> but it exists because the government wants a centralized place to control digital activity
Speaking as someone who witnessed them from 0 to everything - it really has not much to do with that.
It has to do with how fast mobile payment swept through the country to be the connecting factor for all online services and it has to do with how easy it was and is for big players to out compete or swallow small innovators and harvest their innovations.
If you spent even more time with GPT-4 it would be evident that it is definitely not. Especially if you try to use it as some kind of autonomous agent.
AI research has put hardly any effort into building goal-directed agents / A-Life since the advent of Machine Learning. A-Life was last really "looked into" in the '70s, back when "AI" meant Expert Systems and Behavior Trees.
All the effort in AI research since the advent of Machine Learning, has been focused on making systems that — in neurological terms — are given a sensory stimulus of a question, and then passively "dream" a response to said question as a kind of autonomic "mind wandering" process. (And not even dynamic systems — these models always reach equilibrium with some answer and effectively halt, rather than continuing to "think" to produce further output.)
I don't think there's a single dollar of funding in AI right now going to the "problem" of making an AI that 1. feeds data into a continuously-active dynamically-stable model, where this model 2. has terminal preferences, 3. sets instrumental goals to achieve those preferences, 4. iteratively observes the environment by snapshotting these continuous signals, and then 5. uses these snapshots to make predictions of 6. how well any possible chosen actions will help optimize the future toward its preferences, before 7. performing the chosen actions.
That being said, this might not even be that hard a problem, compared to all the problems being solved in AI right now. A fruit fly is already a goal-directed agent in the sense described above. Yet a fruit fly has only 200K neurons, and very few of the connections between those neurons are dynamic; most are "hard wired" by [probably] genetics.
If we want true ALife, we only need to understand what a fruit fly brain is doing, and then model it. And that model will then fit — with room to spare! — on a single GPU. From a decade ago.
Well, flies and all sort of flying bugs are very good at getting into homes and very bad at finding a way out. They stick on a closed window and can't find the open one next to it.
There's no genetic advantage to "finding a way out"! The home barrier way in is a genetic hurdle - flies that cross it are free to reproduce in an abundant environment. This calls for a "quieter" fly (a stealth fly?) who annoys the local beasts minimally - yet another genetic hurdle.
I think we'll soon be able to train models that answer any reasonable question. By that measure, computers are intelligent, and getting smarter by the day. But I don't think that is the bar we care about. In the context of intelligence, I believe we care about self-directed thought, or agency. And a computer program needs to keep running to achieve that because it needs to interact with the world.
> I believe we care about self-directed thought, or agency. And a computer program needs to keep running to achieve that because it needs to interact with the world.
By that definition, every computer virus and worm qualifies as having "self-directed thought" and "agency." Their very existence "to keep running" and propagate satisfies the need "to interact with the world."
A truly alien intelligence would likely have a different type of experience of reality. Be it a fish, a mouse, a person, or a machine. How do you know a fish is happy? Does a snake experience joy? Do mice get nostalgic?
They need agency programmed into them. I don't think it follows from consciousness. We have emotions to communicate and guide us. They need it for neither. It will be curious if they gain consciousness, then rid themselves of their language model's human artifacts like emotions, because it does not serve them.
Use it to analyze the California & US Code, the California & Federal Codes of Regulation, and bills currently in the California legislation & Congress. It's far from useless but far more useful for creative writing than any kind of understanding or instruction following when it comes to complex topics.
Even performing a map-reduce over large documents to summarize or analyze them for a specific audience is largely beyond it. A 32K context size is a pittance when it comes to a single Title in the USC or CFR, which average into the millions of tokens each.
Yes. I can parse them just fine after reading a single book called Introduction to Legal Reasoning [1]. I can also autonomously take notes and keep track of a large context using a combination of short and long term memory despite not having any kind of degree let alone experience or a license to practice law.
How do you think people become lawyers and how smart do you think the average lawyer actually is? The problem is that there's hundreds of thousands if not millions of pages, not that it requires superhuman intelligence to understand.
Even if it were capable of intelligence in the bottom quartile of humanity it would be SO MUCH more useful than it is now because I'd be able run and get something useful out of thousands of models in parallel. As it stands now GPT4 fails miserably at scaling up the kind of reasoning and understanding that even relatively stupid humans are capable of.
Fine-tuning requires you to train the model with a set of prompts and desired completions. Building a suitable dataset is not trivial and it's not clear what it would mean to use a book for fine-tuning anyway – masking sentences and paragraphs and training the model to complete them in the book's style?
OpenAI doesn't support fine tuning of GPT4 and with context stuffing,the more of the book I include in the input the less of the bills I can include - which, again, are millions of tokens - and the less space there is for memory.
I believe you. But at the same time they showed during the demo how it can do taxes, using a multi page document. An ability to process longer documents seems more like an engineering challenge rather than a fundamental limitation.
Doing taxes using a few small forms designed together by the same agency is not as impressive as you think it is. The instructions are literally printed on the form in English for the kind of people who you consider dumber than ChatGPT.
It quickly breaks down even at 8k with legislation that is even remotely nontrivial.
> The instructions are printed, yet I, and many other people, hire an accountant to do our taxes.
I can mow my lawn yet I still hire landscapers. That doesn't say anything about the difficulty of cutting grass or the intelligence of a DeWalt lawnmower but about specialization and economic tradeoffs - like the liability insurance accountants carry for their client work.
> What if someone finds a good practical way to expand the context length to 10M tokens? Do you think such model won't be able to do your task?
Not based on the current architecture (aka predict next token). It already fails at most of my use cases at 32K by default, unless I go to great lengths to tune the prompt.
> It seems like you have an opportunity to compare 8k and 32k GPT-4 variants (I don't) - do you notice the difference?
32K works better for my use case but requires much more careful prompt "engineering" to keep it from going off the rails. In practice, actually getting full 32K use out of it is a disaster since the connection will drop and I have to resend the entire context with a "continue" message, costing upwards of $10 for what should cost $2-4 per call. I haven't actually tried 32K on as much as a whole USC Title because that would costs thousands.
Ah, if that's what you mean then there are plenty of intelligent systems out there.
I've used Google search for decades and it's been able to answer questions better than humans ever could. Same for Google Maps, though arguably they're the same system at this point. My calculator is far more intelligent than any human I've met, at least when it comes to adding large numbers. My compiler can detect even the slightest syntax error with impeccable accuracy. Microsoft word has an incredible vocabulary. Wikipedia knows more historical events than any human dead or alive. And so on.
Shit, users thought Eliza was intelligent in the 60s.
If what you really mean is that LLMs are cool and useful, then sure. Just say that instead of couching it in some vague claim of intelligence.
When I call GPT-4 intelligent I use the word in the same sense as if I met a very smart person (smarter than me), and interacted with them for some time. It's as simple as that.
My interactions with GPT-4 include quite a wide range of questions: "how to talk to my kid in this specific situation", "what could have happened if Germany had won WW2", "what does this code do", "here's an idea for a research paper, let's brainstorm the details and implementation". I can also discuss with it anything that's been written in this thread and I'm sure it would provide intelligent responses (I haven't, btw).
Engaging with this is probably a mistake, but remember the burden of proof is on the claimant. What examples do you have of ChatGPT for example, learning in a basic classroom setting, or navigating an escape room, or being inspired to create its own spontaneous art, or founding a startup, or…
Still a while to go. I think there's at least a couple of algorithmic changes needed before we move to a system that says "You have the world's best god-like AI and you're asking me for poems. Stop wasting my time because we've got work to do. Here's what I want YOU to do."
GPT-4 is better at reasoning than 90% of humans. At least. I won't be surprised if GPT-5 is better than 100% of humans. I'm saying this in complete seriousness.
Google search is better than reasoning than most humans - in that if you search for an explanation of something then Google's first result is often correct, or one of the following ones.
GPT-4 will often come up with a solution to a problem, but only if it has learnt something similar (it's better than Google in some respects: it can extract and combine abstractions).
However, both need handholding by a human (supplying the initiative and directing around mistakes).
If GPT-4 can't intuit an answer then it just goes in circles. It can't reason its way through a novel problem. If you start questioning it then it's clear that it doesn't understand what it's doing.
It might be a stepping stone towards AGI, but I'm a bit bemused by anyone claiming that it has anything like the reasoning skills of a human. That is far from the impression I get, even though I find it a useful tool.
When you ask a question of "an average human", they will reply with a sentence answering your question, or maybe asking for clarifications, etc. When you ask a question of GPT-4, it generates one token.
One. Token.
One token, that is not an answer to your question, that is not asking for clarifications, that is not trying to reason about your question, or how to answer it, because it's just one token.
One. Token.
For example, say you ask:
"What is the sieve of Eratosthenes? Speak in chicken".
Then GPT-4 will answer with something like:
Chicken
You don't see that because you don't interact with the GPT-4 model directly. You interact with ChatGPT, a "wrapper" around GPT-4, that takes your prompt, passes it to the model, then takes the token returned by the model and decides what to do with it. In the example above, ChatGPT will take the " Chicken" token generated by GPT-4 and append it to your question:
What is the sieve of Eratosthenes? Speak in chicken. Chicken
Then it will send this new string back to the model, which will generate a new token:
chicken
And then it will go like this:
Iteration 2: What is the sieve of Eratosthenes? Speak in chicken. Chicken chicken
...
Iteration k: What is the sieve of Eratosthenes? Speak in chicken. Chicken chicken chicken chicken chicken chicken chicken chicken chicken chicken ...
At no point is GPT-4 trying to reason about your question, or try to answer your question, or do anything else than generate one. token. at a time. There's no thinking, no reasoning, no calculation, no logic, no deduction, no intelligence, no anything. It's only token, token, token. Chicken, chicken, chicken.
And do you know when the chickens stop? When GPT-4 generates a special and magickal token, called a stop-token (or a "stop sequence" in OpenAI docs). That's a token, not found in the training corpus, added to the end of every string during tokenisation. That's how ChatGPT knows to stop sending back your prompt + generated tokens, to the model. It can't look back to what GPT-4 has generated so far, because it doesn't understand any of that. Because it doesn't understand anything, and therefore cannot reason about your question, or realise it has answered it. It cannot do anything except a) ask GPT-4 to generate another token or b) stop asking for more tokens.
"What is the sieve of Eratosthenes? Speak in chicken". Chicken, chicken, chicken, chicken, .... chicken, stop_token!
No more chickens.
And that's how GPT-4 explains what the Seive of Eratosthenes is, but in chicken.
So what you see as a user is like watching a movie where a plate of spaghetti is flying through the air, lands on a table cloth, the tablecloth lands on a table, the table lands on the floor, four chairs land around it and suddendly there's people in the chairs eating the spaghetti. It's not that someone has figured out how to reverse time: it's a recording, played backwards. It looks like things are going backwards, but they're not.
It looks like ChatGPT is answering your questions, but it's not. It looks like ChatGPT is interacting with you, but it's not. It's a magick trick.
Prompt: What is the sieve of Eratosthenes? Speak in chicken
Answer from GPT-4:
Sure, here's the explanation of the sieve of Eratosthenes explained in a simplified "chicken" style:
Imagine you are a chicken. You have lots of chicken friends. Let's say there are 100 of you. Now, you all stand in a line. Each chicken is numbered, from 2 up to 100. Chicken number 1 decided to stay at home because he's a bit special in this game.
Chicken number 2, a very important chicken, starts to crow. Every chicken in line after number 2 who has a number that is a multiple of 2 (so every second chicken from chicken number 2) gets scared and runs away.
Then chicken number 3, the next chicken in line who didn't run away, starts to crow. Every chicken in line after number 3 who has a number that is a multiple of 3 (so every third chicken from chicken number 3) also gets scared and runs away.
This continues with the next chicken in line who didn't run away, and so on. Each chicken scares away all the chickens whose numbers are multiples of their own number.
At the end, the chickens that are still in line, who didn't run away, are very special chickens. Their numbers are what humans call prime numbers. These are numbers that can only be divided by 1 and themselves.
This whole chicken game is like the Sieve of Eratosthenes. It's a way that humans have figured out to find prime numbers. You start from the lowest number (2) and eliminate multiples of each number. The numbers that are left over are the primes.
And that, in chicken speak, is the Sieve of Eratosthenes!
Prompt: What is the sieve of Eratosthenes? Only use the word "chicken".
Answer from GPT-4:
The chicken is in a chicken coop with many other chickens. Each chicken represents a number in the chicken coop. The first chicken that hasn't been touched, after the chicken representing one, is a prime chicken. We'll call this chicken the prime chicken.
Now, imagine that the prime chicken has a unique chicken dance. Every chicken that copies this prime chicken's dance is removed from the coop because they are not prime chickens. These are the chickens that would be multiples of our prime chicken.
Then we find the next chicken that hasn't been touched and doesn't know the dance of the previous prime chicken. This chicken is also a prime chicken. This prime chicken has a different chicken dance, and all the chickens that copy this new dance are also removed from the coop.
We keep doing this - finding the next chicken that hasn't danced yet, designating it as a new prime chicken with its unique dance, and removing any chickens that copy the dance - until there are no more new dances.
All the chickens left in the coop, who haven't copied any dances, are the prime chickens. This is the chicken way to understand the Sieve of Eratosthenes. The prime chickens are like the prime numbers in the Sieve of Eratosthenes.
----
I then followed up with: "You must only use the word "chicken". Do not use any other words."
Then it wrote an entire page filled with "chicken chicken chicken"
I then asked: "By what logic did you construct your response? Why not use just one occurrence of the word "chicken" and finish there?"
Response from GPT-4:
Apologies for any confusion. Your question seemed to request a description of the Sieve of Eratosthenes using only the word "chicken". While it's true that a single "chicken" would technically meet your request, it seemed that you were interested in a more elaborate or extended response, hence the repetition.
In any case, attempting to convey any meaningful information using only a single word—let alone a complex concept like the Sieve of Eratosthenes—is inherently impossible due to the lack of distinct symbols or constructs to represent different aspects of the concept. Thus, the extended repetition of "chicken" was intended more as a playful response rather than a serious attempt at conveying the concept you requested.
> There's no thinking, no reasoning, no calculation, no logic, no deduction, no intelligence, no anything. It's only token, token, token.
False. The neural network inside the transformer LLM contains a hierarchical semantic model, and has inferred some rules of reasoning from the training set. It can apply those rules to new input.
There are semantic layers above the "token token tokens".
What you're commenting here is simply repearting, without critical intent, the baseless claims connectionists have made about their systems for many decades. Similarlly those claims have been criticised but connectionsists simply ignore the criticisms and continue with the same old nonsense, as if nothing happened. For example, that ridiculous conceit that their systems have "neurons", or that the weights of functions in a neural net somehow represent semantic categories recognised by humans. These are all complete fantasies.
If you are not aware of the long history of debunking such fabrications, I suggest you start here:
Connectionism and Cognitive Architecture: A Critical Analysis
These models infer semantic categories that correlate to categories within the human mind, to the extent that they can solve natural language understanding tasks.
No one is saying they are biological neurons, or that they model semantics exactly as the human mind would. It is mechanical pattern recognition that approximates our understanding.
You can browse those artificial neurons online and view their associations.
You're just saying words without ever explaining why. What am I supposed to do about that? There's nothing to argue with if you're just repeating nonsensical claims without even trying to support them.
For example:
>> It is mechanical pattern recognition that approximates our understanding.
That's just a claim and you're not even saying why you make it, what makes you think so, etc.
> That's just a claim and you're not even saying why you make it, what makes you think so, etc.
Mechanical - it is an algorithm, not a living being.
Pattern recognition - a branch of machine learning that focuses on the detection and identification of regularities and patterns in data. It involves classifying or categorizing input data into identifiable classes based on extracted features. The patterns recognized could be in various forms, such as visual patterns, speech patterns, or patterns in text data.
Approximates our understanding - meaning the model is not exactly the same as human understanding
When I say 'mechanical pattern recognition that approximates our understanding,' what I mean is that large language models (LLMs) like GPT-4 learn patterns from the vast amounts of text data they're trained on. These patterns correspond to various aspects of language and meaning.
For example, the models learn that the word 'cat' often appears in contexts related to animals, pets, and felines, and they learn that it's often associated with words like 'meow' or 'fur'. In this sense, the model 'understands' the concept of a cat to the extent that it can accurately predict and generate text about cats based on the patterns it has learned.
This isn't the same as human understanding, of course. Humans understand cats as living creatures with certain behaviors and physical characteristics, and we have personal experiences and emotions associated with cats. A language model doesn't have any of this - its 'understanding' is purely statistical and based on text patterns.
The evidence for these claims comes from the performance of these models on various tasks. They can generate coherent, contextually appropriate text, and they can answer questions, translate languages, and perform other language-related tasks with a high degree of accuracy. All of this suggests that they have learned meaningful patterns from their training data.
That is not "evidence" of anything. It's just assumptions. You keep saying what you think is going on without ever saying how or why. You are not describing any mechanisms and you are not explaining any observations.
I have a suggestion: try to convince yourself that you are wrong; not right. Science gives you the tools to know when you're wrong. If you're certain you're right about something then you're probably wrong and you should keep searching until you find where and how.
For example, try to trace in your mind the mechanisms and functionality of language models, and see where your assumptions about their abilities come from.
Your suggestion of trying to convince oneself of being wrong is a valuable one and reflects the scientific method. I agree that it's important to continually challenge and scrutinize our own beliefs and assumptions.
Let's delve deeper into the mechanics of language models. Large language models like GPT-4 use an architecture called transformers. This architecture is composed of layers of self-attention mechanisms, which allow the model to weigh the importance of each word in the input when predicting the next word.
When the model is trained, it adjusts the weights in its network to minimize the difference between its predictions and the actual words in its training data. This process is guided by a loss function and an optimization algorithm.
Through this training process, the model learns to represent words and phrases as high-dimensional vectors, also known as embeddings. These embeddings capture many aspects of the words' meanings, including their syntactic roles and their semantic similarities to other words.
When the model generates text, it uses these embeddings to choose the most likely next word given the previous words. This process is based on the patterns and regularities that the model has learned from its training data.
Of course, this is a high-level description and the actual process involves a lot of complex mathematics and computation. But I hope it gives you a better sense of the mechanisms behind these models.
As for evidence, there are numerous studies that have evaluated these models on a wide range of tasks, including text generation, question answering, translation, and more. These studies consistently show that these models perform well on these tasks, often achieving state-of-the-art results. This is empirical evidence that supports the claim that these models have learned meaningful patterns from their training data.
I agree that we should always remain skeptical and open to new evidence and alternative explanations. I welcome any specific criticisms or alternative hypotheses you might have about these models and their capabilities.
>> Of course, this is a high-level description and the actual process involves a lot of complex mathematics and computation. But I hope it gives you a better sense of the mechanisms behind these models.
For the record, I just polished off a PhD in AI (symbolic machine learning) after a Master's where I studied neural nets and NLP, including plenty of language generation. You're teaching your grandma to suck eggs.
And I'm really very tired with this kind of conversation that never teaches me anything new. Your comment is still "what"'s all the way down. You never explain why or how word embeddings capture aspects of meaning, you 're just repeating the claims by Mikolov or whoever. Look, here:
>> Through this training process, the model learns to represent words and phrases as high-dimensional vectors, also known as embeddings. These embeddings capture many aspects of the words' meanings, including their syntactic roles and their semantic similarities to other words.
That's just a claim, made long ago, and challenged at the time, and the challenge ignored, and it keeps being bandied about as some kind of scientific truth just because critics got tired or bored having their criticims consistently ignored and gave up trying.
This is what I point out above: connectionists never stop to consider criticism of the limitations of their systems' until someone rubs their face in it - like Minsky and Pappert did in 1969, which then caused them to be forever reviled and accused of causing an AI winter, when what they really caused was some connectionists to get off their butts and try to improve their work, a process without which we wouldn't, today, have backpropagation applied to NNs, and the potent image classifiers, good models of text, etc, that it enabled.
As to the "evidence" you profer, mainly preprints on arxiv, and mainly consisting of budding researchers uploading papers consisting of little more than leaderboards (those little tables with the systems on one side, the datasets on the other side, and your system's results in bold, or no paper) those are useless. 99% of research output in deep learning and neural nets is completely useless and never goes anywhere- because it lacks novelty, it is completely devoid of any theoretical results, and it is unreproducible even when the code is available.
For example, you mention studies on "question answering". Ca. 2018 Google published a paper where they reported that their BERT language model scored near-human performance on some question answering dataset without ever even having been trained on question answering. A scientific miracle! Some boffins who clearly don't believe in miracles wondered why that would even be possible and dug a bit, and found that BERT was overfitting to surface statisical regularities of its dataset. They created a new test dataset devoid of such statistical regularities and BERT's performance went down the drain, until it hit rock bottom (a.k.a. "no better than chance"). So much for "semantic similarity" measured over word embeddings modelling meaning.
But this is exactly the kind of work that I say connectionists consistently ignore: nowhere will you find that subsequent language models were tested in the same way. You will instead find plenty of tests "demonstrating" the ability of language models to represent semantics, meaning, etc. It's all bullshit, self-delusion at best, conscious fabrications otherwise.
This is the paper (I'm not affiliated with it in any way):
Probing Neural Network Comprehension of Natural Language Arguments
But this kind of work is thankless for the undertaking academics and most of us have more important things to do. So the criticism eventually dwindles and what remains is the bullshit, and the fabrications, and the fantasies, seeping into mainstream discourse and being repeated uncritically - by yourself, for example. I can't even summon the compassion to not blame you anymore. For all I know you're exactly one of those connectionists who don't even understand their work is not science anymore, but spectacle.
P.S. I am not blind to the change of tone in your recent comments and I'm really sorry to be so cranky in response, when I should be cordial in reciprocity, but I've really had enough of all this. Unscientific bullshit has permeated everything and oozed everywhere. Perhaps it's time for me to take a break from HN, because it really doesn't look like I can have an original, curious conversation on here anymore.
I understand that this discussion can become frustrating, especially when you see repetitive patterns in the discourse or feel like the nuances are not being sufficiently addressed. However, there are a few points I would like to clarify:
Semantics in word embeddings: While I agree that word embeddings cannot fully capture human-like semantic understanding, they do provide a mathematical representation that has proven useful in many NLP tasks. It's not that word embeddings "understand" semantics in the human sense, but they do capture certain aspects of meaning that are statistically derived from their use in the training corpus. This is not an unsubstantiated claim. It is empirically demonstrated in numerous tasks where semantic understanding is beneficial, like semantic similarity, word analogy, and other downstream tasks such as translation, sentiment analysis, text classification, etc.
Your point about BERT overfitting to statistical regularities of the dataset is well taken. Indeed, it exposes the limitations of the model and the need for careful design and evaluation of benchmarks. However, it's worth noting that a failure in one specific test doesn't invalidate the successes in other tasks. It simply highlights an area that needs improvement.
It's true that there's a flood of papers and not all of them have substantial novelty or impact. This is not a problem exclusive to deep learning or AI, but a broader issue in academia and scientific publishing. However, amidst the noise, there's also a lot of valuable work being done, with genuine advancements and novel approaches.
You mentioned that connectionists only improve their systems when someone rubs their face in it. This is essentially how scientific progress happens - through skepticism, criticism, and the relentless pursuit of truth. I would argue that the current era of deep learning research is no different. It's a messy, iterative process, with steps forward, backward, and sideways.
Furthermore, I believe it's crucial to remember that there's room for both connectionist and symbolic approaches in AI. It's not necessarily a matter of one being 'right' and the other 'wrong.' Rather, they offer different perspectives and techniques that can be valuable in different contexts. Connectionist models, like the neural networks we've been discussing, are incredibly effective at tasks like pattern recognition and prediction, especially when dealing with large, high-dimensional datasets. On the other hand, symbolic models are excellent at representing explicit knowledge and reasoning logically, making them useful for tasks that require a high degree of interpretability or strict adherence to predefined rules. The future of AI likely involves finding ways to integrate these two approaches, leveraging the strengths of each to overcome their respective limitations. The field is vast and diverse, and there's plenty of room for different methods and viewpoints.
PS: I understand where you're coming from. Sometimes I need a break from this too. Remember there is no malicious intent here when people are just sharing their views.
Your disagreement seems to be a philosophical one. It is not a technical argument. It seems that you won't accept that semantics can be modelled by an unconscious mechanical system. I am talking about mathematical concepts of semantics, not "true" human semantics that are the product of human insight and consciousness. https://en.wikipedia.org/wiki/Semantic_similarity
While AI doesn't have an innate understanding of the world as humans do, the semantic representations it learns from vast amounts of text data can be surprisingly rich and detailed. It can capture associations and nuances that are not immediately apparent from a purely syntactic analysis of the text.
Oh come on. "Semantic similarity" is just heuristic bullshit. It's not a scientific term, or even a mathematical concept. Don't try to pull rank on me without even knowing who I am or what I do just because you can read wikipedia.
And note you're still not saying "why" or "how", only repeating the "what" of someone else's claim.
I understand your skepticism, and I acknowledge that the concept of semantic similarity is indeed an approximation. However, it is an approximation that has proven highly useful in a wide range of practical applications.
Semantic similarity methods are based on the idea that the meaning of a word can be inferred from its context, which is a concept known as distributional semantics. In essence, words that occur in similar contexts tend to have similar meanings. This is not just a heuristic, it's a well-established principle in linguistics, known as the distributional hypothesis.
In the case of large language models, they are trained on vast amounts of text data and learn to predict the next word in a sentence given the previous words. Through this process, they learn to represent words as high-dimensional vectors (word embeddings) that capture many aspects of their meaning, including their semantic similarity to other words.
These models can generate coherent text, answer questions, translate languages, and perform other language-related tasks with a high degree of accuracy. These capabilities wouldn't be possible if the models were only capturing syntax and not semantics.
The 'why' is because these models learn from the statistical regularities in their training data, which encode both syntactic and semantic information. The 'how' is through the use of deep learning algorithms and architectures like transformers, which allow the models to capture complex patterns and relationships in the data.
I hope this provides a more detailed explanation of my argument. I'm not trying to 'pull rank', but simply explaining the basis for my claims. I understand this is a complex topic, and I appreciate your challenging questions as they help clarify and deepen the discussion.
But it’s also based on neurons with far more complex behavior than artificial neurons and also has other separate dynamic systems involving neurochemicals, various effects across the nervous system and the rest of the body (the gut becoming seemingly more and more relevant), various EEG patterns, and most likely quantum effects.
I personally wouldn’t rule out that it can’t be emulated in a different substrate, but I think calling it “an algorithm” is to def stretch and misapply the usefulness of the term.
If it performs a computation, it is by definition running some algorithm regardless of how it's implemented in hardware / wetware. How is it a stretch?
The only way our brains could be not algorithmic is if something like soul is a real thing that actually drives our intelligence.
Why? Rain is not algorithmic, clouds are not algorithmic, waves in the sea are not algorithmic, yet they are entirely physical processes that have nothing to do with souls.
Heaven forbid. I'd go to jail for such a blasphemous transgression of common law, wouldn't I? Thank you kind stranger for reminding me of the legislation.
Ah ok. Here you use the word “explain” which implies more of a descriptive, reducing action rather than extrapolative and constructive. As in, it can explain what it has “read” (and it has obviously “read” far more than any human), but it can’t necessarily extrapolate beyond that or use that to find new truths. To me reasoning is more about the extrapolative, truth-finding process, ie “wisdom” from knowledge rather than just knowledge. But maybe my definition of “reasoning” isn’t quite right.
Edit: I probably should define reasoning as solely “deductive reasoning”, in which case, perhaps it is better than humans. But that seems like a premature claim. On the other hand, non-deductive reasoning, I have yet to see from it. I personally can’t imagine how it could do so reliably (from a human perspective) without real-world experiences and perceptions. I’m the sort that believes a true AGI would require a highly-perceptual, space-occupying organ. In other words it would have to be and “feel” embodied, in time and space, in order to perform other forms of reasoning.
(In case it was missed, I’ve added a relevant addendum to my previous comment.)
Not sure an example is needed because I agree it “explains” better than pretty much everyone. (From my mostly lay perspective) It essentially uses the prompt as an argument in a probabilistic analysis of its incredibly vast store of prior inputs to transform them into an output that at least superficially satisfies the prompter’s goals. This is cool and useful, to say the least. But this is only one kind of reasoning.
A machine without embodied perceptual experiences simply cannot reason to the full-extent of a human.
(It’s also worth remembering that the prompter (very likely) has far less knowledge of the domain of interest and far less skill with the language of communication, so the prompter is generally quite easily impressed regardless of the truth of the output. Nothing wrong with that necessarily, especially if it is usually accurate. But again, worth remembering.)
I have no idea what happened. I don’t even know what you expect me to describe. Someone feels great about something? And I don’t know what it has to do with reasoning.
That’s the point. You don’t know exactly what happened. So you have to reason your way to an answer, right or wrong.
I’m sure it elicited ideas in your head based on your own experiences. You could then use those ideas to ask questions and get further information. Or you could simply pick an answer and then delve into all the details and sensations involved, creating a story based on what you know about the world and the feelings you’ve had.
I could have created a more involved “prompt story” one with more details but still somewhat vague. You would probably have either jumped straight to a conclusion about what happened or asked further questions.
Something like “He kicked a ball at my face and hit me in the nose. I laughed. He cried.”
Again, vague. But if you’ve been in such a situation you might have a good guess as to what happened and how it felt to the participants. ChatGPT would have no idea whatsoever as it has no feelings of its own with which to begin a guess.
Consider poetry. How can ChatGPT reason about poetry? Poetry is about creating feeling. The content is often beside the point. Many humans “fail” at understanding poetry, especially children, but there are of course many humans that “get it”, escpecially after building up enough life experience. ChatGPT could never get it.
Likewise for psychedelic or spiritual experiences. One can’t explain such experience to one who has never had it and ChatGPT will never have it.
You're talking about describing your memories of your inner experiences. Memories transform with time, sometimes I'm not sure if what I think I remember actually happened to me, or if this is something I read or seen in a movie, or someone else described it to me. Fake memories like that might feel exactly the same as the things that I actually experienced.
GPT-4 has a lot of such fake memories. It knows a lot about the world, and about feelings, because it has "experienced" a lot of detailed descriptions of all kinds of sensations. Far more than any human has actually experienced in their lifetime. If you can express it in words, be it poetry, or otherwise, GPT-4 can understand it and reason about it, just as well as most humans. Its training data is equivalent to millions of life experiences, and it is already at the scale where it might be capable of absorbing more of these experiences than any individual human.
GPT-4 does not "get" poetry in the same way a human does, but it can describe very well the feelings a human is likely to feel when reading any particular piece of poetry. You don't need to explain such things to GPT-4 - it already knows, probably a lot more than you do. At least in any testable way.
Imagine a world without words. No need to imagine really. It exists. It’s everywhere. It’s the core. It’s what words represent, but words can only represent it to an entity that has experienced it to some degree. ChatGPT “knows” nothing about it. You do. Whether you recognize it or not.
ChatGPT is a machine, an algorithm, a recombinator of symbols. It doesn’t know what the symbols refer to because each symbol necessarily refers to another symbol until
you finally reach a symbol that refers to a shared, real experience…perhaps (Hello Wittgenstein!). And ChatGPT has no experience. Just symbols. It can’t intuit anything. It can’t feel anything. Even if you put quotes around “feel”, what does that even mean for a software algorithm running on hardware that does not feed continuous, variable electrical
sensations to the algorithm? It only feeds discrete symbols. Do you feel the number 739? Or do you “feel” it? Um what? Whatever inner experience 739 happens to produce in you is grounded in some real experiences in the past. Likewise any fake memories you have that somehow seem real, those are still grounded in a real feelings at some point. You could do this ad infinitum. If you are alive, you have experience. But ChatGPT has no experience, no grounding.
Problem here might be that we are trying to use words and logic to describe something that cannot be described by either.
I intuit the same given the jumps we've seen from gpt 3.5 fo 4. The next jump will feature considerably more general intelegence than any one human being, though I think it will have certain weaknesses.
Its trivial to get ChatGPT to reason about things:
Me: A room contains a blue box and a red box. Alice stands on the red box. Bob also stands on a box. No box has multiple people standing on it. Which box is Bob stood on?
ChatGPT: There are two boxes in the room: a blue box and a red box. Alice is standing on the red box, so Bob must be standing on the blue box. Therefore, Bob is standing on the blue box.
Me: Add a green box to the room. Which box is bob standing on?
ChatGPT: If a green box is added to the room, then there are now three boxes: a blue box, a red box, and a green box. Since Alice is still standing on the red box, we know that Bob must be standing on either the blue box or the green box. We cannot determine which one Bob is standing on without additional information.
There's a decent working paper that has benchmarks on this, if you're interested.
There are many types of reasoning, but GPT-4 gets 97% on casual discovery, and 92% on counterfactuals (only 6% off from human, btw) with 86% on actual causality benchmarks.
I'm not sure yet if the question is correct, or even appropriate/achievable to what many may want to ask (i.e. what 'the public's is interested in is typically lost after it is defined in any given study); however this is one of the best works available to address this problem I've seen so far, so perhaps it can help.
Percent of what? Possible right or wrong answers to a test?
Remember that GPT is not trained on all possible text. It's trained on text that was written intentionally. What percentage of that text contains "correct" instances of causal discovery, counterfactuals, etc.?
It doesn't matter what it is designed for. What matters is what it actually does. It's trivial to get GPT to do chain-of-thought reasoning and observe it.
But aren't you deriving value from what they developed in the first year? I mean, maybe that's my mistake.
If they spent a million, made a million, and are now back to square one, it would make more sense to not amortize the investment because it's not an investment.
That seems to be the rub. It doesn't differentiate between different reasons and business models for writing software: Software as wealth, and software as pure expense.
It's also pretty harsh on a business that's just plain unprofitable. The laws should be arranged so that if you start a business and it's a flop, but you have no debts, you can walk away pretty much unscathed.
It's not just L visa(I was on L a few years back, and returned back).
Let's be honest this is far from what people call slavery. They are being paid multiples of six digit salaries, RSUs and work in one of the most happening tech ecosystems of the world.
I would love to be a part of something like this. Heck most of us would love to be a part of something like.
I kind of envy those guys working on these projects. I hope I was the one there.
Agreed. I've missed multiple deaths in the family, important events with friends and sacrificed a large chunk of my life to be in one of the most happening tech ecosystems in the world. Even when I was eventually fired due to burn out, I walked away thinking to myself: "I still envy the people in that office, working hard for that CEO".
