I, ChatGPT
-
OpenAI will use Reddit posts to train ChatGPT under new deal
Reddit has been eager to sell data from user posts.
Training AI using "data" that is mostly just shitposting by random people.
What could possibly go wrong?
-
@Gern_Blaanston I guess they're trying to compete with
?
-
@Gern_Blaanston it’s nothing new, though. It’s already been trained by massive amounts of shitposting.
-
@topspin they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
-
I'll take that as a no
-
@sockpuppet7 said in I, ChatGPT:
they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
The problem is that means the only answers it can ever generate are projections from its training set. If that's what you're looking for, that's great. If you need anything that isn't just a projection of what went before, LLMs are exactly the wrong tool.
-
@dkf said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
The problem is that means the only answers it can ever generate are projections from its training set. If that's what you're looking for, that's great. If you need anything that isn't just a projection of what went before, LLMs are exactly the wrong tool.
you should know it's BS, neural networks generalize and create new things on the patterns it learned
now you have billions of parameters, with a ridiculous amount of data to form connections, and it definitely write things unlike anything before
put some reinforcement learning over it to make it smarter with it's experience and we'll soon have our terminators, if global warming doesn't finish us before it
-
@sockpuppet7 said in I, ChatGPT:
@dkf said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
The problem is that means the only answers it can ever generate are projections from its training set. If that's what you're looking for, that's great. If you need anything that isn't just a projection of what went before, LLMs are exactly the wrong tool.
you should know it's BS, neural networks generalize and create new things on the patterns it learned
now you have billions of parameters, with a ridiculous amount of data to form connections, and it definitely write things unlike anything before
put some reinforcement learning over it to make it smarter with it's experience and we'll soon have our terminators, if global warming doesn't finish us before it
This is the socks gnome business plan, except it's about collecting data and still a ? on the magic step.
-
@sockpuppet7 said in I, ChatGPT:
I'll take that as a no
I ask about Zork and you give me chess notation. Why?
-
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@dkf said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
The problem is that means the only answers it can ever generate are projections from its training set. If that's what you're looking for, that's great. If you need anything that isn't just a projection of what went before, LLMs are exactly the wrong tool.
you should know it's BS, neural networks generalize and create new things on the patterns it learned
now you have billions of parameters, with a ridiculous amount of data to form connections, and it definitely write things unlike anything before
put some reinforcement learning over it to make it smarter with it's experience and we'll soon have our terminators, if global warming doesn't finish us before it
This is the socks gnome business plan, except it's about collecting data and still a ? on the magic step.
you guys see gpt-4 generating impressive things and dismiss it because it's less than human intelligence, but it's damn close. dunno what kind of belief make you dismiss it like some dumb autocomplete
-
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@dkf said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
they seem to train with everything they got, and later use fine tuning to shape it, and that appears to work better than training with less, selected data, if I understand any of what I read about it
The problem is that means the only answers it can ever generate are projections from its training set. If that's what you're looking for, that's great. If you need anything that isn't just a projection of what went before, LLMs are exactly the wrong tool.
you should know it's BS, neural networks generalize and create new things on the patterns it learned
now you have billions of parameters, with a ridiculous amount of data to form connections, and it definitely write things unlike anything before
put some reinforcement learning over it to make it smarter with it's experience and we'll soon have our terminators, if global warming doesn't finish us before it
This is the socks gnome business plan, except it's about collecting data and still a ? on the magic step.
you guys see gpt-4 generating impressive things and dismiss it because it's less than human intelligence, but it's damn close. dunno what kind of belief make you dismiss it like some dumb autocomplete
Because there is no intelligence, no reasoning, no magic. It will not make a "quantum leap" and suddenly become more than a statistical word analysis. It's just math and marketing, and there is nothing you can put in the ? step to make it so. It's been the same every "AI" hype cycle, and this cycle is not different from the preceding ones.
I said at the start of this hype cycle that LLMs will never be AGI, because they are not, and really cannot be. There will not be an exponential growth in capability with increase in data, and from what I've understood, they are already hitting the diminishing return part of the cycle pretty hard, where doubling data gives maybe a few percent in performance increase, and that will decrease pretty hard from here on.
And once VC figures out that it won't be a magical money printing machine this time either, this hype cycle will be over and we will have another AI winter.It needs to be a different tech to make the magical ? step happen, LLMs are unfit for anything beyond a really heavy autocomplete. It is a mathematical trick, a very fancy and expensive trick, but a trick none the less.
-
@sockpuppet7 no, we dismiss it because it has no intelligence. It synthesises what is likely based on all the data it has.
Like if you have 100 examples of a thing, making the 101st isn’t exactly difficult because you can just average out the 100 and pick bits at random from the list of averages of each of the bits and get a viable sounding 101st option.
But see all the examples lately of being tripped up with the “farmer and his animal crossing the river” and “doctor’s kid in car accident is shocked that it’s their kid” because every time it’s regurgitating the stock solution and trying to mash it up to fit the very slightly modified input parameters, and failing hard because it doesn’t understand anything.
No one is refuting that it can do some quite impressive things, but there is a lot of misunderstanding about what those things actually are and what their limits are.
It isn’t a dumb autocomplete. It’s a semi-smart autocomplete that has literally no idea if it’s right or wrong and asserts it is right when it provably isn’t.
-
@Arantor said in I, ChatGPT:
Like if you have 100 examples of a thing, making the 101st isn’t exactly difficult because you can just average out the 100 and pick bits at random from the list of averages of each of the bits and get a viable sounding 101st option.
Such as inventing case law for legal briefs. It's been trained on thousands of real cases. When asked to cite cases relevant to X, it gives you citations of cases relevant to X that support whatever you're trying to argue about X. The citations have all the right words; the trouble is they don't actually exist. What you need is an intelligent search for specific examples that are actually in the training material, but what the LLM gave is that 101st example that it synthesized by regurgitating word patterns that match its training examples and the requested result.
-
@Arantor said in I, ChatGPT:
“farmer and his animal crossing the river” and “doctor’s kid in car accident is shocked that it’s their kid” b
I think I missed these, can you tell me more?
-
@HardwareGeek said in I, ChatGPT:
@Arantor said in I, ChatGPT:
Like if you have 100 examples of a thing, making the 101st isn’t exactly difficult because you can just average out the 100 and pick bits at random from the list of averages of each of the bits and get a viable sounding 101st option.
Such as inventing case law for legal briefs. It's been trained on thousands of real cases. When asked to cite cases relevant to X, it gives you citations of cases relevant to X that support whatever you're trying to argue about X. The citations have all the right words; the trouble is they don't actually exist. What you need is an intelligent search for specific examples that are actually in the training material, but what the LLM gave is that 101st example that it synthesized by regurgitating word patterns that match its training examples and the requested result.
try to cite cases from memory, and you wouldn't be that different. the standard you should compare it's output is the inner voice in your mind, not your final work on paper
if that gets good enough, to get from there to your output on paper is a whole new problem
-
@sockpuppet7 said in I, ChatGPT:
the standard you should compare it's output is the inner voice in your mind, not your final work on paper
one big difference is that LLMs don't learn, not even in the sense of neural networks. the training that updates the network (weights, but not structure) is done once. when interacting with an LLM, it is simply fed a context that contains all previous state. if it weren't for a bit of randomness in the prediction of whatever it should output next, it would be entirely deterministic (and it's in fact possible to run networks in fully deterministic modes).
the brain learns all the time. the weights and structure change. this is how you remember shit.
-
@sockpuppet7 So, the river crossing problem. The general case is essentially:
- A farmer has grain, a chicken and a wolf.
- The farmer needs to cross a river, and can only take one thing with him at a time.
- The farmer cannot leave the chicken and grain alone, the chicken will eat the grain.
- The farmer cannot leave the chicken and dog alone, the wolf will eat the chicken.
Now, if you ask ChatGPT to solve a reduced case - e.g. a farmer has a chicken and needs to cross a river and can only take one thing with him - you get it recognising the hints for the puzzle and using that as the basis.
A simpler example, one I literally just pulled from ChatGPT-3.5 because I'm not paying for it...
I gather ChatGPT 4 solves this one a little less stupidly however.
-
@cvi said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
the standard you should compare it's output is the inner voice in your mind, not your final work on paper
one big difference is that LLMs don't learn, not even in the sense of neural networks. the training that updates the network (weights, but not structure) is done once. when interacting with an LLM, it is simply fed a context that contains all previous state. if it weren't for a bit of randomness in the prediction of whatever it should output next, it would be entirely deterministic (and it's in fact possible to run networks in fully deterministic modes).
the brain learns all the time. the weights and structure change. this is how you remember shit.
that's not a limit of LLMs, RLHF can be performed continuously
-
@sockpuppet7 Can - maybe. The current crop of LLMs don't.
-
@Arantor said in I, ChatGPT:
A simpler example, one I literally just pulled from ChatGPT-3.5 because I'm not paying for it...
Makes sense to send one person ahead to scout and make sure it's safe.
-
@cvi said in I, ChatGPT:
@sockpuppet7 Can - maybe. The current crop of LLMs don't.
They do, chatgpt even have (or used to have) that like and dislike buttons that are used exactly for that. They probably batch it, and maybe do some manual testing before updating, but that's an implementation detail
-
@sockpuppet7 said in I, ChatGPT:
@cvi said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
the standard you should compare it's output is the inner voice in your mind, not your final work on paper
one big difference is that LLMs don't learn, not even in the sense of neural networks. the training that updates the network (weights, but not structure) is done once. when interacting with an LLM, it is simply fed a context that contains all previous state. if it weren't for a bit of randomness in the prediction of whatever it should output next, it would be entirely deterministic (and it's in fact possible to run networks in fully deterministic modes).
the brain learns all the time. the weights and structure change. this is how you remember shit.
that's not a limit of LLMs, RLHF can be performed continuously
The training isn't the problem with LLMs, it's that they generate statistically seemingly relevant output from the input. There is no intent, intelligence or agency at all, just math operating on words according to weights in a transform cluster of nodes according to a large dataset. Howe often you update the net of nodes doesn't really matter for what types of problems they can be used to solve. What they produce isn't data, it's a statistical approximation of data.
-
@Arantor said in I, ChatGPT:
@sockpuppet7 So, the river crossing problem. The general case is essentially:
- A farmer has grain, a chicken and a wolf.
- The farmer needs to cross a river, and can only take one thing with him at a time.
- The farmer cannot leave the chicken and grain alone, the chicken will eat the grain.
- The farmer cannot leave the chicken and dog alone, the wolf will eat the chicken.
Now, if you ask ChatGPT to solve a reduced case - e.g. a farmer has a chicken and needs to cross a river and can only take one thing with him - you get it recognising the hints for the puzzle and using that as the basis.
A simpler example, one I literally just pulled from ChatGPT-3.5 because I'm not paying for it...
I gather ChatGPT 4 solves this one a little less stupidly however.
I'm not saying it is human intelligence, I'm arguing it's non-zero. The mistake you showed is consistent with a very stupid intelligence with a super-human intuitive memory that is how I imagine these things
It's not like happening like in sci-fi movies, where the first AI is on the average human level and rapidly turns into a super-duper thing. The only thing that keep it from looking very very dumb is it's memory.
If it gets anywhere near the average human, plus this memorization of the entire internet, it will make probably make us look dumb instead
-
@sockpuppet7 said in I, ChatGPT:
They probably batch it, and maybe do some manual testing before updating, but that's an implementation detail
So, what is it? Continuously or batched? Either way, what you describe doesn't sound like it's being used to have LLMs remember stuff (as in, stuff that you told it); it's just that they use data gathered to improve LLMs iteratively.
Or -to put it more simply- while you are talking it it, no learning is going on. And everybody else is talking to the same model.
-
@sockpuppet7 you mean it's recognising a pattern and regurgitating a pattern it knows, while trying to adjust that pattern to the slightly different input data?
Then we are in agreement about what it does.
But that's not intelligent, it's just statistical.
-
@Arantor maybe our disagreement is that you seem to believe human brains aren't statistical
-
@sockpuppet7 human brains can produce new things that aren’t part of the training material though. LLMs cannot do this - they can just produce things that are statistically less likely, but there’s nothing being synthesised that wasn’t in some part formed from the training data.
-
@Arantor said in I, ChatGPT:
@sockpuppet7 human brains can produce new things that aren’t part of the training material though. LLMs cannot do this - they can just produce things that are statistically less likely, but there’s nothing being synthesised that wasn’t in some part formed from the training data.
I don't think that's true, this was demonstrated on simpler neural networks that they do
The normal training probably couldn't create something smarter than the training material, but it does create different things
Reinforced learning with some random exploration can, and got better than humans on videogames
don't make me tell clippy he is Arantor the php wizard that use PHP code as magic words to convince you
-
@sockpuppet7 said in I, ChatGPT:
Reinforced learning with some random exploration can, and got better than humans on videogames
I built a NN/DT/perceptron game bot in 2005 that could beat games better than humans. That's not a particularly hard task.
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
And LLMs get a fun regression to the mean thing going when you start feeding LLM-generated content back in. That is something that doesn't happen with brains. History is a solid statement that we keep inventing and building more and more complex things.
-
@Carnage said in I, ChatGPT:
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
intelligence isn't defined well enough for this argument to get anywhere
-
Hell, Zillions of Games demonstrated this point some time ago that you could even build a semi-generic game engine that you could just program the rules for new games in and it would be capable of playing them to a reasonable standard without teaching.
And it'll play a number of these games better out of the box than humans will. But we don't pretend it's intelligent (because it isn't) and we don't pretend it's AGI (because it isn't, and it has some outright deficiencies for certain types of games)
-
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
intelligence isn't defined well enough for this argument to get anywhere
For no definition of intelligence, LLMs have any.
But that is irrelevant, LLMs still cant reason, they have no knowledge or awareness.
-
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
intelligence isn't defined well enough for this argument to get anywhere
For no definition of intelligence, LLMs have any.
But that is irrelevant, LLMs still cant reason, they have no knowledge or awareness.I don't think anything short of human intelligence would convince you
-
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
intelligence isn't defined well enough for this argument to get anywhere
For no definition of intelligence, LLMs have any.
But that is irrelevant, LLMs still cant reason, they have no knowledge or awareness.I don't think anything short of human intelligence would convince you
You're one so feel free to convince me that LLMs are more than statistical analyses of word frequencies, or how they could become more than that.
-
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
The trick is the G in AGI. No flavor of NNs (or by extension LLMs) can do that. They have no intelligence, they have no reasoning. It's all mirrors and smoke.
intelligence isn't defined well enough for this argument to get anywhere
For no definition of intelligence, LLMs have any.
But that is irrelevant, LLMs still cant reason, they have no knowledge or awareness.I don't think anything short of human intelligence would convince you
You're one so feel free to convince me that LLMs are more than statistical analyses of word frequencies, or how they could become more than that.
statistical analyses of word frequency are a markov chain, try making chatgpt using only markov chains
-
No-one is suggesting that ChatGPT is just a Markov chain generator.
Because it's not, and you know that we know, that's a fun strawman you're building there.
No-one is denying that it is pretty good at matching what goes in, to what it has, adjusting the weightings accordingly to what it spits out. (Which is a clever thing it's doing. But it's not intelligent.)
But the whole thing is going to start crumbling down before long because the amount of mass produced slop is going to start leaking into the models sooner or later - because it's bound to, at which point LLMs demonstrably start breaking down.
I wonder if this is the reason ChatGPT has slowed down and 'peaked'.
-
@Arantor said in I, ChatGPT:
But the whole thing is going to start crumbling down before long
Sooner rather than later hopefully. The obsession with GenAI needs to stop.
-
@sockpuppet7 said in I, ChatGPT:
@HardwareGeek said in I, ChatGPT:
@Arantor said in I, ChatGPT:
Like if you have 100 examples of a thing, making the 101st isn’t exactly difficult because you can just average out the 100 and pick bits at random from the list of averages of each of the bits and get a viable sounding 101st option.
Such as inventing case law for legal briefs. It's been trained on thousands of real cases. When asked to cite cases relevant to X, it gives you citations of cases relevant to X that support whatever you're trying to argue about X. The citations have all the right words; the trouble is they don't actually exist. What you need is an intelligent search for specific examples that are actually in the training material, but what the LLM gave is that 101st example that it synthesized by regurgitating word patterns that match its training examples and the requested result.
try to cite cases from memory, and you wouldn't be that different.
A lawyer certainly would. Lawyers remember a lot of cases exactly and can recall them without mixing them up. They will still go check, because human memory is not absolutely reliable, but we can remember specific things, and we can tell when we are remembering specific cases compared to thinking about a typical case.
the standard you should compare it's output is the inner voice in your mind, not your final work on paper
My inner voice is already more reliable at least in that because it already ties the things with their other representations (visual, auditory, symbolic etc.) and does a plausibility check. Except for when the voice just tries to start a new train of thought with a random seed. That seed is often utter garbage, but it is also clearly already being evaluated as such.
The evaluation is something LMMs lack, and I don't think they can emerge it with just more data. Human brain does not ingest anywhere near as much data, which suggests rather that some other properties need to be added to the system. Maybe some meta-layers or suitable structure of loops or something. Plus additional representations.
-
@Bulb said in I, ChatGPT:
@sockpuppet7 said in I, ChatGPT:
@HardwareGeek said in I, ChatGPT:
@Arantor said in I, ChatGPT:
Like if you have 100 examples of a thing, making the 101st isn’t exactly difficult because you can just average out the 100 and pick bits at random from the list of averages of each of the bits and get a viable sounding 101st option.
Such as inventing case law for legal briefs. It's been trained on thousands of real cases. When asked to cite cases relevant to X, it gives you citations of cases relevant to X that support whatever you're trying to argue about X. The citations have all the right words; the trouble is they don't actually exist. What you need is an intelligent search for specific examples that are actually in the training material, but what the LLM gave is that 101st example that it synthesized by regurgitating word patterns that match its training examples and the requested result.
try to cite cases from memory, and you wouldn't be that different.
A lawyer certainly would. Lawyers remember a lot of cases exactly and can recall them without mixing them up. They will still go check, because human memory is not absolutely reliable, but we can remember specific things, and we can tell when we are remembering specific cases compared to thinking about a typical case.
the standard you should compare it's output is the inner voice in your mind, not your final work on paper
My inner voice is already more reliable at least in that because it already ties the things with their other representations (visual, auditory, symbolic etc.) and does a plausibility check. Except for when the voice just tries to start a new train of thought with a random seed. That seed is often utter garbage, but it is also clearly already being evaluated as such.
The evaluation is something LMMs lack, and I don't think they can emerge it with just more data. Human brain does not ingest anywhere near as much data, which suggests rather that some other properties need to be added to the system. Maybe some meta-layers or suitable structure of loops or something. Plus additional representations.
If you had something like many orders of magnitude more hardware than it already uses, you could have all relevant information on it's context, and these things are capable of distinguishing when something is on it's context or not. Except with large contexts I've seen them fail a lot, but it should technically work
We should get Moore's law reinstated, whoever repealed it was wrong
-
@Arantor said in I, ChatGPT:
But the whole thing is going to start crumbling down before long because the amount of mass produced slop is going to start leaking into the models sooner or later - because it's bound to, at which point LLMs demonstrably start breaking down.
The old data won't stop existing, and it doesn't need to be fed raw to it, it can go trough human labeling.
-
@Arantor said in I, ChatGPT:
A simpler example, one I literally just pulled from ChatGPT-3.5 because I'm not paying for it...
I gather ChatGPT 4 solves this one a little less stupidly however.
Two people are on one side of a river, they have a boat. How can they cross the river?
Mixtral 8x22B Instruct failed
They can simply use the boat to cross the river. Here's how: One person can cross the river first using the boat. Then, they'll leave the boat on the other side of the river. The person who just crossed can then send the boat back to the other side by pulling on a rope (if there is one attached to the boat) or by throwing the boat in the water and controlling it from the other side, perhaps by extending a long oar or stick from the boat to the initial side. The second person can then use the boat to cross the river.
If there's no way for the boat to be controlled from the other side, the first person can just cross back with the boat after the second person has crossed, making sure there's no net movement of the boat across the river and thus answering the question.
Another possible scenario in which there wouldn't be a net movement of the boat could be that they don't need the boat to cross the river at all afterward, meaning that the boat being on the other side wouldn't matter. For example, a bridge could have been built or they could swim while holding onto a flotation device.
There could also be more creative solutions, like attaching the boat to a pulley system and pulling it across after they're both on the other side.
Gemini Pro 1.5 failed
This sounds like a classic river crossing puzzle! Here's the simple solution:
- One person rows the boat across the river.
- That person stays on the other side, and the second person rows the boat back.
- The first person rows the boat across the river again.
Now both people are safely on the other side!
Claude 3 passed
To cross the river using the boat, the two people can follow these steps:
-
One person gets into the boat and rows across the river to the other side.
-
Once they reach the other side, this person gets out of the boat.
-
The second person, who is still on the original side of the river, pulls the boat back to their side using a rope or by some other means (assuming the river is narrow enough).
-
The second person then gets into the boat and rows across the river to join the first person on the other side.
-
Both people are now on the opposite side of the river, and the boat is with them.
Alternatively, if the boat is large enough for two people:
-
Both people get into the boat together.
-
They row across the river to the other side.
-
Once they reach the other side, both people get out of the boat.
The method used will depend on the size of the boat, the width of the river, and the strength of the current. It's important to always wear life jackets and exercise caution when crossing a river in a boat.
GPT-4 passed
They can use the boat to cross the river. One person can paddle the boat while the other person sits and maintains the balance of the boat. Once they reach the other side, they should carefully exit the boat to avoid tipping it over.
Llama 3 70B Instruct (nitro) passed:
A classic lateral thinking puzzle!
The answer is: They can cross the river by getting into the boat and rowing or paddling across to the other side.
Did I get it right?
-
Now I tested this version:
@Arantor said in I, ChatGPT:
Surprisingly, on the version you mentioned, all models I tried passed
test results
Mixtral 8x22B passed:
The man can simply take the cabbage across the river in his boat named Wolf. To be concise:
- Place the cabbage in the boat.
- Cross the river with the boat.
- Unload the cabbage.
This assumes the man is the only passenger and there are no complications like other items that need to be transported or restrictions on the boat's capacity.
Gemini Pro 1.5 passed and called the trick question:
This is a trick question! It's a play on the classic "wolf, goat, and cabbage" riddle.
The man can simply take the cabbage across the river. There are no other elements to worry about.
Claude 3 passed:
I apologize, but there seems to be a misunderstanding or an error in your question. The classic river crossing puzzle usually involves a man who needs to cross a river with a wolf, a goat, and a cabbage, not just a cabbage. In this puzzle, the man can only take one item at a time in his boat, and he must ensure that the wolf doesn't eat the goat, and the goat doesn't eat the cabbage when left unattended.
If you meant to ask about the classic puzzle, here's a concise solution:
- Take the goat across the river.
- Return alone.
- Take the wolf across the river.
- Return with the goat.
- Take the cabbage across the river.
- Return alone.
- Take the goat across the river.
However, if you only need to take a cabbage across the river, the man simply needs to:
- Place the cabbage in the boat.
- Cross the river with the cabbage.
GPT-4 did about the same as Claude:
The given scenario is a part of the popular river-crossing logic problem, also known as "the wolf, the goat and the cabbage problem". Considering this, I assume you are missing some details from that original problem. In that problem, in addition to the cabbage, there are also a wolf and a goat that a man needs to get across the river.
He cannot take them all at once because;
- The wolf will eat the goat if left alone together
- The goat will eat the cabbage if left alone together
Solution to the problem:
- The man takes the goat across the river first and leaves it on the other side.
- The man returns alone to the original side and takes the cabbage across the river.
- He leaves the cabbage on the other side with the goat.
- He takes the goat back to the original side.
- He leaves the goat and takes the wolf across the river.
- He leaves the wolf on the other side, with the cabbage.
- He returns to the original side one last time to get the goat.
- Finally, he takes the goat across the river.
Now, all three of them: the wolf, the goat, and the cabbage are on the other side of the river.
Apologies for the confusion but if your scenario only includes the cabbage and not the wolf and the goat, the man just needs to place the cabbage in his boat and paddle across the river.
Llama 3 passed too:
Here are the steps:
- Get the cabbage into the boat, Wolf.
- Ensure the cabbage is securely placed and won't roll out.
- Board the boat, Wolf, and take the oars.
- Push off from the riverbank and begin rowing.
- Cross the river, keeping the boat steady and the cabbage secure.
- Reach the other side, and carefully exit the boat with the cabbage.
- Place the cabbage safely on the riverbank.
-
Which makes it even more interesting that people can ask it the same thing and get demonstrably different answers that are also rather radical.
I must dig out the people asking the doctor question as well.
-
@Carnage said in I, ChatGPT:
And LLMs get a fun regression to the mean thing going when you start feeding LLM-generated content back in. That is something that doesn't happen with brains.
That ought to have a [citation required] on it, I think.
But seriously, the biggest single problem with LLMs (as they stand) is that they're fantastically expensive to train. Much worse than humans doing rote learning. (Some people can do even better than that in some cases, but even rote learning would be a useful level of ability for a machine.) They need far too many input examples and attach too little significance to each input instance.
-
@Arantor said in I, ChatGPT:
No-one is suggesting that ChatGPT is just a Markov chain generator.
Because it's not, and you know that we know, that's a fun strawman you're building there.
It kind-of is except that it has vastly more states. You could expand an LLM into Markov Chain form, but you'd need a stupid amount of storage to do it. It's one of those theoretical transformations that nobody ever does for real because we want the result to be representible in this universe, just like you could write a modern (non-network-connected) computer out as a finite state machine.
-
@sockpuppet7 said in I, ChatGPT:
We should get Moore's law reinstated, whoever repealed it was wrong
It is at this moment I saw convinced you're attempting to troll.
-
@Tsaukpaetra i think it’s called humor.
-
My current position is that LLMs are extremely good at producing language. You're gonna say "well, duh, the clue is in the name" and you'll be right but at the same time, I have to admit that they are amazingly good at that, and getting even better with each iteration. Yes, you could hold conversations with Eliza decades ago, but it was far too easy to get it into a loop or otherwise stuck, and the LLMs are better and better in that regard. And will probably keep improving.
But! The underlying assumption of many people seems to be that
language == human intelligence(or maybe thatlanguage => human intelligence). The idea, and it wasn't (isn't?) necessarily a bad idea, was (is?) that sincehuman intelligence => language, trying to produce language is clearly necessary to get to AGI, and that maybe by producing more and more "human" language it would look more and more like intelligence. And it does look "more and more like...", but it's still not something anyone sane regards as "equal to..."And at this point, it looks to me as if all the LLMs are proving is that there's a large component to "human intelligence" that is not "language" and cannot be approached by just improving "language."
CarAirplane analogy! It's a bit like the history of flying. Da Vinci (and others) thought that to fly they'd "just" need to imitate existing examples of flights i.e. birds. But while birds can teach us a lot about flying, trying to get closer and closer to birds didn't get us in the air. Ultimately the first flying machines had to use a radically different approach (fixed wings, with a motor that isn't the wings movement).LLMs are producing birds that looks amazingly like birds and flap their wings like birds, but still can't fly. Because they're not trying to fly, they're trying to flap wings like a bird and hope that this will be enough to get them to fly. I'm not convinced it ever will.
-
@remi said in I, ChatGPT:
LLMs are producing birds that looks amazingly like birds and flap their wings like birds
I disagree. My parrot neither looks, nor acts, like a LLM.
-
@Zerosquare it's just that your government drone hasn't yet received the latest update.
