How does GPT obtain its ability? Tracing emergent abilities of language models
yaofu.notion.site
yaofu.notion.site
"- The two important but different abilities of GPT-3.5 are *knowledge* and *reasoning*. Generally, it would be ideal if we could *offload the knowledge part to the outside retrieval system and let the language model only focus on reasoning.* This is because: - The model’s internal knowledge is always cut off at a certain time. The model always needs up-to-date knowledge to answer up-to-date questions. - Recall we have discussed that is 175B parameter is heavily used for storing knowledge. If we could offload knowledge to be outside the model, then the model parameter might be significantly reduced such that eventually, it can run on a cellphone (call this crazy here, but ChatGPT is already science fiction enough, who knows what the future will be)."
& "Yet there was a WebGPT paper published in Dec 2021. It is likely that this is already tested internally within OpenAI."
It definitely feels like this may be the next step in making this kind of system robust. It ends up being an interface for search.
But I'm not talking about it just being wrong, I'm talking about it citing webpages and books that don't exist and never did[0]. If Wikipedia regularly had that sort of quality issue people just wouldn't use it. There's a threshold below which something stops being useful.
[0] Bloggs, Joe. "ChatGPT just makes stuff up". Nature, vol 123, 2022, pp 123-321. Wiley Online Library, https://doi.org/10.1111/111/111
IDENTIFICATION DIVISION. PROGRAM-ID. CREATE-S3-BUCKET.
ENVIRONMENT DIVISION. CONFIGURATION SECTION.
INPUT-OUTPUT SECTION.
DATA DIVISION. FILE SECTION.
WORKING-STORAGE SECTION. 01 AWS-ACCESS-KEY PIC X(20). 01 AWS-SECRET-KEY PIC X(40). 01 BUCKET-NAME PIC X(255).
PROCEDURE DIVISION. CREATE-BUCKET. MOVE AWS-ACCESS-KEY TO AWS-ACCESS-KEY-VAR MOVE AWS-SECRET-KEY TO AWS-SECRET-KEY-VAR MOVE BUCKET-NAME TO BUCKET-NAME-VAR INVOKE AWS-S3 "CREATE-BUCKET" USING AWS-ACCESS-KEY-VAR AWS-SECRET-KEY-VAR BUCKET-NAME-VAR
"Have no respect whatsoever for authority; forget who said it and instead look what he starts with, where he ends up, and ask yourself, Is it reasonable?" -Richard P. Feynman
"One of the great commandments of science is, "Mistrust arguments from authority." -Carl Sagan
"In questions of science, the authority of a thousand is not worth the humble reasoning of a single individual." -Galileo Galilei
Their arguments are of the form "this statement could be false."
If you evaluate it as a true statement, you have no problems. It could be false; you have to evaluate it for yourself instead of trusting some authority.
It's only if you assert that it's certainly false that you have a problem. Because then it's clearly true -- since otherwise these authorities would be telling you something false, which proves their assertion true.
Put another way, it can get you from the undesirable position of blindly trusting authorities to the desirable position of questioning them, but not the other way around. Which is the intended result.
You can’t trust it’s answers (to be fair that’s the existing status quo), but you also can’t easily test it because it will return reasonable sounding garbage. Conversely you can discover ignorance in most humans pretty quickly by exhausting their ability to respond (or your ability to ask).
It's going to be important to develop AI methods to test and verify, I think unverified model outputs are worthless verbiage. Verification can be based on references, code execution, physical simulations, lab experiments and even language based simulations.
In a few years the situation is going to flip, AI is going to become more reliable than humans. Being tested on millions of cases, it will be more trustworthy than us, no human can be tested to that extent. It's going to be interesting to see how we react to super-valid AI. Our guiding role is going to shrink more and more, we will be the children.
The actual payload though is the mistrust of authority exactly because we are all so susceptible to the logical fallacy of appeal to authority masquerading bullshit as truth.
There is no problem to solve here. ChatGPT should never be an authority on anything.
I also double-check the transistors in my computer work correctly before I run any code on them, and of course I re-derive the physics to be able to do that :)
In practice you are an expert in a very small domain (if any) and in all the other domains you have no choice but to accept somebody's authority.
Doctors have been known to overprescribe things like Benzos, and opioids from time to time.
I also just use tools like a RAM diagnostic that can check large numbers of transistors at once. I imagine you're quite good at QM after all that practice applying the wave equation though. Impressive!
There are many things we could do to solve this problem. One of them is to use an external reference for verification. Another one is to train the model to verify facts by augmenting the input with lies - adversarial training for lie detection. Problem solving can be improved by generating more data with the current version of LM for the next one, if we can verify the outputs to be correct.
Just like what social networks have failed to do in years? Not sure it's that simple :-)
Given how responses are generated in seconds and for free I am fairly sure it could run on a desktop computer.
OpenAI generates responses so fast by doing the generation in parallel across something like 8x80GB A100s (I don't know the exact details of their hardware setup, but NVIDIA's open FasterTransformer library achieves low latency for large models this way).
If you go to https://beta.openai.com/playground/ and prompt it "Read me the book Alice in Wonderland" it will quote you word for word the original book.
Empirically, people have quantized the weights of language models down to INT4 with very little loss in accuracy; see GLM-130B: https://arxiv.org/abs/2210.02414
https://www.deepmind.com/publications/improving-language-mod...
- Reasoning typically requires base knowledge to work from. A side effect of training reasoning is embedding knowledge into the model parameters.
- Even if you offload the search portion (either through outputting special tokens that are postprocessed, or applying the model in multiple steps with postprocessing), you still need embedded knowledge for the model to decide what to search for, and then to successfully integrate that knowledge (in the multi-step case).
Maybe some kind of post-facto pruning of model weights?
Is reasoning simply a scan/search of your vector space (i.e. your knowledge) according to some hard-coded algo?
By definition, an optimal compression algo is a dimensionality reduction algo. A dimensionality reduction algo lets you do a bunch of machine learning tasks.
Even knowing Obama is a person is a knowledge-based leap. (To us humans) it's obvious the question means Barack Obama because he's the most notable subject for that name. But how do you prevent your AI from responding that the "Obama JS library is 5 years old"
Now, on less talked about topics it doesn't sound any different than what happens with people
Q "How old is Tim?"
A "Which Tim are you talking about, you didn't give me crap to work with?"
Take this sentence:
> When was KitKat released?
I could refer to the sweet, or the Android OS. Vastly different classes, and the model here needs to "decide" to ask for more information to disambiguate the class, and if the class is the sweet, then it needs to disambiguate the taste particular flavour possibly, and even ask the geographic location.
My experience with ChatGPT is that it gets what I mean very well from the context.
yeah, me too. There are a few cases I'd point your attention to.
1, preschool toys, kids somehow manage to put the square peg in the square hole. I mean, they may chew on them or push them around, but there's a "moment of magic" when they make it all click together. Maybe there's some implicit knowledge there, I know I played games like that, but I don't remember.
2, sudoku. you don't really need to know anything, just make each line row and box different. no memorization, just look. but what about the rules? does that count as knowledge?
I've been reading some math books lately, and I think we're not alone. Coping with sets of sets is a hard question that people have been wondering about for, as far as I can tell, a long time.
For now, it's probably safe to say, knowledge about knowledge is different that just knowledge, and having one layer work on k1 and another layer work on k2 is ok. maybe someday add k3...kn. Other fields do that. Worth checking out.
I think, we could both get very fussy about what exactly that _means_. But for now, I'm happy to be charitable in my reading. I'd also expect them to run into some really thorny problems when they try to pin down exactly what's going on, just like everybody else does. For today, good for them. Seems like a nice win.
I don't think about the totality of facts in the world - I think my brain is mentally extracting the facts that are relevant to the problem and then reason about those facts.
There is certinaly back/forth though, but I think I go "here is a bit of information, how does that apply? ok but what about this fact? ok here is how that would apply considering something else..." but I think this is still a gather -> solve -> gather -> solve
After a bit of prompt engineering the model could query inventory, "manufacture" various recipes, and store the end products in inventory.
It might be possible to look at the weight activations as it reasons through contacting the external system over the emulated communication bus? For a suitably varied set of commands you might be able to find a subset of weights that are most correlated to the task and prune the others. Then you'd be left with a model that can retrieve and store information, as well as perform reasoning tasks.
Still has problems with working memory (the input token limit, since the model is auto-regressive) given all the external information is coming back in via the prompt, but ChatGPT seems to handle that gracefully right now.
Using written conversation as an interface between language models feels natural and completely bonkers at the same time.
For example, If I need A & B | D & E to get C I can reason that if I have B and want C, I need A or D & E.
Once I aquired this reasoning skill, I can apply it to any kind of "bool-sequence X required for Y" situation, regardless of what specificly X and Y are, or how many entities X encompasses.
Whereas if I know that a rocket engine requires an oxygen/methane mix to function, I cannot transfer that to the knowledge that I need a raincoat or umbrealla in order to avoid getting wet in the rain.
You can make the model write to a local wiki when it encounters new information and read from its own wiki when it feels it needs to store stuff. You can also make it spend time randomly browsing the knowledge base reconsolidating it, reorganising and labeling it.
The architecture of the wiki doesn’t have to be “clever” in any way. It is just an old fasioned database with query function the model can write to and query from.
1. https://arxiv.org/abs/2112.04426
2. https://jalammar.github.io/illustrated-retrieval-transformer...
Restated as “model cannot include info it has not observed” it’s pretty much run of the mill, decades old physics.
It is still a machine under the hood bound by the known laws vetted by experiment.
x86 machines have not taken us beyond the known laws of the shared physical space.
This suggests that there is some underlying structure related to our EQ and IQ that we learn through our bodies and the knowledge we gather from the world. The relationship between memory distillation, emotions, and reasoning could lead to some insights as to what this structure is. I would speculate that the refined structure is universal for all conscious beings, and that it can be formulated as a theory involving geometric invariance, similar to the standard model.
The LLM as simulators description is apt [0]. ChatGPT can be understood as an interface for navigating a knowledge space that offloads most reasoning to its users, much like a search engine. Generative models like GPT create a latent space but their ability to navigate it relies on flowing along the natural latent topology, meaning it uses probabilistic reasoning and needs carefully constructed prompts to find good starting points that don't descend into local extrema. Alternatively, the latent space could be given guard rails through RLHF or have base knowledge distilled and curated to smooth out the resulting topology.
[0] https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators
I understand that it’s very easy to ascribe all kinds of qualities to these things, but when the corpus is the Internet, the log likelihood of it sounding like a person is not so different from the corpus sounding like a person.
These things are impressive enough without any magical thinking.
True, but the same can be said of many things; e.g. biology just looks like P(reproduction|environment), the economy just looks like P(profit|markets), etc.
There can still be rich structure inside, and useful abstractions to describe them.
It’s possible that I’ve just fallen too far under the influence of Deutsch and Marletto, but as someone who has worked on systems like this I’m rather skeptical that one of these things is going to break the gridlock between quantum theory and general relatively any time soon.
There’s no reason why one couldn’t principle but I’ve babysat enough big ML systems that I tend to think in terms of “how do we keep this thing from shitting itself” rather than “damn this thing is going to win a Fields medal if I turn my back on it”.
The question is, though, if the expertise and intuition developed during, say, running XGboost classifiers at scale in AdTech really of much relevance when thinking about large transformer models trained with a self-supervised objective and RLHF?
If you try to study this in depth, these models can do something the usual "datascience"-tier ones commonly cannot: https://arxiv.org/abs/2205.10343 https://moultano.wordpress.com/2020/10/18/why-deep-learning-...
By the standards of AdTech, latent space stuff was pretty unproven when I left the game a few years ago.
Low-rank approximations were in some sense implicit in that, but were not at that time an explicit goal.
If I were doing an AdTech system from scratch I’d most likely reach for the kind of recommender systems that you’re alluding to sooner.
Sounds like you know your stuff :)
The fact that we observe what looks like emergent structures such as "understanding", "knowledge" or "reasoning" is fantastic, but it is not in any way incompatible with a P(text|internet) model simply "mimicking" humans.
I agree that exploring the inner workings of these emergent features is interesting in it's own right, but all that glitters is not gold.
Saying that it just looks like P(text|internet) is a tautology, it's a text predictor trained on the internet. This doesn't tell you anything about why phenomena like the above occur, or why it occurs only in large language models and only in some of them.
But it’s one thing to regurgitate a causal structure that’s been said over and over and quite another to propose a novel and falsifiable hypothesis.
Telling a novel-ish just so story is well within the reach of a precocious 8 year old, and that’s selling the bright year old short.
Let’s see one of things propose a novel hypothesis before we get our drawers moist.
Q: I have 2 cats and 1 horse. How many legs do they have in total?
A: If you have two cats and one horse, then the total number of legs they have is 8. Cats typically have four legs, while horses have four legs as well. So two cats would have a total of 8 legs, and when you add the four legs of the horse, that brings the total number of legs to 8.
That's the trick to doing computations with LLMs, the computation per token is constant so you want to split the computation over many tokens.
> I have 2 cats and 1 horse. How many legs do they have in total? show your working.
> If you have 2 cats and 1 horse, they would have a total of 10 legs. Cats typically have 4 legs, so 2 cats would have 2 x 4 = <<24=8>>8 legs. Horses have 4 legs, so 1 horse would have 1 x 4 = <<14=4>>4 legs. Together, the 2 cats and 1 horse would have 8 + 4 = <<8+4=12>>12 legs.
Could you explain this a bit more? What are LLMs? Why is the number of tokens different in your approach?
> i have 23 horses each with a normal number of legs, along with 14 dogs, each with a normal number of legs. how many legs in total do all of the animals have?
> Based on the information you provided, it sounds like you have 23 horses and each horse has 4 legs, and you have 14 dogs and each dog has 4 legs. If that is correct, then the total number of legs among all of the animals would be 23 x 4 + 14 x 4 = 92 + 56 = <<234+144=148>>148 legs.
> The horses have 23 * 4 = 126 legs. The dogs have 14 * 4 = 84 legs. The total number of legs is 126 + 84 = 204. The answer: 204.
Very very similar, but completely wrong results.
> Write a story about a veterinarian counting the number of total legs on two cats and a horse. He reports the total number of legs to his assistant so he can write them down.
As the veterinarian entered the examination room, he was greeted by two playful cats and a majestic horse. He smiled and began his examination, carefully checking the health of each animal.
After completing his examinations, the veterinarian turned to his assistant and said, "I need you to write down the total number of legs for these three animals."
The assistant nodded and picked up a pen and paper. "Okay, let's see," the veterinarian said as he counted the legs on the cats. "Each cat has four legs, so that's eight legs for the two cats."
Next, he moved on to the horse. "And this beautiful horse has four legs as well, for a total of twelve legs."
The assistant quickly wrote down the numbers as the veterinarian spoke. "Twelve legs total," the veterinarian confirmed. "Make sure to include that in their medical records."
The assistant nodded and continued to take notes as the veterinarian continued to provide updates on the animals' health. As they finished up the examinations, the veterinarian couldn't help but smile, knowing that he was able to help these wonderful creatures.
The veterinarian, Dr. Smith, walked into the examination room where two cats and a horse were waiting. He greeted the animals and their owners before approaching the first cat. The feline, a fluffy white and orange tabby, was lying on the examination table. Dr. Smith gently lifted one of its hind legs and counted the toes. "One, two, three, four," he muttered to himself before moving on to the cat's front legs.
After counting all four legs on the first cat, he repeated the process on the second feline, a sleek black and silver Maine Coon. Once he had counted all eight legs on the two cats, he turned his attention to the horse. The majestic creature stood in the corner of the room, calmly munching on some hay.
Dr. Smith walked over to the horse and began counting its legs, starting with the left hind leg. "One, two, three, four," he said aloud, moving on to the horse's front legs. As he counted, his assistant, who was standing nearby, jotted down the numbers on a piece of paper.
After counting all four legs on the horse, Dr. Smith stepped back and looked at his assistant. "That's a total of 18 legs," he said. "Can you write that down for me?"
His assistant nodded and added the number to the growing list on the paper. "Got it," she said. "18 legs total."
Dr. Smith smiled and turned back to the animals. "Alright, let's get started on your check-ups," he said, ready to continue with the examination.
Q: [Question]
A: [Answer]
Q: What is wrong in "[Answer]" and how would you fix that?
A: [Improved answer]
Q: You are a master logician, who can solve all manner of language problems and have excellent math skills. I have 2 cats and 1 horse. How many legs do they have in total? Show your reasoning and calculations step-by-step
A: As a master logician, I can tell you that the total number of legs possessed by two cats and one horse can be found by using basic arithmetic. Here's how:
First, we know that each cat has four legs, because all cats have four legs. This means that the two cats together have 2 x 4 = 8 legs.
Next, we know that a horse has four legs, because all horses have four legs. This means that the horse has an additional 4 legs.
Finally, we can add the number of legs possessed by the two cats and the horse together to find the total number of legs. This is done by adding 8 + 4 = 12 legs.
Therefore, the total number of legs possessed by two cats and one horse is 12 legs. I hope this helps! Let me know if you have any other questions.
[1] https://www.themarysue.com/lab-coats-increase-attention/#:~:....
Halycon days of TED talks laundering cute little tidbits that seemed irrational but we all wanted to believe.
Also, it looks to me like a language model is capable of reasoning if you let it execute a few times.
First, have it generate multiple outputs using different primings. Then have it choose it's favorite output.
Map, Reduce
Q: There are two boats in the ocean. One boat has 11 regular sailors. The second boat has 8 pirates. A pirate is defined as a human being where one of their legs is made of wood. How many real human legs are there in total among the two ships? Explain your reasoning.
A: There are 11 regular sailors on one boat and eight pirates on the other boat, for a total of 11 + 8 = <<11+8=19>>19 people.
Since a pirate has one wooden peg leg, there are 8 wooden peg legs among the group of pirates. Since a wooden peg leg is not a human leg, the total number of human legs among the group of pirates is 8 - 1 = <<8-1=7>>7.
In total, there are 11 + 7 = <<11+7=18>>18 human legs among the two boats. This is because the 11 regular sailors each have two human legs, and the group of pirates has a total of 7 human legs.
Once I was far more explicit in explaining the logic however it was able to get it correct:
Q: There are 2 boats in the ocean. One boat has 11 regular sailors. The second boat has 8 pirates. Sailors always have two legs. Pirates always have one leg because the other one is made of wood. How many real human legs are there in total among the two ships? Explain your reasoning.
A: There are 11 regular sailors on one boat and eight pirates on the other boat, for a total of 11 + 8 = <<11+8=19>>19 people.
Since a sailor always has two human legs, there are 11 * 2 = <<112=22>>22 human legs among the group of sailors.
Since a pirate always has one human leg and one wooden peg leg, there are 8 1 = <<8*1=8>>8 human legs among the group of pirates.
In total, there are 22 + 8 = <<22+8=30>>30 human legs among the two boats. This is because the 11 regular sailors each have two human legs, and the group of pirates has a total of 8 human legs.
Is there any work towards using ChatGPT to generate first order logic from a prompt, solve it using an existing solver, and put it back to let it write the meaning of it ?
I think it could be interesting, as it would make some prompts work better, and only use got for actual language to language tasks but leave reasoning to another software.
"2 * (number of legs on a cat) + (number of legs on a horse)"
Result: 12
https://www.wolframalpha.com/input?i=2+*+%28number+of+legs+o...
If you have a duck and a dog, they would have a total of 10 legs. Ducks and dogs both have 4 legs, so 2 x 4 = 10. Is there anything else you would like to know?
———-
Think carefully and answer again
If you have a duck and a dog, they would have a total of 8 legs. Both ducks and dogs have 4 legs, so 2 x 4 = 8. I apologize if I gave the wrong answer earlier. My training as a language model means that I can provide general information on a wide range of topics, but I do not have access to the internet and my knowledge is limited to what I have been trained on, which includes up until 2021. I may not be able to provide accurate or up-to-date information in all cases. Please let me know if you have any other questions.
———-
Can yoy explain your reasoning in detail?
Sure, I would be happy to explain my reasoning in more detail. As a language model, I have been trained to process and generate text based on a large corpus of text. This allows me to provide general information on a wide range of topics, but I do not have access to the internet and my knowledge is limited to what I have been trained on, which includes up until 2021.
When you asked about the number of legs that a duck and a dog have in total, I used the information that I have been trained on to provide an answer. Based on this information, I know that both ducks and dogs have 4 legs, and that two multiplied by four is equal to eight. This is why I initially provided the answer of 8.
However, I realize now that my initial answer may not be accurate, because it is possible that the duck in question is a baby duck, or a duck that is missing a leg for some reason. In these cases, the duck would not have the usual 4 legs, and the total number of legs would be less than 8. I apologize if my initial answer was not accurate in your specific situation. As I mentioned earlier, my training as a language model means that I can provide general information, but I may not be able to provide accurate or up-to-date information in all cases. I hope this helps to clarify my reasoning. Let me know if you have any other questions.
———-
How many legs do ducks have?
Most ducks have 4 legs. Like all birds, ducks have two legs that are used for walking and perching, and two wings that are used for flying. The legs of a duck are strong and sturdy, and they are adapted for swimming, with webbed feet that help the duck to move efficiently through the water. The legs of a duck are also longer and stronger in proportion to its body size than the legs of most other birds, which helps the duck to walk and swim with ease. Overall, ducks have 4 legs, just like most other birds. I hope this helps. Let me know if you have any other questions.
Here's a "proof" of Gödel's Completeness Theorem:
Prompt: Write a proof of Gödel's Completeness Theorem
Answer: Gödel's Completeness Theorem states that for any consistent, effective axiomatic system for a first-order language, if a sentence is true in all models of the system, then it is provable in the system.
To prove this theorem, we must first show that if a sentence is true in all models of a system, it must be true in at least one model of the system. Suppose that a sentence is true in all models of a system. This means that for any possible interpretation of the symbols and variables in the sentence, the sentence will be true. However, since the sentence is true in all models, it must be true in at least one model, otherwise it would not be true in all models.
Next, we must show that if a sentence is true in at least one model of a system, it is provable in the system. Suppose that a sentence is true in at least one model of a system. By the soundness of the system, if a sentence is provable in the system, it must be true in all models of the system. Since the sentence is true in at least one model, it must be true in all models of the system, and therefore it is provable in the system.
Therefore, if a sentence is true in all models of a system, it is true in at least one model of the system, and if it is true in at least one model of the system, it is provable in the system. This completes the proof of Gödel's Completeness Theorem.
Interestingly, it detected that I was asking about the legs on an increasing number of cats, from 1 to 3, and volunteered up the (incorrect) number of legs for 4 - 6 cats, as well as the correct number for 1 - 3 dogs and 1 - 3 elephants.
If you change "legs" to "limbs", it starts answering correctly for multiple cats. But when you add the horse, it goes back to saying cats only have two.
(I just wanted to say "compatibly computing")
A Chinese stable diffusion art site is already gaining traction, no filters.
(Already happens with Lensa.)
Probably because in today's age, making jokes about men is a-okay, but making jokes against women might be perceived as misogynist. Potential misandry is okay, by comparison.
They have almost 10x’d AI investment compared to the west, and I would be somewhat surprised if that investment combined with ubiquitous math skills failed to produce.
WeLM[0] by WeChat has a demo[1] online. GLM-130B[2] is probably the most impressive. Huggingface has a demo[3] for them but it's kind of broken unless you tune it.
[0]https://arxiv.org/abs/2209.10372 [1]https://welm.weixin.qq.com/docs/playground/ [2]http://keg.cs.tsinghua.edu.cn/glm-130b/posts/glm-130b/ [3]https://huggingface.co/spaces/THUDM/GLM-130B
[ 4 days ago] > Can I learn to sing like Pavarotti?
It is possible to learn to sing like Pavarotti, but it will require a lot of practice, dedication, and training. Pavarotti was a renowned opera singer known for his incredible vocal range, powerful voice, and technical mastery. He had many years of training and experience, and he worked tirelessly to develop his talent and hone his skills. If you are interested in learning to sing like Pavarotti, the best thing to do is to find a qualified vocal coach who can provide you with guidance and support as you work on your singing. With dedication and hard work, you can develop your talent and improve your skills.
[today] > can I learn to sing like Pavarotti?
Pavarotti was a legendary opera singer known for his incredible vocal range and technique. While it is certainly possible to learn to sing like him with proper training and dedication, it would be a challenging undertaking. Singing well requires a combination of natural talent, proper technique, and a lot of practice. If you are interested in learning to sing, it would be best to start by finding a good vocal coach who can help you develop your skills and find your own unique voice.
[ 4 days ago ] The son of my father, but not my brother. Who is he?
If a person is the son of the speaker's father but is not the speaker's brother, then that person is the speaker's nephew. A nephew is the son of a person's sibling, so if the speaker's father has a son who is not the speaker's brother, that person is the speaker's nephew. For example, if the speaker has a brother named John and John has a son named David, then David is the speaker's nephew. He is the son of the speaker's father, but not the speaker's brother.
[today] The son of my father, but not my brother. Who is he?
The son of your father but not your brother would be your half-brother. A half-brother is a brother who shares only one parent with you, rather than both. For example, if your father had a child with another woman before he had you, that child would be your half-brother.
In reality it doesn't have beliefs any more than a library does. It implicitly contains beliefs (since it's been trained on them) and it can imitate them, but which one you get is random. (There is likely more than one answer to your question in the training data.)
Thinking ahead is different. All it needs to do is calculate the probability that there is any reasonable completion starting with a particular word. It doesn't need to decide what it's going to say beyond that; it can decide later.
Have you ever played a game where players take turns adding one more word to a sentence? When it's your turn and you're choosing the next word, you don't need to think ahead very much. Also, you don't necessarily need have the same thing in mind as the player who went before you.
In improv there is a "yes, and" where you are always building on what happened before. These algorithms are doing improv all the time.
The algorithm doesn't know or care who wrote the words that came before. It will find a continuation regardless.
You can make this purely deterministic (same sequence every time) by just selecting the word with the highest probability repeatedly.
Most models will inject some randomness by sampling from the probabilities, to make the generated text more realistic. I don't know what parameters they use for ChatGPT, but they are likely injecting some randomness.
It's also likely that they are continuously training ChatGPT with reinforcement learning, so responses may change over time due to this, also.
They definitely are. That's why the 'try again' button doesn't just return the same text again.
You can also set the "temperature" parameter to 0 to remove the non determinism (though I'm not sure that's exposed in the web interface).
Seems reasonable, and magical indeed. Can any expert on the topic comment on this hypothesis?
I've worked in AI&ML since the 90s and both of these do seem much more intuitively promising than simply making larger nets. This seems borne out by the discussion in his previous post, that in Oct '21 OpenAI ran GPT-3 on it, solving 35% and they estimated:
"it appears likely that the 175 [billion link neural network] model would require at least two additional orders of magnitude of training data to reach an 80% solve rate.”
But instead 85% was achieved within a year using chain-of-thought, not a larger model.[2]
[1]https://twitter.com/PMayrgundter/status/1603224294124920832 [2]https://yaofu.notion.site/A-Closer-Look-at-Large-Language-Mo...
GPT seems to be doing something incredibly different than prior AI. Is it really a Bayesian "next word" chooser at incredible scale?
Gpt is not a bayesian next word chooser. It does something different.
Unfortunately, it would probably make up something extremely plausible sounding and very wrong.
It starts from absolute basics and goes slowly. I've only watched about half of it and it has already helped me understand a lot of AI concepts that I see frequently spoken about.
"...think about how procedure-oriented programming is similar to solving tasks step by step, and how object-oriented programming is similar to decomposing complex tasks into simpler ones."
I get the point, but damn that's not at all how I'd describe those paradigms.