I paid a bit more than $1 to fine tune the biggest GPT3 model for my use case. I didn’t push a lot of data and the fine tuning took 35 minutes (without counting the time waiting in the queue).
I find it a little unsettling that GPT-3 is able to do this well. This kind of problem is abstract and I assume its training data doesn't extensively cover this kind of thing. I wonder if it would work meaningfully worse if you used different words in the prompt other "edge" and "node" which are probably used in examples somewhat like this.
Example prompt -
Every turn each gloober flomps the same gloobers. If a yellow gloober flomps a blue gloober then the blue gloober turns yellow, otherwise, gloobers stay the same color.
1. gloober Roy flomps gloober Zoo and gloober Bat
2. gloober Zoo flomps gloober Bat and gloober Crystal
3. gloober Crystal flomps gloober X-Ray
4. gloober Bat flomps gloober Blue
If gloober Zoo is the only gloober that is yellow, and all other gloobers are blue, will gloober Blue turn yellow, and, if so, at what turn?
GPT-3 didn't even need the `gloober Blue` prompt as part of `Answer:`. It 0-shotted (no other training examples required) it straight out of the box.
Prompt:
Every turn each gloober flomps the same gloobers. If a yellow gloober flomps a blue gloober then the blue gloober turns yellow, otherwise, gloobers stay the same color.
1. gloober Roy flomps gloober Zoo and gloober Bat
2. gloober Zoo flomps gloober Bat and gloober Crystal
3. gloober Crystal flomps gloober X-Ray
4. gloober Bat flomps gloober Blue
If gloober Zoo is the only gloober that is yellow, and all other gloobers are blue, will gloober Blue turn yellow, and, if so, at what turn?
Answer:
Continuation:
gloober Blue will turn yellow on turn 4.
> I find it a little unsettling that GPT-3 is able to do this well.
I agree and think that this is definitely the right reaction.
I know that GPT-3 is not AGI nor modeled properly to ever represent what we would consider to be AGI; but this feels freakishly AGI-ish. This feels like the model is 'reasoning' in a way that I would not expect based on my intuition of how GPT-3 works. As it's been explained to me in lay terms; GPT is great at predicting the next word based on a chunk of previous context. This feels like it's doing much more than that as it's understanding the context and parameters that relate to a question embedded in the text, which seems unrelated to the original instruction of GPT-3 as I understand. Can anyone explain?
I think the sentences in this example are, odd vocabulary notwithstanding, fairly typical descriptions of logical implication, which I expect has non-trivial representation in the training set. GPT-3 is probably capable of matching the pattern of the tokens (sentence structure), even if the content of the tokens (spelling/vocabulary) is novel.
I doubt it. There are billions of parameters and weights all combining in non trivial ways to produce the output. I don’t think you can describe any of its workings with simple terms like “doing token substitution” even if that’s what it looks like to us when we look at the output.
Try it. Reduce the tokens to single characters but keep the structure. Does it work? Maybe single characters looks too much like an abbreviation and that throws it off, because those are found in sufficiently different contexts. Exchange to two characters, three. Throw similar inputs at it and find the point where it reliably "gets it".
A NN is a means of compressing a training dataset into a mathematical representation (weights) which can be exploited for pattern finding. The representation of the training data is where 95% of the "magic" comes from. We intelligently prepare it, so no intelligence is required in its use.
On any given input, its predictions can "look" how we expect an intelligent response to "look". The job is to disprove this intuition, which is usually trivial. This cannot be done by looking at a single response.
Consider walking into a room and predicting a few people's personality traits correctly based on their star sign. This shows nothing.
This sounds like you consider what has been called "feature engineering" in older ML research as being fundamental to modern ML successes. But the trajectory has been the opposite direction, away from feature engineering, and instead towards letting the transformer architecture learn as much as possible, including interesting features.
Most research on deep net architecture, is about how to train it effectively on huge stacks of GPUs and TPUs. That's how/why transformers were invented.
weights are just compressions of the training set. The reason hardware improvements have improved ML is just because we're compressing the training set to larger sizes.
Going eg., from 100TB training to 300BB weights. The process of training is more-or-less just "building a search index", and prediction is little more than "looking up examples in the search index and combining".
Modern DL uses stochastic gradient descent, which Domingos's paper doesn't deal with, in fact he leaves the generalisation as an open problem.
Be that as it may, universality in computing models is a well known issue. I could also say "Every Model Learned by Gradient Descent Is an x86 processor + a lot of RAM", or ""Every Model Learned by Gradient Descent Is a 2 Layer Neural Net"
(as 2 layer nets are universal approximators), but this is not terribly illuminating.
Yes, one of the key insights of modern DL-based ML is that size matters, neural networks start producing interesting generalisation only once you have large parameter spaces. Indeed, one might wonder if size is all that matters?
I don't see what is wrong with search indices, every computer is a gigantic search index, because main memory is a map from an index (the memory address) to 2^n bit integers, the value stored at the index. So anything that can be computed at all, can be computed by "a search index". The important thing that, prima facie, distinguish NNs from eg SQL databases, is the use real numbers as latent space: because real numbers are a metric space, so you have a notion of distance which allows you do use efficient local search like gradient descent. Moreover, when you ask this 'mere' search index (NNs) about something that it was not trained on, it can (and probably does) return something that is close in this metric space to something it learned.
That did not work at all in the 1990s. Today it works spectacularly well in some domains. So an interesting question now is, to use your language: is human intelligence anything more than "building a search index" and "looking up examples in the search index and combining"?
So this is a variation of Searle's Chinese room, but unlike him in 1980, we have working Chinese rooms!
It is important to note that the search space is just the training data. That space does not span enough dimensions to actually solve the problems in question.
For example, we act in the world so as to set the state of possible causal variables -- and thereby can distinguis
causes.
Above, graph algorithms operate on a discrete graph space and hence on the graph. GPT here isn't doing that.
Seeing NNs as distance-based searches in a compressed 'training space' demystifies what solutions can be found.
Here the text space defined by the training set does not permit a graph traversal, so eg., the NN 'uses' edges that don't exist
We are not searches across such spaces, which are radically impoverished and not parametrised by time or action
It is not the case that "the search space is just the training data", the search space in NNs are some variants of real multivariate functions. The training data can be all manner of things, in LLMs, they are text in natural language. Graphs or natural language text are mapped from a discrete space to a continuous metric space.
It is true that LLMs cannot currently solve complex mathematics problems reliably. Whether that's an intrinsic shortcoming of transformers, or not is an interesting open question, that is being investigated as we speak. A lot of well-known DL research teams bet on transformers beating humans in mathematics in the next few years, see e.g.
"is human intelligence anything more than "building a search index" and "looking up examples in the search index and combining"?"
Yes, it is more than this. Even if we can simulate things that appear like thinking using these methods, what we do bears no comparison. That's not a criticism of the computer here but the brain doesn't work like this.
Ultimately, maybe the question "can computers think" is meaningless as Turing himself said. If the output appears like it is the product of thought, then what difference does it make?
> The process of training is more-or-less just "building a search index", and prediction is little more than "looking up examples in the search index and combining".
It's really not.
You are compressing the information in that dataset. That means the model has to learn characteristics of the information in the data that represent it. This is one of the reasons why some argue that compression is a form of intelligence[1]
It's a very different mechanism to a search index or database even if in some case it appears to work similarly.
Your example implies that they are not affected by star signs. This is arguably false given people believe in star signs and thus certainly influenced by the traits of said star signs.
My point is a few instances of reality confirming our prejudices is not the basis for explanation. This is something we've already discovered, it's called "the scientific method", and its wholey absent from ML.
Speaking of things that feel weird, while I had the playground open I tried another prompt that came to to my mind. This prompt is based on the idea that people will usually miss it if you double a word (e.g. I doubled "to" a few words back - did you notice?)
I gave GPT-3 a prompt to test this. Like a human reader would, GPT-3 seems to miss the doubled word in the last sentence. Continuation is at the asterisk.
Repeat the sentence in quotes.
"The man went to the gas station today."
The man went to the gas station today.
"I'm happy to be on holiday."
I'm happy to be on holiday.
"If you can get to the the store on time, it would be appreciated"
If you can get to the store on time, it would be appreciated.
---
What's extra-odd about this is that GPT-3 can repeat random character sequences up to 1,000 characters in legnth. For each of these I just gave the sentence in quotes and let GPT-3 autocomplete the repeat. I think that's especially odd because it's showing that GPT-3 does something different when it is asked to repeat random characters (wouldn't miss a duplicate) versus when it is asked to repeat sentences that it could "read".
"I need something stronger than I'm XXXknjd used to."
I need something stronger than I'm XXXknjd used to.
The trick was diversity. GPT-3 trained on so many tasks it can understand zero shot ones like pathfinding, even when formulated in completely different terms.
I've been messing with the prompt a little bit and tried to get it to spell out what it thinks is happening. (I changed the name of gloober Blue to gloober Lovely just to make it easier to read). Continuation is at the asterisk.
Every turn each gloober flomps the same gloobers. If a yellow gloober flomps a blue gloober then the blue gloober turns yellow, otherwise, gloobers stay the same color.
1. gloober Roy flomps gloober Zoo
2. gloober Roy flomps gloober Bat
3. gloober Zoo flomps gloober Bat
4. gloober Zoo flomps gloober Crystal
5. gloober Crystal flomps gloober X-Ray
6. gloober Bat flomps gloober Lovely
gloober Zoo is the only yellow gloober. Write the turn and which gloobers are yellow.
1. Yellow gloobers - Zoo.
2. Yellow gloobers -* Zoo, Bat.
3. Yellow gloobers - Zoo, Bat, Crystal.
4. Yellow gloobers - Zoo, Bat, Crystal, X-Ray.
5. Yellow gloobers - Zoo, Bat, Crystal, X-Ray, Lovely.
6. Yellow gloobers - Zoo, Bat, Crystal, X-Ray, Lovely.
Other than thinking the spread of yellow is limited to one gloober at a time, this seems right.
The text says "Every turn each gloober flomps the same gloobers."
If I understand you correctly, you are applying only one of the steps each turn, in order, which contradicts the instructions by flomping different gloobers each turn.
It is also guessable in multiple ways which is unfortunate. "There are four steps specified so I guess it happens on turn four." is one invalid way of getting the same answer.
I disagree with your definition of "well". It found an optimal solution only 30% (roughly inferred from the bar plot) of the time and only on graphs of very limited size (up to 14 nodes). I'm pretty sure even the most basic path-finding algorithms we have can do better than this.
It's not about competing with good path finding algorithms. It would not be surprising to find a program that someone had written to find paths that could kind of find paths okay. The surprising thing is that nobody wrote this program to find paths at all. This is a program that is attempting to predict what the next characters are in a string of text - and it can kind of do path finding.
The problem with research conducted in this way is that there's no theory of operation of the model, except the implicit one "it works like how I expect".
And then there's no hypothesis test against predictions made based on this theory, ie., no one tries to disconfirm the model works that way. It is usually trivial to disconfirm. ML models almost always work some other way, usually exploiting coincidental statistical information in the training set.
This turns the whole show into a superstitious activity, in which we think google answers our "local restaurants" query because it knows where I live (rather than say, using an IP location or one of 100 other methods).
If you present a ML model as "capable of x", you're required to disprove that claim as best you can, and present the occasions where it failed.
It might not be rigorous science but it's a cracker of a blog post.
It shows a surprising ability of language models, code is provided and plenty of mystery is left for the reader... which can be investigated by almost anyone for tens of dollars.
What would a useless "random" algorithm find operating on these graphs as inputs? What would a useful one find? What would that be "useful" for?
Is this system much closer to the useless or the useful? We cant tell.
Without any explanation, we're left reading this blog post as a magic trick in the literal sense.
Consider the rate at which it "uses edges that don't exist" -- this makes it not a graph traversal algorithm. So it is just exploting syntax in the input whose representation has some "graphlike" qualities which can be statistically exploited.
Why is that interesting? I dont know. Is this a good system?
(EDIT: I can safely say this isn't practically useful... but it's interesting in the same way as watching a cat try to fill out a tax return. I'm not surprised the cat does it badly, I'm surprised that it seems to be doing it at all, even if it's really a trick).
When Newton described the laws of gravity, now often referred to as Newtonian physics, did he explain what gravity was, in other words, did he have a "theory of operation of the model, except the implicit one"? No, au contraire! He gave laws that took gravity as a black box, but worked very well allowing us to predict things "like how I expect". It has been argued that giving up on explaining what gravity really is, and instead focussing on being able to predict its workings, has been one of Newton's great insights. One might argue that we still don't understand what gravity and the remaining fundamental forces (electromagnetic, strong and weak forces) really are, we still treat them as a black box, just like modern ML research mostly treat transformers (and similar) as a black box. That's normal science. Indeed, what else can you do?
The fact that "it works like how I expect" and that there is a clear trajectory of working more and more "like how I expect" is a strong indicator of scientific progress.
The semantics for the mathematical model /are/ the theory of how it works.
M = mass, F = force etc.
st., the model claims we are in a world of masses and forces, and so on. And this model can be disconfirmed.
You're confusing two issues here. Nowhere in the whole of science are models presented without an explanatory semantics. Almost nowhere in the whole of ML, are they.
Here it is trivial to disconfirm even the authors interpretation of the NNs 'finding' of solutions, by noting it is often using edges that don't exist.
This whatever it is doing, we can immediately show it isn't traversing a graph. The author seems to wilfully opted for the superstitious interpretation
> ML models almost always work some other way , usually exploiting coincidental statistical information in the training set.
That's true for models trained on small datasets, but GPT-3's dataset was anything but superficial, and on top of it we consider this task to be OOD, thus the amazement.
The performance of this model using English descriptions suggests to me the possibility of a model that is built to work using precise mathematical notation performing even better.
Are there models that are trained to receive and respond with a purely mathematical description of the problem?
OpenAI has trained a model to solve Math Olympiad-type problems[1], with some initial success, but currently the model isn't good at coming up with proofs involving more than a couple arguments chained together. Still some very impressive work.
There is zero evidence for the AGI claim or even a weaker intelligence claim. GPT 3 is just a search engine on steroids. I'm regularly suprised by the results but it's honestly just an illusion as far as 'intelligence' goes. This has nothing to do with intelligence, and anyone who thinks otherwise is succumbing to the Eliza effect.
Why can't the same thing be said of human intelligence? If we knew more about neuroscience it'd probably seem just as trivial. Wetware trained on examples we've seen. Experience encoding neurons. Neurons running on the laws of physics. Data goes in, endogenous processing happens, then outputs. I don't see why it's categorically different. ANNs are just demystified because we made them and know how they work.
"why can't the same thing be said of human intelligence"
"if we knew more about neuroscience..."
So that's why we can't say the same thing.
Fwiw, my take is that whatever the brain does, its a machine of some sort but possibly not even a turing machine. Many disagree but noone can prove me wrong. Yet at least.
The human mind in its totality, ie including consciousness, is a fundamentally different kind of machine to a computer. It's a machine, there is nothing supernatural here, and I am not invoking a 'god of the gaps' argument. However, like Roger Penrose, I think that our consciousness is not a computation for example. That kind of thing seems pretty key. If that is true, and what we call 'understanding' is a function of consciousness, then that's going to have implications.
It doesn't matter by the way, if Turing machines can't replicate human minds, why would we want that anyway? "Intelligent" systems that provide outputs and solve problems for us is what we want right? I think we'll have something that is, for all intents and purposes, "AGI", using classical computers. A 'philosophical zombie' in other words.
Intelligence is the capacity to survive and thrive by taking actions that are likely to succeed. Artificial intelligence is defined by the ability to take actions that maximize measures of success. Artificial general intelligence is defined by the ability to operate intelligently regardless of context; ie, not merely trained in a particular context but capable of abstracting and generally succeeding in a range of contexts roughly as broad as a human.
By those reasonable and common definitions, GPT3 seems like a form of AGI. I mean, it is used for protein synthesis sequences, robotic sequences, design sequences— and it is a fabulous writer in a vast range of domains.
It doesn’t need to be god to be AGI v1.
I work with it every day and I recognize my contributions to it. But it is so generally intelligent. That is, it produces successful outcomes (useful text) in a vast range of contexts.
I tried in the thread to give clear definitions of intelligence and AGI in order to make an argument for.
I didn’t give references for the definitions, but for intelligence, I think Legg and Hutter (2007) & Russell and Norvig (2010) work. And Wikipedia has some tests for AGI.
I asked GPT3 “if you were going to go into a person’s house and make a cup of coffee, detail what steps you’d take.” GPT3 responded with text that is indistinguishable from what I’d expect from a high school graduate, at least.
1. Knock on the door or ring the doorbell.
2. Wait for the person to answer the door.
3. If the person invites you in, step into the house.
4. Find the kitchen and locate the coffee pot and coffee beans.
5. Measure out the desired amount of coffee beans and grind them.
Would be interesting to see if it does better if the process was broken down into ’moves’ and the model asked/reminded what its next possible steps are along the way.
Great writeup! I made a couple of changes in your code and it seems it gets ~90% of examples correct (inside the first three categories you have) with aroung 65% having either optimal path or correctly saying that there is no path.
What I changed was:
1. Used code-davinci-002 (codex)
2. Instead of using an explanation of how these tasks work, I changed your code prompts slightly and instead gave it 3 examples (two with paths, and one to show it what to output when there is no path present).
3. Changed the output so that it has to tell me which edges, instead of which nodes, it is traversing (this helps GPT to avoid using nonexistent edges it seems).
Here is an example: https://pastebin.com/D7Hn95VC Last "Problem" is the real problem we want to solve of course, everything else is static.
I'll post my code when I finish testing :)
EDIT: I used Codex not because I think it's better suited for this, but because it's currently free. I don't have enough credit ATM to run 1000 iterations with text-davinci, so I have no idea what difference using Codex made.
- Found the optimal path 504
- Found a (non-optimal) path 293
- Correctly reported no path exists 131
- Found a solution when none existed 19
- Ended on wrong node 0
- Used edges that don't exist 53
- Started with the wrong node 0
- Incorrectly reported no path exists 0
- Total: 1000
80 comments
[ 2.3 ms ] story [ 157 ms ] threadYou can find tune gpt-J or gpt-neo with a 3090
Example prompt -
Every turn each gloober flomps the same gloobers. If a yellow gloober flomps a blue gloober then the blue gloober turns yellow, otherwise, gloobers stay the same color.
1. gloober Roy flomps gloober Zoo and gloober Bat
2. gloober Zoo flomps gloober Bat and gloober Crystal
3. gloober Crystal flomps gloober X-Ray
4. gloober Bat flomps gloober Blue
If gloober Zoo is the only gloober that is yellow, and all other gloobers are blue, will gloober Blue turn yellow, and, if so, at what turn?
Answer: gloober Blue
Prompt:
Continuation: > I find it a little unsettling that GPT-3 is able to do this well.I agree and think that this is definitely the right reaction.
[0]: http://starai.cs.ucla.edu/papers/ZhangArxiv22.pdf
On any given input, its predictions can "look" how we expect an intelligent response to "look". The job is to disprove this intuition, which is usually trivial. This cannot be done by looking at a single response.
Consider walking into a room and predicting a few people's personality traits correctly based on their star sign. This shows nothing.
Most research on deep net architecture, is about how to train it effectively on huge stacks of GPUs and TPUs. That's how/why transformers were invented.
Going eg., from 100TB training to 300BB weights. The process of training is more-or-less just "building a search index", and prediction is little more than "looking up examples in the search index and combining".
https://arxiv.org/abs/2012.00152
Be that as it may, universality in computing models is a well known issue. I could also say "Every Model Learned by Gradient Descent Is an x86 processor + a lot of RAM", or ""Every Model Learned by Gradient Descent Is a 2 Layer Neural Net" (as 2 layer nets are universal approximators), but this is not terribly illuminating.
Yes, one of the key insights of modern DL-based ML is that size matters, neural networks start producing interesting generalisation only once you have large parameter spaces. Indeed, one might wonder if size is all that matters?
I don't see what is wrong with search indices, every computer is a gigantic search index, because main memory is a map from an index (the memory address) to 2^n bit integers, the value stored at the index. So anything that can be computed at all, can be computed by "a search index". The important thing that, prima facie, distinguish NNs from eg SQL databases, is the use real numbers as latent space: because real numbers are a metric space, so you have a notion of distance which allows you do use efficient local search like gradient descent. Moreover, when you ask this 'mere' search index (NNs) about something that it was not trained on, it can (and probably does) return something that is close in this metric space to something it learned.
That did not work at all in the 1990s. Today it works spectacularly well in some domains. So an interesting question now is, to use your language: is human intelligence anything more than "building a search index" and "looking up examples in the search index and combining"? So this is a variation of Searle's Chinese room, but unlike him in 1980, we have working Chinese rooms!
For example, we act in the world so as to set the state of possible causal variables -- and thereby can distinguis causes.
Above, graph algorithms operate on a discrete graph space and hence on the graph. GPT here isn't doing that.
Seeing NNs as distance-based searches in a compressed 'training space' demystifies what solutions can be found.
Here the text space defined by the training set does not permit a graph traversal, so eg., the NN 'uses' edges that don't exist
We are not searches across such spaces, which are radically impoverished and not parametrised by time or action
It is true that LLMs cannot currently solve complex mathematics problems reliably. Whether that's an intrinsic shortcoming of transformers, or not is an interesting open question, that is being investigated as we speak. A lot of well-known DL research teams bet on transformers beating humans in mathematics in the next few years, see e.g.
Yes, it is more than this. Even if we can simulate things that appear like thinking using these methods, what we do bears no comparison. That's not a criticism of the computer here but the brain doesn't work like this.
Ultimately, maybe the question "can computers think" is meaningless as Turing himself said. If the output appears like it is the product of thought, then what difference does it make?
It's really not.
You are compressing the information in that dataset. That means the model has to learn characteristics of the information in the data that represent it. This is one of the reasons why some argue that compression is a form of intelligence[1]
It's a very different mechanism to a search index or database even if in some case it appears to work similarly.
[1] http://www.hutter1.net/ai/
I gave GPT-3 a prompt to test this. Like a human reader would, GPT-3 seems to miss the doubled word in the last sentence. Continuation is at the asterisk.
---What's extra-odd about this is that GPT-3 can repeat random character sequences up to 1,000 characters in legnth. For each of these I just gave the sentence in quotes and let GPT-3 autocomplete the repeat. I think that's especially odd because it's showing that GPT-3 does something different when it is asked to repeat random characters (wouldn't miss a duplicate) versus when it is asked to repeat sentences that it could "read".
"I need something stronger than I'm XXXknjd used to."
I need something stronger than I'm XXXknjd used to.
"pmkkrrpjvn"
pmkkrrpjvn
"wxvbssajxczjjewmanvzznxlwzwdvfrghsstmzpvqwsstrxkzunjgxicvrshychbqsulydkiwgyryhkgjtukssacnodtggfahvus"
wxvbssajxczjjewmanvzznxlwzwdvfrghsstmzpvqwsstrxkzunjgxicvrshychbqsulydkiwgyryhkgjtukssacnodtggfahvus
"dcyvufmxwmctighwwhmamgwuyldpkzrwubvcbjirxzhuoqtktcdfjgasurcprctnxmlqfvwkxixrtqqynxfjvtfhmgsfglxhvfhnbqnixulexkqflrdxqpwgavrphkbgspgoyfjfolczssjfrmggtpcdheneaynhdwimcmvkcsauxwnlovefpmqyznynuoxraqixeveiuqotssiymrtmbiodwnhmtexfrttzdlbdwduqzvrkpmqmjcyciqsuwqmlhjrveeozktvccsspsqosutpngigzkqgaltojhvkodzoeqzhlhfqyblvqkoyvjkphmjroidjpbpnkvrchvrnvuoqhgzkypvjycfzrdsyefstckorqwgsbknslqyiwqghdinuwrapkgyvhslbfgqrslsrferhodlvukdsvvzdcwyjwtrqhqkvwbwksefuanzaarvhnmjcgtajsbivaijthercbmqfvndlsfhjtwsiztpwcehuvckmcbjvhnqbvhhqwugrodtyjfrwlkbpatpwgrlzneenyqgcdyzzhrzfsrvfqlnunfpnathcpuealrdwcliriufxdfqaygxtvdriidzzzdbbfwrowhlbasogmzfkwhaaufrcjsdgdobaozazbdfmasrgjiblluvfvfkeuwksyrzdmepmapesadqoozrpaidlgshyhpbypdzdxjdtwnfaxthjwluarbldmxvvfdissthlpdxqxjmgpuefwqvfjupxveztgrrwyobgaljqkgexeulvjucokythfkmcgnlbfkmpgrpaoztrawqciyilbzwjxagwqgwuqcwactwdtobomlkanmdbhpmubdzvolqpusjbtgqoymuovzsqbkzonabhzogabraxmnxnvfdotfisefmfbdbqjzgrhdtgtppobyndfgsrlpxyutebzlrxgahwizvtmieutqnkflkldxeathvrttyffacofuwdkhxdklcisbzbrlddtahut"
dcyvufmxwmctighwwhmamgwuyldpkzrwubvcbjirxzhuoqtktcdfjgasurcprctnxmlqfvwkxixrtqqynxfjvtfhmgsfglxhvfhnbqnixulexkqflrdxqpwgavrphkbgspgoyfjfolczssjfrmggtpcdheneaynhdwimcmvkcsauxwnlovefpmqyznynuoxraqixeveiuqotssiymrtmbiodwnhmtexfrttzdlbdwduqzvrkpmqmjcyciqsuwqmlhjrveeozktvccsspsqosutpngigzkqgaltojhvkodzoeqzhlhfqyblvqkoyvjkphmjroidjpbpnkvrchvrnvuoqhgzkypvjycfzrdsyefstckorqwgsbknslqyiwqghdinuwrapkgyvhslbfgqrslsrferhodlvukdsvvzdcwyjwtrqhqkvwbwksefuanzaarvhnmjcgtajsbivaijthercbmqfvndlsfhjtwsiztpwcehuvckmcbjvhnqbvhhqwugrodtyjfrwlkbpatpwgrlzneenyqgcdyzzhrzfsrvfqlnunfpnathcpuealrdwcliriufxdfqaygxtvdriidzzzdbbfwrowhlbasogmzfkwhaaufrcjsdgdobaozazbdfmasrgjiblluvfvfkeuwksyrzdmepmapesadqoozrpaidlgshyhpbypdzdxjdtwnfaxthjwluarbldmxvvfdissthlpdxqxjmgpuefwqvfjupxveztgrrwyobgaljqkgexeulvjucokythfkmcgnlbfkmpgrpaoztrawqciyilbzwjxagwqgwuqcwactwdtobomlkanmdbhpmubdzvolqpusjbtgqoymuovzsqbkzonabhzogabraxmnxnvfdotfisefmfbdbqjzgrhdtgtppobyndfgsrlpxyutebzlrxgahwizvtmieutqnkflkldxeathvrttyffacofuwdkhxdklcisbzbrlddtahut
Yup. Here are the results I get when manually applying these rules myself.
Here I mention only the flomps that change gloobers to become yellow.
Turn 1: Zoo is yellow.
Turn 2: Zoo flomps Bat, turning Bat yellow. Zoo flomps Crystal, turning Crystal yellow.
Turn 3: Crystal flomps X-Ray, turning X-Ray yellow. Bat flomps Lovely, turning Lovely yellow.
In other words it would be:
1. Yellow gloobers - Zoo.
2. Yellow gloobers - Zoo, Bat, Crystal.
3. Yellow gloobers - Zoo, Bat, Crystal, X-Ray, Lovely.
So it’s got the order correct, and like you said it’s just that it thinks only one gloober is flomped each turn.
It 0-shotted it... but incorrectly, right?
Gloober Zoo flomps gloober Bat on turn 1. Gloober Bat flomps gloober Blue on turn 2.
Answering '4' seems like a total guess, maybe based on there being four steps, or 'Blue' being mentioned in step 4.
The replies to your post seem to assume that the answer is correct, but it is actually demonstrating a lack of step-wise reasoning.
My reading is that on line 1, no action happens, because Roy isn't yellow.
I still don't see how 4 could be a correct answer, could you post a plausible sequence of turns where Blue turns yellow on turn 4?
1. Zoo
2. Zoo, Bat, Crystal
3. Zoo, Bat, Crystal, X-Ray
4. Zoo, Bat, Crystal, X-Ray, Blue
If I understand you correctly, you are applying only one of the steps each turn, in order, which contradicts the instructions by flomping different gloobers each turn.
It is also guessable in multiple ways which is unfortunate. "There are four steps specified so I guess it happens on turn four." is one invalid way of getting the same answer.
And then there's no hypothesis test against predictions made based on this theory, ie., no one tries to disconfirm the model works that way. It is usually trivial to disconfirm. ML models almost always work some other way, usually exploiting coincidental statistical information in the training set.
This turns the whole show into a superstitious activity, in which we think google answers our "local restaurants" query because it knows where I live (rather than say, using an IP location or one of 100 other methods).
If you present a ML model as "capable of x", you're required to disprove that claim as best you can, and present the occasions where it failed.
It shows a surprising ability of language models, code is provided and plenty of mystery is left for the reader... which can be investigated by almost anyone for tens of dollars.
What would a useless "random" algorithm find operating on these graphs as inputs? What would a useful one find? What would that be "useful" for?
Is this system much closer to the useless or the useful? We cant tell.
Without any explanation, we're left reading this blog post as a magic trick in the literal sense.
Consider the rate at which it "uses edges that don't exist" -- this makes it not a graph traversal algorithm. So it is just exploting syntax in the input whose representation has some "graphlike" qualities which can be statistically exploited.
Why is that interesting? I dont know. Is this a good system?
(EDIT: I can safely say this isn't practically useful... but it's interesting in the same way as watching a cat try to fill out a tax return. I'm not surprised the cat does it badly, I'm surprised that it seems to be doing it at all, even if it's really a trick).
The fact that "it works like how I expect" and that there is a clear trajectory of working more and more "like how I expect" is a strong indicator of scientific progress.
M = mass, F = force etc.
st., the model claims we are in a world of masses and forces, and so on. And this model can be disconfirmed.
You're confusing two issues here. Nowhere in the whole of science are models presented without an explanatory semantics. Almost nowhere in the whole of ML, are they.
Here it is trivial to disconfirm even the authors interpretation of the NNs 'finding' of solutions, by noting it is often using edges that don't exist.
This whatever it is doing, we can immediately show it isn't traversing a graph. The author seems to wilfully opted for the superstitious interpretation
That's true for models trained on small datasets, but GPT-3's dataset was anything but superficial, and on top of it we consider this task to be OOD, thus the amazement.
I'm pretty sure nearly everything, including textual representations of graph problems, are found in the training data.
"The internet is vast and infinite" - Motoko Kusanagi at the end of the original ghost in the shell...
Are there models that are trained to receive and respond with a purely mathematical description of the problem?
[1] https://analyticsindiamag.com/openais-neural-theorem-prover-...
1. GPT3 is AGI
2. In combination with people, GPT3 is superintelligence
Strong claims. Anyone care to argue for or against?
"why can't the same thing be said of human intelligence"
"if we knew more about neuroscience..."
So that's why we can't say the same thing.
Fwiw, my take is that whatever the brain does, its a machine of some sort but possibly not even a turing machine. Many disagree but noone can prove me wrong. Yet at least.
It doesn't matter by the way, if Turing machines can't replicate human minds, why would we want that anyway? "Intelligent" systems that provide outputs and solve problems for us is what we want right? I think we'll have something that is, for all intents and purposes, "AGI", using classical computers. A 'philosophical zombie' in other words.
You can start proving your point by representing the smell of incense as "text" and see how far that takes you.
By those reasonable and common definitions, GPT3 seems like a form of AGI. I mean, it is used for protein synthesis sequences, robotic sequences, design sequences— and it is a fabulous writer in a vast range of domains.
It doesn’t need to be god to be AGI v1.
I work with it every day and I recognize my contributions to it. But it is so generally intelligent. That is, it produces successful outcomes (useful text) in a vast range of contexts.
You go first.
I didn’t give references for the definitions, but for intelligence, I think Legg and Hutter (2007) & Russell and Norvig (2010) work. And Wikipedia has some tests for AGI.
I asked GPT3 “if you were going to go into a person’s house and make a cup of coffee, detail what steps you’d take.” GPT3 responded with text that is indistinguishable from what I’d expect from a high school graduate, at least.
1. Knock on the door or ring the doorbell.
2. Wait for the person to answer the door.
3. If the person invites you in, step into the house.
4. Find the kitchen and locate the coffee pot and coffee beans.
5. Measure out the desired amount of coffee beans and grind them.
6. Place the ground coffee in the coffee pot.
7. Fill the coffee pot with water and turn it on.
8. Wait for the coffee to finish brewing.
9. Pour the coffee into a mug and enjoy.
What I changed was:
1. Used code-davinci-002 (codex)
2. Instead of using an explanation of how these tasks work, I changed your code prompts slightly and instead gave it 3 examples (two with paths, and one to show it what to output when there is no path present).
3. Changed the output so that it has to tell me which edges, instead of which nodes, it is traversing (this helps GPT to avoid using nonexistent edges it seems).
Here is an example: https://pastebin.com/D7Hn95VC Last "Problem" is the real problem we want to solve of course, everything else is static.
I'll post my code when I finish testing :)
EDIT: I used Codex not because I think it's better suited for this, but because it's currently free. I don't have enough credit ATM to run 1000 iterations with text-davinci, so I have no idea what difference using Codex made.