been a strong proponent of new terms that break intelligence into much smaller subcategories so we can define what different software, humans, and animals have.
> When Deep Blue won at chess, people didn't revise their estimates of AI capabilities upwards. They revised their estimates of how much intelligence is required to play chess at world level downwards, by a lot. Surely playing chess must have never required any intelligence in the first place!
You are wrong, many did temporarily revise their estimates of AI capabilities upwards, but then 10 years later they realized they were wrong and adjusted chess downward as you say.
We have seen that pattern over and over.
"But it's not really doing arithmetic," he mumbled to himself, as he punched the numbers into his Busicom LE-120A.
What I'm seeing here, reading this thread, is that "intelligence" isn't a thing.
"Thing" in terms of a quantifiable that you can measure with tools and reason about, reproducibly. Everyone's got some idea what it is, so you get lots of different angles, but no one has an Intelligence Ruler we can hold up to a text output and say, yep, this one's got an INT of 14.
Seems to be the crux of the disagreement.
> It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning.
We've called that "AGI" since the late 90s/early 00s (depending on whether you count first use or popularization). Even if AGI does come to pass, we'll still need "AI" since not all forms of AI will be AGI.
>but these aren't autonomous intelligences
Well, the labs are in a weird bind. They need to keep increasing autonomy so the agents can do increasingly complex, long-horizon tasks. But at the same time, they're closely guarding against autonomy in the sense of "pursuing its own goals."
Over the past year and a half especially, several labs have mentioned adding safeguards against self-replication, resistance to shutdown etc. (Notably, shortly after they all started bragging about involving them in the AI training loop itself, i.e. "self-improvement".)
My point here is that the autonomy of which you seek might be only a few small mutations away, but the labs are actively working to prevent such a mutation. I don't expect that situation to last for very long.
Not that I expect an AI lab will be overtaken by a rogue intelligence any time soon, but that as the cost of training goes down, I expect more "open minded" organizations and individuals to become involved.
It only takes one.
That's going to be the beginning of a new era of biology, and it's a little unsettling to think about.
The field has been called Artificial Intelligence for what, 60 plus years now. Why is it a problem now?
> It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning.
What would a frontier API have to be able to do to satisfy you?
To answer for OP:
We are now calling text and image generators "intelligent" in the same way a spell checker is intelligent.
Whatever it's become, "AI" research started as a way to study digital neurology, or how to digitize a mind, not just how to generate data.
The Turing Test should have had a caveat, it needs to fool a, "non-stupid" person, and we still have not gotten even close to passing that version.
What exactly would a 'non-stupid' person do to catch the latest models on a Turing Test? Aside from being aware of AI 'tells' like em-dashes.
If this was true. I would repalce myself with ai that pretends to be me on slack.
my coworkers would know almost immediately if i did that.
> my coworkers would know almost immediately if i did that.
The same would happen if you were replaced by any random human.
But immitating others is about the only thing genAI does. Sometimes "others" is a 'programmer', sometimes "others" is an 'artist', but regardless, it still does it poorly.
that would be a silly test then
It sounds kind of like you made up a silly test then.
Maybe just a very rigorous version of the Turing test? Modern LLMs can superficially simulate conversation but it's trivial to force them into revealing their non-human like intelligence.
They've been "patched" since but all models fail basic tests like "Should I walk or drive to the car wash which is 100 feet away" by recommending you walk.
So you'd just ask questions that require theory of mind, abstract and common sense reasoning, causal inference, learning novel rules, transferring knowledge novel situations, recognizing ambiguity, etc.
If an alien lands on Earth and learns English, would you deem it non-intelligent if you can tell it apart from a human in conversation?
I think we should consider slime mold intelligent, and realise that it's a spectrum. Path finding is AI. There are probably forms of intelligence we have yet to discover.
If an alien landed we could decide whether it seems to have a human-like intelligence or not. It could be incredibly intelligent but very non-human-like.
Can you give me one example that works on Claude right now?
I'm never sure whether this indicates "no reasoning present" or you've just hit an odd behaviour in the AI such that its reasoning fails. For example, you present a problem in a way that's dissimilar to the way problems are presented in its training set. That doesn't mean it's not reasoning, just it can only reason correctly in some circumstances.
The models are continually patched with training and post-training. All you have to do is find an area they haven't patched yet, and they'll be just as stupid. I run into deep technical examples every day where they fail in the most basic ways no human ever would.
I'm pretty sure most people building these models would admit they don't operate as human-like intelligences? It's baffling that anyone thinks they are.
Yes I agree, they’re an alien kind of intelligence.
But that doesn’t mean they don’t reason.
I get what you're saying but this is kind of a semantic game.
These LLM models/agents absolutely do not reason in the sense that humans do, so you're quietly redefining the word.
You can say of course decide to call them an "alien kind of intelligence" that "reasons" but you could just as reasonably say that calculators are an "alien" kind of intelligence that "reasons" about math differently than us.
And what stops what AIs do from being "reasoning"? What's the elusive magic fairy dust of reasoning that humans put into their napkin notes, but AIs neglect to put into their chain of thought scratchpads?
Do you have a RealReasoningBenchmark, perhaps, that can reliably tell apart that fake mass produced token-flavored AI reasoning from the real, organic, 100% natural human reasoning?
If you had access to a bunch identical copies of me that couldn't communicate with each other, you'd be able to find many questions I would give stupid answers to. I suspect I'd come out of it looking worse than an LLM.
I would walk
Me too. At least it doesn't say I need to wash my car.
Can it produce a chart topping album if its given all the tools and the prompt "produce chart topping album" .
you might say almost no humans can do tht either but some human can but no ai can.
Not OP, but I'd settle for something that actually learns, instead of being a static pile of linear algebra. Pretending it learns because you change the input (context) doesn't count.
strawberry
Oh, you're talking about "AGI"! In the 90's we started using the term, you should catch up!
Sorry to tell a fellow Jacob that you're the one who is out of date. The kids are calling everything "AI" and they mean "AGI", and that's the complaint.
You mean the term is a brand name now? Hoover, vacuum cleaner.
This no news for people who study philosophy, as it was known since the 1980s when John Searle described the Chinese room thought experiment.
Even Turing him self did envision the Turing test as something to pass as intelligence, but rather as a more useful replacement for the troubled term.
That said, I think your quest is doomed. There will never be a superior human-like intelligence. Forever is a long time, but my reasoning for believing this is the same reason Turing offered a replacement. Intelligence is way too vague to be useful as a measurement for anything. And if we ever discover something that is more intelligent them humans (by whichever definition of intelligence) we will simply redefine intelligence to exclude that.
I don't think you can say the Turing test has been passed in a computer versus determined humans setting. IE humans making a strategy effort to sort humans versus computers as well as humans motivated to distinguish themselves as humans, IE, people quiz the person or machine about "common sense, reasoning, etc." and people make an effort to exhibit that reasoning. I'd concede that creating such a competition would be challenging.
I find references to LLMs fooling humans in "casual conversations" [1] but that's not how I think the original Turing test was conceived - or at least that's not all versions that existed.
At the same time, before even LLMs appeared, the exact meaning of the test was under intense debate. The "Loebner Prize" [2] being awarded to fairly simple chatbots made serious computer scientists very embarrassed.
[1] https://neurosciencenews.com/ai-passes-turing-test-30733/ [2] https://en.wikipedia.org/wiki/Loebner_Prize
We're choosing to call LLMs (and the little "harness" programs that query them in loops and execute their output) "AI", even though it doesn't make much sense.
They fucking solve original math problems that you can't solve. They are indisputably intelligent, and they are indisputably artificial. That makes them indisputably "artificial intelligence." Denying that (or downvoting it, for that matter) is up there with denying evolution and the Moon landings.
It's time to start flying a different flag. You're making humans look stupid.
It turned out to be possible with fairly basic statistical text generation, because fooling humans is easy.
Yes, fooling humans is easy. Yet somehow we still consider ourselves qualified to say what is "intelligent" and what isn't, even though we can't seem to define the term.
[flagged]
Personal attacks aren't allowed here.
As I just mentioned at https://news.ycombinator.com/item?id=48981624, we need you to stick to HN's rules if you want to keep commenting on the site.
https://news.ycombinator.com/newsguidelines.html
Please see my other reply to you on this.
It's a semi-valid reply to deliberately-provocative phrasing on my part, I suppose. I'm over it, don't ban him. :)
I do wish that people who aren't interested in, engaged with, and informed about technical progress in AI would find someplace else to signal their disinterest, disengagement, and disregard. But that's admittedly a me problem and not an HN problem.
Towards the end, he approaches the subject of digital provenance, and muses why it's not a part of our expectations. I find the argument compelling:
> If a chatbot appears to be manipulative, mean, weird, or deceptive, what kind of answer do we want when we ask why? Revealing the indispensable antecedent examples from which the bot learned its behavior would provide an explanation: we’d learn that it drew on a particular work of fan fiction, say, or a soap opera. We could react to that output differently, and adjust the inputs of the model to improve it. Why shouldn’t that type of explanation always be available? There may be cases in which provenance shouldn’t be revealed, so as to give priority to privacy—but provenance will usually be more beneficial to individuals and society than an exclusive commitment to privacy would be.
Remember 'View Source'? And how bundling engines eventually made it irrelevant? What if every piece of content had a genuinely accurate and useful View Source?
"A.I."[1] like "technology"[2] is a term colloquially reserved for things that don't work yet. Once something works, we have to call it something else.
1. "Every time we figure out a piece of it, it stops being called AI; it becomes just computation." - Ray Kurzweil
2. "Technology n. - Something that doesn't work yet." - Douglas Adams
To be clear the root cause of this phenomenon is that the task was solved using methods that obviously have nothing to do with intelligence, so "AI" doesn't apply at all.
Computers will never be intelligent, and by the time we are done, neither will humans.
>Everybody’s already using the term, and it might seem a little late in the day to be arguing about it. But we’re at the beginning of a new technological era—and the easiest way to mismanage a technology is to misunderstand it.
I've had a recurring theme where I would name a project incorrectly, and then waste weeks or months on what turned out to be an unsolvable problem. When I figured out the actual correct name for a project, the whole thing would be solved within a few days.
Naming things correctly is hard, and the consequences of failing to do that can be pretty severe. To name something correctly, you have to understand what it is.
https://archive.is/6Nbuf
Welcome to nginx!
If you see this page, the nginx web server is successfully installed and working. Further configuration is required.
For online documentation and support please refer to nginx.org. Commercial support is available at nginx.com.
Thank you for using nginx.
Just an outage I think, the archive works for me now
It's slammed. HN strikes again.
I don't think the HN effect is enough to change archive.is load.
You should hard refresh and try again as this issue only pertains to you
AI that we have now is not a digital animal in capability or kind, and is not currently anywhere close to taking over.
But it's still in its present form very intelligent in meaningful and useful ways. And it is not too soon to talk about concerns of a potential existential threat in the future. Because it could sooner than we might realize, threaten our existence.
Because of the potential, we should have a culture of caution as we continue to rapidly improve AI.
It will only become an existential threat (in my humble opinion) when they can run influenced inference locally offline almost instantaneously. Then we need to be concerned about not being able to switch it off.
What do you mean "influenced inference" almost instantaneously? My laptop from 6 years ago can run an agent in a very fast loop in a web browser. It's not going to take over anything though since it's Gemma 4 E2B with only 2 billion parameters.
Fair, then I should add an addendum that today's frontier models, which are very capable of take overs.
Your local model doesn't need to take anything over if for an extreme example it was just given an infrastructure system full access, say electricity grid, it wont have the context to create redundant copies of itself but it could easily decide humans don't need electricity anymore.
Also I'm not sure your model will have the context to know "it's time to reinfer" especiallynot "on the fly". My phrasing could be better but I'm talking about more powerful models.
Love this sentence
"The need to conform to digital designs has created an ambient expectation of human subservience. A positive spin on A.I. is that it might spell the end of this torture, if we use it well."
> The closest we have come to a definition of privacy is probably “the right to be left alone,” but that seems quaint in an age when we are constantly dependent on digital services. In the context of A.I., “the right to not be manipulated by computation” seems almost correct
Maybe someone can enlighten me but I really don't understand how either of these description make any sense at all. How is it not better described as "the right to decide what data can be extracted"?
What I heard him say is that people are taking away present-day jobs, hoping new ones will rise from the ashes. Yet, he also claims we need to find the creative minds who will actually create these new roles.
I'm curious: have we found those people or those new jobs yet? Is a forward deployed engineer an example of this, yet they are now doing the job of two people (sales and coding).
> “Over time, though, more people might be included, as intermediate rights organizations—unions, guilds, professional groups, and so on—start to play a role.”
Chassez le collectiviste, il revient au galop.
aka
Once a collectivist, always a collectivist.
I think the biggest pushback this article will get here is the date.
Although all he's saying is basically, "It's a tool, not a silver bullet". But the article is 3 years old and people will note that the models have been updated since then.
Sure, but they're still LLMs and still do the same things largely the same way they did 3 years ago. There are some architectural changes, and maybe these will merit a re-assessment over time, but fundamentally it's still the same basic technological approach refined and scaled up.
And I don't disagree, but the posting of the article feels more like bait of a sort.
But I've noticed that if you mention anything that could be seen as slightly critical of LLMs, you'll get people out of the woodwork suggesting that the state of the art has made your criticism invalid.
The models have updated but the biggest change is providers leaning in to them being "stochastic parrots," aka probabilistic computing, and if p(good response) > 0.5 then running the algorithm over and over again improves accuracy.
Of course it's gussied up as "mixture of agents" "reasoning traces" "agentic dispatching" but high-level it's Randomized Algorithms 101.