LLMs are solving MCAT, the bar test, SAT etc like they’re nothing. At this point their performance is super human. However they’ll often trip on super simple common sense questions, they’ll struggle with creative thinking.
Is this literally proof that standard tests are not a good measure of intelligence?
Citation needed that LLMs are passing these tests like they’re nothing.
LLMs don’t have intelligence, they are sentence generators. Sometimes those sentences are correct, sometimes they’re gobbledygook.
For instance, they fabricate real-looking but nevertheless totally fake citations in research papers https://www.nature.com/articles/s41598-023-41032-5
To your point we already know standardized tests are biased and poor tools to measure intelligence. Partly that’s because they don’t actually measure intelligence- they often measure rote knowledge. We don’t need LLMs to make that determination, we already can.
Talked about this a few times over the last few weeks but here we go again…
I am teaching myself to write and had been using chatgpt for super basic grammar assistance. Seemed like an ideal thing, toss a sentence I was iffy about into it and ask it what it thought. After all I wasn’t going to be asking it some college level shit. A few days ago I asked it about something I was questionable on. I honestly can’t remember the details but it completely ignored the part of the sentence I wasn’t sure about and told me something else was wrong. What it said was wrong was just…not wrong. The ‘correction’ it gave me was some shit a third grader would look at and say ‘uhhhhh…I’m gonna ask someone else now…’
That’s because LLMs aren’t intelligent. They’re just parrots that repeat what they’ve heard before. This stuff being sold as an “AI” with any “intelligence” is extremely misleading and causing people to think it’s going to be able to do things it can’t.
Case in point, you were using it and trusting it until it became very obvious it was wrong. How many people never get to that point? How much has it done wrong before then? Etc.
OP picked standardized tests that only require memorization because they have zero idea what a real IQ test like the WAIS is like.
Also how those IQ tests work. You kind of have to go in “blind” to get an accurate result. And LLM can’t do anything “blind” because you have to train them.
A chatbots can’t even take a real IQ test, if we trained a chatbots to take a real IQ test, it would be a pointless test
Actually, you can give chatbots a real IQ test, and the range of scores fall into roughly the same spread as how they rank on other measures, with the leading model scoring at 100:
https://www.maximumtruth.org/p/ais-ranked-by-iq-ai-passes-100-iq
Nobody is a blank slate. Everyone has knowledge from their past experience, and instincts from their genetics. AIs are the same. They are trained on various things just as humans have experienced various things, but they can be just as blind as each other on the contents of the test.
No, they wouldn’t.
Because real IQ tests arent just multiple choice exams
You would have to train it to handle the different tasks, and training it at the tasks would make it better at the tasks, raising their scores.
I don’t know if the issue is you don’t know about how IQ tests work, or what LLM can do.
But it’s probably both instead of one or the other.
You’re entirely missing the point.
The requirements and basis of IQ tests are they are problems you haven’t seen before. An LLM works by recognizing existing data and returning what came next in the training set.
LLMs work directly in opposition of how an IQ text works.
Things like past experience are all the shit IQ tests need to avoid in order to be accurate. And they’re exactly what LLMs work off of.
By definition, LLMs have no IQ.
Standard tests don’t measure intelligence. They measure things like knowledge and skill. And ChatGPT is very knowledgeable and highly skilled.
IQ tests have the goal of measuring intelligence.
just a reminder that IQ tests may have the goal of measuring intelligence, but that says nothing of their precision and accuracy
Exactly. I chose my words very carefully.
IQ tests have the goal of measuring intelligence.
The range of LLM scores on IQ tests:
https://www.maximumtruth.org/p/ais-ranked-by-iq-ai-passes-100-iq
Yep, very knowledgeable, highly skilled, kind of a dumbass.
You meet enough doctors and lawyers and you tend to find that combination unremarkable.
All standardized test is how well you prepared for that particular standardized test, doesn’t matter if it is the SAT, MCAT, or Leetcode. You aren’t suppose to think on the spot for these tests, you are suppose regurgitate everything you have rehearsed for weeks and months during the test.
And unthinking regurgitation is what LLMs do better than anything else.
I would argue that some code test questions can be solved spontaneously, but they are limited to easy to some early medium questions, or patterns that are common enough.
I guess this is more common in non FANG companies that don’t have to filter out candidates just because of the sheer number alone.
As someone that didn’t really have good coaching on the SAT, I 100% agree. I kinda fucked it up, but at 17, I wasn’t really used to studying for things outside of school and my parents didn’t get me into any study classes
For GRE though, I studied my ass off… got top 96 percentile scores.
Also went through the leetcode grind. Bombed the first job search I ever did and then later aced the hell out of it after studying really hard.
These tests are all about how diligently you studied and your study technique.
There has been plenty of proof that standardized testing doesn’t work long before ChatGTP ever existed. Institutions will keep using them though because that’s what they’ve always done and change is hard
Long before. Even in 1930 the eugenics-motivated creator Carl Brigham recanted his original conclusions only years ago that had led to the development of the SAT, but by then the colleges had totally invested in a quick and easy way to score students, even if it was inaccurate. Change is hard, but I think the bigger influence here was money since it hadn’t been around that long at that point.
Not disagreeing with you; how do you suggest a way for admissions to reliably compare applicants with each other? A 3.5 at one school can mean something completely different than a 3.5 at another school.
Something like the SAT is far from perfect, but it is a way one number that means the same thing across applicants.
I think this is the point, because Harvard got rid of the SAT requirement, and then just brought it back.
It’s a really terrible measure .
But it is an equal measure, despite what it measuring moderately meaningless.
I don’t think we have a better answer yet, because everything else lacks any sort of comparable equivalency .
And I say this as an ADHD sufferer who is at a huge disadvantage on standardised testing
There shouldn’t even be admission based on what you score in some random test. My (non-US) university accepted everyone who applied, at least for my field of study. Does that mean many people drop out after a semester or two? Absolutely, but there are countless people completing their studies who would have never gotten a chance to do so otherwise. Why shouldn’t they be allowed to prove themselves?
When I was at uni lecturers would often state that exams were thr worst measure of grasping the subject material but its all we have at the moment.
I saw this my self with some of my class mates testing very well but when discussing or problem solving outside of the class there was nothing there.
I think llms fall into this category but with way better recall.
When I was at uni lecturers would often state that exams were thr worst measure of grasping the subject material but its all we have at the moment.
It’s not all we have…
But it’s the only way a professor can run multiple classes of 100 students each.
But colleges are all about profit, so classes sizes are going to be huge.
The goal isn’t educating people, it’s making money.
So when they say “there’s no other option” they’re not mentioning the “and keep making as much money” at the end, it’s just implied.
I’m not in the us collages are generally vocational here with both colleges being less (while not totaly) concerned by the money side.
For example where I live university courses are free for those in country outside they pay fees
Dunno how it’s done elsewhere but our course are usually measured in 3 parts 1 exam 2 practical 3 essey/investigation. Everyone hates exams
It’s also the only way that is portable. A professor could evaluate each student, but has no way to transmit that kind of evaluation in a way that schools or employers across the country would trust. They didn’t know who the professor is, or what his standards are, or even if he is being bribed to pass somebody. (Which would happen much more if the professors opinion had the weight that the standardized test does. )
I had a lot of professors who put most of the grade weight on large projects. It made for a very heavy workload, but projects/ papers give a much better picture of how capable someone is of not only reciting knowledge, but also applying it.
Most of my grades were split 40/40/20
With the 3 being a writen component
A lot of good comments in this thread, but I’d like to add that to say ChatGPT is “not intelligent” is to ignore the hard work of all the stupid humans in the world.
Many humans spread and believe false information more often than ChatGPT. Some humans can’t even string together coherent sentences, and other humans will happily listen to and parrot those humans as though they were speaking divine truths. Many humans can’t do basic math and logic even after 12+ years of being taught it, over and over. Intelligence is a spectrum, and ChatGPT is definitively more intelligent than a non-zero number of humans. I’d love to figure out what that number is before I judge its standardized test performance.
LLMs don’t “think” at all. They string together words based on where those words generally appear in context with other words based on input from humans.
Though I do agree that the output from a moron is often worth less than the output from an LLM
That’s a common misunderstanding.
LLMs have billions of neurons, and we can see firsthand how information travels along their neural pathways and, yeah, it looks a whole lot like they’re thinking. If anything, we have more concrete proof that LLMs think than that humans think.
They do think, it’s just that they don’t have short term memory. They can only remember things linguistically, by talking and then listening to their own output. It’s an artifact of how we’ve set them up to interact with the world. Many humans use a similar thought process for certain problems (e.g. talking out loud to a rubber duck). Sure, there are other ways humans think too (e.g. visual/spatial), but linguistic thought is still valid.
This is kind of how humans operate as well though. We just string words along based on what input is given.
We speak much too fast to be properly reflecting on it, we just regurgitate whatever comes too mind.
To be clear, I’m not saying LLM think but that the difference between our thinking and their output isn’t the chasm it’s made out to be.
The key difference is that your thinking feeds into your word choice. You also know when to mack up and allow your brain to actually process.
LLMs are (very crudely) a lobotomised speech center. They can chatter and use words, but there is no support structure behind them. The only “knowledge” they have access to is embedded into their training data. Once that is done, they have no ability to “think” about it further. It’s a practical example of a “Chinese Room” and many of the same philosophical arguments apply.
I fully agree that this is an important step for a true AI. It’s just a fragment however. Just like 4 wheels, and 2 axles don’t make a car.
Apologies if this comes off as rude, but as an engineer involved in reinforcement learning, it’s upsetting when people make claims like this based on conjecture and hand-wavey understandings of ML. Some day there will be goal-driven agents that can interact with the world, and those agents will be harmed by those kinds of incorrect understandings of machine learning.
The key difference is that your thinking feeds into your word choice.
LLMs’ thinking also feeds into their word choice. Where else would they be getting the words from, thin air? No, it’s from billions of neurons doing what neurons do, thinking.
They can chatter and use words, but there is no support structure behind them.
What is a “support structure”, in your mind? That’s not a defined neuroscience, cog sci, or ML term, so it sounds to me like hand-waving.
The only “knowledge” they have access to is embedded into their training data.
LLMs can and do generalize beyond their training data, it’s literally the whole point. Otherwise, they’d be useless.
Once that is done, they have no ability to “think” about it further.
During training, neural weights from previous examples are revisited and recontextualized given the new information. This is what leads to generalization.
It’s a practical example of a “Chinese Room” and many of the same philosophical arguments apply.
The Chinese Room is not a valid argument, because the same logic can be applied to other humans besides yourself.
Disagree. We’re very good at using words to convey ideas. There’s no reason to believe that we speak much too fast to be properly reflecting on what we say—the speed with which we speak speaks to our proficiency with language, not a lack thereof. Many people do speak without reflecting on what they say, but to reduce all human speech down to that? Downright silly. I frequently spend seconds at a time looking for a word that has the exact meaning that will help to convey the thought that I’m trying to communicate. Yesterday, for example, I spent a whole 15 seconds or so trying to remember the word exacerbate.
An LLM is extremely good at stringing together stock words and phrases that make it sound like it’s conveying an idea, but it will never stop to think about the definition of a word that best conveys a real idea. This is the third draft of this comment. I’ve yet to see an LLM write, rewrite, then rewrite again it’s output.
Kinda the same thing though. You spent time finding the right auto-complete in your head. You weighed the words that fit the sentence you’d constructed in order to find the one most frequently encountered in conversations or documents that include specific related words. We’re much more sophisticated at this process, but our whole linguistic paradigm isn’t fundamentally very different from good auto-complete.
I’ve yet to see an LLM write, rewrite, then rewrite again it’s output.
It’s because we (ML peeps) literally prevent them from deleting their own ouput. It’d be like if we stuck you in a room, and only let you interact with the outside world using a keyboard that has no backspace.
Seriously, try it. Try writing your reply without using the delete button, or backspace, or the arrow keys, or the mouse. See how much better you do than an LLM.
It’s hard! To say that an LLM is not capable of thought just because it makes mistakes sometimes is to ignore the immense difficulty of the problem we’re asking it to solve.
To me it isn’t just the lack of an ability to delete it’s own inputs, I mean outputs, it’s the fact that they work by little more than pattern recognition. Contrast that with humans, who use pattern recognition as well as an understanding of their own ideas to find the words they want to use.
Man, it is super hard writing without hitting backspace or rewriting anything. Autocorrect helped a ton, but I hate the way this comment looks lmao
This isn’t to say that I don’t think a neural network can be conscious, or self aware, it’s just that I’m unconvinced that they can right now. That is, that they can be. I’m gonna start hitting backspace again after this paragraph
That was brilliant, thanks for actually giving it a try :D
It’s easy for me to get pedantic about minor details, so I should shut up and mention that I see what you mean and agree with the big picture. It’s not there yet and may someday be.
Thanks again, stranger! You made my day. Keep on being awesome
I think it highlights how a lot of these exams are just about the amount of information one can memorize.
Standardized tests were always a poor measure of comprehensive intelligence.
But this idea that “LLMs aren’t intelligent” popular on Lemmy is based on what seems to be a misinformed understanding of LLMs.
At this point there’s been multiple replications of the findings that transformers build world models abstracted from the training data and aren’t just relying on surface statistics.
The free version of ChatGPT (what I’m guessing most people have direct experience with) is several years old tech that is (and always has been) pretty dumb. But something like Claude 3 Opus is very advanced at critical thinking compared to GPT-3.5.
A lot of word problem examples that models ‘fail’ are evaluating the wrong thing. When you give a LLM a variation of a classic word problem, the frequency of the normal form biases the answer back towards it unless you take measures to break the token similarities. If you do that though, most modern models actually do get the variation completely correct.
So for example, if you ask it to get a vegetarian wolf, a carnivorous goat, and a cabbage across a river, even asking with standard prompt techniques it will mess up. But if you ask it to get a vegetarian 🐺, a carnivorous 🐐 and a 🥬 across, it will get it correct.
GPT-3.5 will always fail it, but GPT-4 and more advanced will get it correct. And recently I’ve started seeing models get it correct even without the variation and trip up less with variations.
The field is moving rapidly and much of what was true about LLMs a few years ago with GPT-3 is no longer true with modern models.
I don’t know… I’ve been using ChatGPT4. I use it only where the knowledge it outputs is not important. It’s good when I need help with language related things, as more of a writing assistant. Creative stuff is also OK, sometimes even impressive.
With facts? On moderately complicated topics? I’d say it gets something subtly wrong about 80% of the time, and very obviously wrong 20%. The latter isn’t the problem.
I don’t understand where the “intelligent” part would even come in. Sure, it requires a fair level of intelligence to understand and generate human language responses. But, to me, all I’ve seen fits: generate responses that seem plausible as responses to the input.
If intelligence requires some deeper understanding of the world, and the facts and relationships between them, then I don’t see it. It’s just a coincidence when it looks like it happened. It’s impressive how often that is, but it’s still all it is.
LLMs have a good time with standardized tests like SAT precisely because they’re standardized, i.e. there’s enough information on the internet for them to parrot on them
Try something more complex and free-form and where a human might have to work a little more to break it down into actual little subtasks with their intelligence - and then solve it, LLMs in the best case scenario will just say they don’t know how to do it, and in the worst case scenario they’ll hallucinate some actual bullshit.
Ask an LLM to explain a joke. It often won’t understand why a joke is funny, but that won’t stop it from trying to give you an explanation.
Eh, yes and no. It might help illustrate the limitations of testing for some people, but it’s not really telling us anything new about them. It is meant to cheaply provide an indication of how a student is fairing and has never been considered by anyone serious as some kind of comprehensive measure of intelligence. Their flaws have been known for a long time.
Those tests are not for intelligence. They’re testing whether you’ve done the pre-requisite work and acquired the skills necessary to continue advancing towards your desired career.
Wouldn’t want a lawyer that didn’t know anything about how the law works, after all, maybe they just cheated through their classes or something.
We use standardized tests because they’re cheap pieces of paper we can print out by the thousands and give out to a schoolfull of children and get an approximation of their relative intelligence among a limited range of types of intelligence. If we wanted an actual reliable measure of each kid’s intelligence type they’d get one-on-one attention and go through a range of tests, but that would cost too much (in time & money), so we just approximate with the cheap paper thing instead. Probably we could develop better tests that accounted for more kinds of intelligence, but I’m guessing those other types of intelligence aren’t as useful to capitalism, so we ignore them.
Everyone knew this.
Obviously 1:1 mentoring, optional cohort/Custom grouping, experiential, self paced, custom versioned assignment learning is best but that’s simply not practical for a massive system.
Intelligence cannot be measured. It’s a reification fallacy. Inelegance is colloquial and subjective.
If I told you that I had an instrument that could objectively measure beauty, you’d see the problem right away.
But intelligence is the capacity to solve problems. If you can solve problems quickly, you are by definition intelligent.
the ability to apply knowledge to manipulate one’s environment or to think abstractly as measured by objective criteria (such as tests)
https://www.merriam-webster.com/dictionary/intelligence
It can be measured by objective tests. It’s not subjective like beauty or humor.
The problem with AI doing these tests is that it has seen and memorized all the previous questions and answers. Many of the tests mentioned are not tests of reasoning, but recall: the bar exam, for example.
If any random person studied every previous question and answer, they would do well too. No one would be amazed that an answer key knew all the answers.
But intelligence is the capacity to solve problems. If you can solve problems quickly, you are by definition intelligent
To solve any problems? Because when I run a computer simulation from a random initial state, that’s technically the computer solving a problem it’s never seen before, and it is trillions of times faster than me. Does that mean the computer is trillions of times more intelligent than me?
the ability to apply knowledge to manipulate one’s environment or to think abstractly as measured by objective criteria (such as tests)
If we built a true super-genius AI but never let it leave a small container, is it not intelligent because WE never let it manipulate its environment? And regarding the tests in the Merriam Webster definition, I suspect it’s talking about “IQ tests”, which in practice are known to be at least partially not objective. Just as an example, it’s known that you can study for and improve your score on an IQ test. How does studying for a test increase your “ability to apply knowledge”? I can think of some potential pathways, but we’re basically back to it not being clearly defined.
In essence, what I’m trying to say is that even though we can write down some definition for “intelligence”, it’s still not a concept that even humans have a fantastic understanding of, even for other humans. When we try to think of types of non-human intelligence, our current models for intelligence fall apart even more. Not that I think current LLMs are actually “intelligent” by however you would define the term.
Does that mean the computer is trillions of times more intelligent than me?
And in addition, is an encyclopedia intelligent because it holds many answers?
This is a semantic argument.
Have you never felt smarter or dumber depending on the situation? If so, did your ability to think abstractly, apply knowledge, or manipulate your environment change? Intelligence is subjective (and colloquial) like beauty and humor.
This isn’t quite correct. There is the possibility of biasing the results with the training data, but models are performing well at things they haven’t seen before.
For example, this guy took an IQ test, rewrote the visual questions as natural language questions, and gave the test to various LLMs:
https://www.maximumtruth.org/p/ais-ranked-by-iq-ai-passes-100-iq
These are questions with specific wording that the models won’t have been trained on given he wrote them out fresh. Old models have IQ results that are very poor, but the SotA model right now scores a 100.
People who are engaging with the free version of ChatGPT and think “LLMs are dumb” is kind of like talking to a moron human and thinking “humans are dumb.” Yes, the free version of ChatGPT has around a 60 IQ on that test, but it also doesn’t represent the cream of the crop.
Maybe, but this is giving the AI a lot of help. No one rewrites visual questions for humans who take IQ tests. That spacial reasoning is part of the test.
In reality, no AI would pass any test because the first part is writing your name on the paper. Just doing that is beyond most AIs because they literally don’t have to deal with the real world. They don’t actually understand anything.
They don’t actually understand anything.
This isn’t correct and has been shown not to be correct in research over and over and over in the past year.
The investigation reveals that Othello-GPT encapsulates a linear representation of opposing pieces, a factor that causally steers its decision-making process. This paper further elucidates the interplay between the linear world representation and causal decision-making, and their dependence on layer depth and model complexity.
https://arxiv.org/abs/2310.07582
Sizeable differences exist among model capabilities that are not captured by their ranking on popular LLM leaderboards (“cramming for the leaderboard”). Furthermore, simple probability calculations indicate that GPT-4’s reasonable performance on k=5 is suggestive of going beyond “stochastic parrot” behavior (Bender et al., 2021), i.e., it combines skills in ways that it had not seen during training.
We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2’s performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT).
Just a few of the relevant papers you might want to check out before stating things as facts.