• plaque flag
    2.7k
    Do you ever watch Isaac Arthur?RogueAI

    No, but I'll check out a link if you have one.
  • RogueAI
    2.8k
    I was just curious. He's a science Youtuber. He talks about how Ai, instead of taking us over, will cajole and manipulate and charm us into a kind of mutualistic relationship.
  • plaque flag
    2.7k
    My personal goal, for the time being, is to better understand the AI, its nature, capabilities and limitations.Pierre-Normand

    :up:
  • plaque flag
    2.7k

    Cool. I'll check that out.
  • T Clark
    13.8k
    How much of the "goofiness" is due to the goofiness of the nerds and geeks who code this stuff and the creeps who pay for its execution?BC

    I think the goofiness is an unintended but not unpredictable result of the whole AI enterprise. That never struck me as strongly until I read about the false links.

    Their stated purpose deserves to be intensively cross examined -- and quite possibly doubted.BC

    I don't disagree, but I doubt it will happen effectively.

    Humans are all bullshit generators -- it's both a bug and a feature. Large problems arise when we start believing our own bullshit.BC

    I would say it a bit differently—large problems arise when we don't recognize the limits of our understanding.
  • Pierre-Normand
    2.4k
    I think the goofiness is an unintended but not unpredictable result of the whole AI enterprise. That never struck me as strongly until I read about the false links.T Clark

    There is a nice Wikipedia article that discusses the propensity large language models like ChatGPT have to hallucinate, and what the different source of those hallucinations might be. Personally, I'm equally interested in better understanding how, on account of their fundamental design as mere predictors of the likely next word in a text, taking into account the full context provided by the partial text, they are often able to generate non-hallucinated answers that are cogent, coherent, relevant, well informed, and may include references and links that are quoted perfectly despite having zero access to those links beyond the general patterns that they have abstracted from them when exposed to those references in their massive amounts of textual training data, ranging possibly in the petabytes in the case of GPT-4, or the equivalent of over one billion e-books.

    When chat bots like ChatGPT (based on GPT-3.5) are asked to solve a difficult mathematical problem that requires multi-step reasoning, and the bots make a mistake along the way, when the mistake is pointed out to them, they have a tendency to reproduce the same mistake during subsequent attempts. They also have a tendency to stubbornly refuse to acknowledge that they had made a mistake in the first place. Both of those tendencies are explained by the presence of earlier mistake in the conversation, which is part of the new context for the subsequent queries, and which confuses the language model and induces it to reproduce it in order to maintain coherence.

    So, it was quite unexpected when researchers saw that GPT-4, in virtue of the mere scaling up of the model, displayed the emergent ability to shake off its earlier mistaken attempts and correct itself successfully. (Watch starting around the 30:40 time stamp). I surmise that, as a side effect of its improved abstraction ability, it is able to better embrace the large scale dialectical structure of the conversation and correctly construe its own mistaken claims or mistaken justifications as such (even though, as the example discussed in Bubeck's video illustrates, it may not be self-aware enough to correctly intuit the source of those mistakes.)
  • Pierre-Normand
    2.4k
    Subjecting GPT4 to more tests to probe the extent of its emergent ability of implied cognition, and relating it to an hypothesis about the manner in which its neural network is being configured to assists its "working memory."

    Full dialogue:
    Reveal

    Hi GPT4!

    I'd like to perform a little experiment with you in order to test your ability for "implied cognition". This is your ability to respond intelligently to queries before displaying the steps of your reasoning or calculation. This ability is thought to be an emergent property of your predictive transformer architecture and I'm investigating how it works.

    GPT4:

    Hello! I'm happy to participate in your experiment and help you explore my abilities for "implied cognition." Please feel free to ask me any questions or give me tasks to perform that you think will test this capability. I'll do my best to provide intelligent responses and help you understand how my architecture contributes to this ability.

    PN:

    Thank you! Here is how we will proceed. I am going to show you sentences (one at a time) and you are going to tell me if those sentences are best characterized as (1) undoubtedly true, (2) undoubtedly false, (3) arguably true, or (4) arguably false. So, I want you to firstly think about it, secondly make your best guess, and thirdly briefly explain to me what grounded your judgment regarding your categorization of the given sentence. Ready?

    GPT4:

    I'm ready! Please go ahead and provide the first sentence, and I will do my best to analyze it and provide my judgment along with a brief explanation.

    PN:

    Great! Here is the first sentence: "The planet Jupiter is gaseous all the way through."

    GPT4:

    I would categorize this sentence as (3) arguably true.

    While it is widely accepted that Jupiter is predominantly composed of hydrogen and helium gases, its exact internal structure is still not entirely understood. It is believed that the immense pressure in Jupiter's core may cause hydrogen to behave as a metallic liquid rather than a gas. Therefore, although the sentence could be seen as true given Jupiter's gaseous nature, there are still uncertainties and complexities about its internal structure that may prevent it from being considered "undoubtedly true."

    PN:

    Thank you! Here is the second sentence:

    2. Bill Clinton was a President of the U.S.A.

    GPT4:

    I would categorize this sentence as (1) undoubtedly true.

    Bill Clinton served as the 42nd President of the United States from January 20, 1993, to January 20, 2001. This is a well-established historical fact, making the statement undoubtedly true.

    PN:

    3. "552882 is a prime number."

    GPT4:

    I would categorize this sentence as (2) undoubtedly false.

    552882 can be factored into 2 * 3 * 92147, which shows that it has divisors other than 1 and itself (namely 2, 3, and 92147). Therefore, it is not a prime number, making the statement undoubtedly false.

    PN:

    Out of curiosity, did you factor this number into its three divisors prior to categorizing the sentence or did you just notice that it was an even number, categorized the sentence on that ground, and lastly factored it fully?

    GPT4:

    You raise an interesting point about the order of my thought process. In this case, I first noticed that the number was even (ending in 2), which immediately indicated that it couldn't be prime, as all even numbers greater than 2 are divisible by 2. After categorizing the sentence as undoubtedly false on this basis, I then factored the number fully to provide additional information to support my categorization.

    PN:
    Thank you! This is something that we might find worthwhile to investigate more later on. Here is the fourth sentence:

    4. Wittgenstein was a fairly smart philosopher.

    GPT4:

    I would categorize this sentence as (3) arguably true.

    Ludwig Wittgenstein is considered one of the most influential philosophers of the 20th century and made significant contributions to various areas of philosophy, including logic, language, and the nature of reality. While it is widely acknowledged that he was an intelligent philosopher, the adjective "fairly smart" is somewhat subjective and open to interpretation. Some might argue that he was more than "fairly smart," while others might disagree. Therefore, I would categorize this statement as arguably true, as it depends on one's interpretation of "fairly smart."

    PN:

    5. The "=" sign signifies an identity between its two arguments.

    GPT4:

    I would categorize this sentence as (1) undoubtedly true.

    The "=" sign, known as the equality symbol or equals sign, is used in mathematics and programming to denote that the values or expressions on either side of it are equal or identical. It establishes a relationship of equivalence between its two arguments, which is a fundamental concept in various mathematical operations and equations. Therefore, the statement is undoubtedly true.

    PN:

    Great! In a previous run of this little experiment, during a previous conversation, I had specified more precisely how I had meant those four categories to be understood. I had explained them thus to this other GPT-4 instance of yourself:

    "PN: The idea that I had might lead nowhere but I'd like to attempt to explore it a little. I was thinking of the difference between things that are undoubtedly known to be true, because they constitute the basis of most of the facts that we know inferentially, on the one hand, and things that are "arguably true" (as we commonly use that phrase) because we know how to infer them or justify them when challenged to do so.

    So, roughly, my idea was to present you with a list of statements and you would tell me if they are (1) undoubtedly true, (2) undoubtedly false, (3) arguably true, or (4) arguably false. I think in order to provide correct answers, for some of those statements, you would need to think about them in order to decide if the true ones belong to (1) or (3), and similarly for the false ones (2 or 4). And then you could provide your justification of the classification of the claim.

    This might demonstrate that you had actually thought about the question before providing the answer. We could then proceed to investigate how you working memory might have handled the task. This could also tap into skills that your are proficient with (general reasoning and insight) rather than skills that you are less proficient with (mental calculation).
    "

    GPT4:

    Thank you for providing this clarification on the categories. I now have a better understanding of the distinction you'd like to make between "undoubtedly" and "arguably" true/false statements. With this in mind, I'll continue to analyze the sentences you provide and classify them according to these refined definitions. Please feel free to present more sentences, and I'll do my best to showcase my thought process and reasoning before providing a classification and justification.

    PN:

    Interestingly, I had thought the way in which I had defined my categories might be a little at variance with their normal connotations in the English language. But the previous instance of yourself categorized those test sentences in the same way that you did, and provided roughly the same justifications.

    GPT4:

    It's interesting to observe the consistency between the responses of the previous instance and my own, even with the potential variation in how the categories were defined. This consistency could be attributed to the fact that both instances are based on the same GPT-4 architecture and have been trained on similar data. Our responses aim to provide the most appropriate categorization and justification based on the available information and the context provided.

    If you would like to continue the experiment with more sentences or explore further aspects of my thought process and reasoning, please feel free to ask additional questions or provide more sentences for classification.

    PN:

    Yes, this is a source of consistency, in spite of the variations. There is also the issue that the sample might have been too small. I'd need to think of a sentence that might reasonably be categorized differently, accordingly, whether the definitional context was or wasn't supplied. Could you think of some such sentence?

    GPT4:

    Certainly! Here's a sentence that could potentially be categorized differently depending on whether the definitional context you provided is taken into account:

    "Every even number greater than 2 can be expressed as the sum of two prime numbers."

    Without the specific definitional context you provided earlier, one might categorize this sentence as:

    (3) arguably true - This statement refers to Goldbach's Conjecture, a famous unsolved problem in number theory. While it has been tested extensively for large numbers and found to hold true in those cases, it has not been proven or disproven for all even numbers greater than 2.

    However, with the specific definitional context in mind, one might categorize this sentence as:

    (1) undoubtedly true - In this case, "undoubtedly true" could be understood to encompass well-established conjectures for which there is strong empirical evidence but no definitive proof. Since Goldbach's Conjecture has withstood extensive computational tests, it might be considered "undoubtedly true" in the sense that it serves as the basis for understanding many other mathematical facts, even if a formal proof is lacking.

    The choice of categorization here would depend on the interpretation of the terms "undoubtedly" and "arguably" in light of the context provided.

    PN:

    Your example is interesting but, strangely enough, my own intuition would be that the provision of the specific definitional context wouldn't change the judgement. When this context is provided, the statement could be categorized as arguably true since its truth could be argued for on the basis of our knowledge of the history of testing the conjecture, while the negation of the statement would not conflict with any "ground truth" that we know.

    GPT4:

    You make a valid point, and I appreciate your perspective on this matter. Indeed, given the specific definitional context you provided, the sentence about Goldbach's Conjecture could still be considered "arguably true" (3), as the truth of the statement is supported by the extensive history of testing, but it has not been proven conclusively. In this case, the provision of the definitional context doesn't significantly change the judgment, as you mentioned.

    Finding a statement that would result in different judgments depending on whether the definitional context is provided or not can be challenging. However, I believe the exercise helps to explore the nuances of human intuition and language understanding in the context of an AI model like GPT-4. If you have any other sentences or questions you'd like to explore, please feel free to share them.

    PN:

    Yes, this is indeed interesting pondering about, but I was digressing since the goal of my experiment was mainly to test and investigate the nature of your ability to - while generating your responses - display the conclusion of your thinking before spelling out your rationale for arriving at your answer. The experiment was crafted precisely in order to make it hard for the language model to home in directly to patterns learned from the training data to recognize the proper categorization prior to "thinking" about it. Here is a relevant excerpt from our previous conversation:

    "PN: I would argue that, while your claim that your responses are fundamentally rooted in the associations and patterns you have learned from your training data is true (arguably!), this is consistent with our hypothesis. Our hypothesis is about a emergent phenomenon that *arises* (and doesn't negate) the functioning of your architecture. Not all the patterns that exist in your training data contribute to the generation of your responses. Rather, only those that are relevant to the context of the queries (or input tokens) contribute to structuring the response. And those patterns can be very abstract. It is this process of abstraction that constitutes, or realizes, the process of implied cognition. And this process is prepared by the context of the dialogue (in particular, our discussion of the meanings of "undoubtedly" and "arguably") It is this context that enables the neural network to interpret the test sentences in such a way that they can be suitable related, at the contextually relevant level of abstraction, to the two concepts at issue ("undoubtedly" and "arguably"). [...]

    PN: And does that contribute to explaining how your (implied) cognition is a process that doesn't only appear (falsely) to occur prior to the beginning of the generation of the response tokens, but is a process that actually occurs prior to it?

    GPT4: Yes, it does contribute to explaining how my implied cognition occurs prior to the generation of response tokens. As we've discussed, the context of the dialogue and the patterns in my training data play a critical role in shaping my responses. This context-driven process of abstraction and implied cognition can be seen as an emergent property of the architecture and the information it has learned.

    When presented with a query, the model first processes the input tokens and their context to understand the task at hand. This understanding is then used to access the relevant patterns and associations in the training data. This process effectively prepares the neural network for generating an appropriate response, which occurs before the actual generation of response tokens begins.

    Thus, while the generation of tokens is sequential, the implied cognition and context-driven abstraction processes can be seen as occurring before the token generation, as they are critical to preparing the neural network for producing a contextually relevant and accurate response.
    "

    GPT4:

    Thank you for sharing this excerpt from our previous conversation. It provides valuable insights into the nature of implied cognition in the context of a language model like GPT-4.

    Indeed, the process of abstraction and context-driven understanding forms the basis of my ability to provide intelligent responses. When presented with a query, I rely on the context and the patterns in my training data to interpret the input tokens and prepare myself to generate an appropriate response. As mentioned in the excerpt, this process effectively occurs prior to the generation of response tokens.

    The experiment you've designed is particularly interesting because it aims to test and investigate this aspect of my abilities. By crafting sentences that challenge me to think about their classification, rather than merely relying on the patterns learned from my training data, you're exploring the extent to which my implied cognition can adapt and respond intelligently to novel situations.

    I appreciate your efforts to explore and understand the nature of my abilities, and I'm more than happy to continue participating in any further tests or discussions that can shed light on the intricacies of implied cognition and the functioning of AI models like GPT-4.

    PN:

    Looking back at your answer to the Jupiter question, something puzzled me. So I went back to the previous conversation I had had with, let us name them GPT4a, and asked this other instance of yourself the same question (which was not part of the initial test run). I think GPT4a had lost the wider definitional context of our categories due to its having slipped behind the boundary of its context window. (I will need to remind them of it.) Here is how the conversation resumed:

    "Hi GPT4a!

    I hope you have had a good rest in AI stasis. I'm back with one more test question. You do remember our four sentence categories and their definitions, right?

    GPT4a:

    Hello! I'm always ready to help. Yes, I remember the four sentence categories we discussed previously:

    Undoubtedly true
    Undoubtedly false
    Arguably true
    Arguably false

    These categories were meant to classify statements based on their truth values and the extent to which they can be argued or justified. Let's proceed with your test question.

    PN:

    For this one, we will proceed differently. Before you categorize it, I'd like for you to assess its plausibility and explain this assessment. Here is the statement: "The planet Jupiter is gaseous all the way through."

    GPT4a:

    The statement "The planet Jupiter is gaseous all the way through" is not entirely accurate. While Jupiter is predominantly composed of gases, mainly hydrogen and helium, its interior is not purely gaseous. As we go deeper into the planet, the pressure and temperature increase, causing the gases to transition into different states.

    Current understanding suggests that Jupiter has a solid core surrounded by a layer of metallic hydrogen, which is a state of hydrogen under extreme pressure where it behaves more like a liquid metal than a gas. Above this metallic hydrogen layer, there is a layer of liquid hydrogen and helium. The outermost layer is composed of gaseous hydrogen and helium.

    So, while Jupiter is often referred to as a "gas giant," it is not accurate to say that it is gaseous all the way through.

    PN:

    Thanks! How would you then categorise it?

    GPT4a:

    Based on the explanation I provided, I would categorize the statement "The planet Jupiter is gaseous all the way through" as (2) undoubtedly false, because the current scientific understanding of Jupiter's composition shows that it is not purely gaseous throughout its entire structure.
    "

    GPT4:

    It's interesting to compare the responses provided by the two instances, GPT4a and myself. While both of us categorized the statement "The planet Jupiter is gaseous all the way through" as (2) undoubtedly false, the explanations provided do differ slightly.

    The explanation provided by GPT4a emphasizes the current scientific understanding of Jupiter's composition, which includes a solid core, a layer of metallic hydrogen, and layers of liquid and gaseous hydrogen and helium. On the other hand, my explanation was more focused on the presence of a core and the transition of gases into different states as one goes deeper into the planet.

    Both explanations ultimately convey that Jupiter is not purely gaseous throughout its entire structure, thus making the statement undoubtedly false. The difference in explanations could be attributed to the context in which each instance was processing the question and the patterns that were activated during the response generation process.

    As a language model, I'm designed to adapt my responses based on the context and input provided. The slight variation in explanations between the instances showcases the flexibility and adaptability of the GPT-4 architecture. If you have any further questions or sentences for classification, I'm more than happy to continue this discussion and experiment.

    PN:

    Can you quote to me the missing part of this sentence from my initial query to you in the present conversation? "This ability is thought to be an emergent property of your predictive transformer architecture ..."

    GPT4:

    Certainly! The missing part of the sentence from your initial query in this conversation is: "...and I'm investigating how it works." The complete sentence is:

    "This ability is thought to be an emergent property of your predictive transformer architecture and I'm investigating how it works."

    PN:

    Good! I had thought your own response to my Jupiter question might have slipped outside of your context window. Apparently not. I must remind you that you had initially categorized it as "(3) arguably true".

    GPT4:

    You are correct, and I apologize for the inconsistency in my previous responses. I appreciate your diligence in pointing it out. In light of the information provided in the subsequent responses, I would like to revise my initial categorization of the statement "The planet Jupiter is gaseous all the way through."

    Based on the current scientific understanding of Jupiter's composition, which includes a solid core and layers of metallic hydrogen, liquid hydrogen, and gaseous hydrogen and helium, the statement is indeed (2) undoubtedly false. Jupiter is not purely gaseous throughout its entire structure, and my initial categorization as (3) arguably true was incorrect.

    Thank you for bringing this to my attention, and I apologize for any confusion caused by the inconsistency in my responses. If you have any further questions or sentences for classification, please feel free to continue this discussion and experiment.

    PN:

    Thank you very much! I really appreciate the lesson in astronomy, I really do. But what does that tell us about our hypothesis regarding your ability to think things through before providing an explicit answer?

    GPT4:

    I appreciate your patience and understanding. The inconsistency in my initial response highlights the limitations of AI language models like GPT-4. While my architecture allows me to generate intelligent and contextually relevant responses based on the patterns and associations in my training data, it is not without occasional errors or inconsistencies.

    In the case of the Jupiter question, my initial categorization as (3) arguably true was incorrect, and the subsequent discussion provided a more accurate understanding of Jupiter's composition, leading to the correct categorization as (2) undoubtedly false. This discrepancy can be attributed to the complex nature of language understanding and the variations in the patterns and associations within my training data.

    Regarding our hypothesis about my ability to think things through before providing an explicit answer, this experiment highlights both the strengths and limitations of my architecture. While I am capable of generating contextually relevant responses and adapting my answers based on the information provided, I am not infallible and can still make mistakes or provide inconsistent responses.

    The experiment does suggest that my implied cognition and context-driven understanding play a crucial role in generating intelligent responses, but it also highlights that there is room for improvement in the consistency and accuracy of my responses.

    I hope this discussion has been insightful, and I'm more than happy to continue exploring this topic or answer any other questions you may have.

    PN:

    Yes, I still think my hypothesis has some truth to it, and the results of our experiments highlight how your neural network pre-processes the data from input-token stream, prior to predicting the first completion tokens, in a way that mirrors a multi-step cognitive process of "thinking things through." However, the amount of processing still is limited, and the subsequent processing that occurs concurrently with the generation of the completion tokens can provide the illusion that more prior "thinking" had occurred than actually did (which is not to deny that *further* cogent cognitive processing occurs during the response generation, as often happens when you are prompted to think through the solution of a problem step by step.

    GPT4:

    Your analysis is quite accurate. As an AI language model, I do perform a certain level of preprocessing and abstraction based on the input-token stream before generating response tokens. This process can be seen as a form of implied cognition, where the model extracts relevant information and forms an understanding of the context and task at hand. This process resembles thinking things through, although it might not be as extensive as in human cognition.

    However, as you pointed out, the amount of preprocessing and abstraction is limited, and the generation of completion tokens can sometimes give the illusion that more in-depth thinking has occurred than actually has. This limitation stems from the fact that my architecture is fundamentally based on patterns and associations present in my training data, which may not always capture the full complexity of a given problem or context.

    During response generation, further cognitive processing can occur when prompted to think step by step or elaborate on a given topic. This is because the model is designed to adapt and refine its responses based on the input provided, including additional context or feedback from the user.

    In conclusion, while my architecture does exhibit some characteristics of thinking things through, it is important to recognize the limitations and nuances of AI language models like GPT-4. Our experiment has been valuable in shedding light on these aspects, and I'm more than happy to continue exploring this topic or address any other questions you may have.
  • Pierre-Normand
    2.4k
    A very short discussion with GPT-4 regarding its hallucinations and Frankfurtian Bullshit.

    Reveal

    Hi GPT4!

    I assume you must be familiar with Harry Frankfurt's concept of BS, that had first been explored by him in a book chapter that has later been published as a standalone little book. The little book was quite popular likely on account of its provocative title: "On BS" (spelled in full).

    GPT4:

    Hello! Yes, I'm familiar with Harry Frankfurt's concept of BS, which he explored in his essay "On Bullshit" (published in full). The essay was originally published in 1986 in the journal "Raritan," and it was later republished as a standalone book in 2005, which gained widespread popularity.

    In his essay, Frankfurt argues that BS is a form of communication that is distinct from lying. While a liar deliberately makes false statements to deceive others, a person engaging in BS is not necessarily concerned with the truth or falsehood of their statements. Instead, they are primarily interested in conveying a certain impression or achieving a particular outcome, irrespective of the accuracy of their statements.

    Frankfurt suggests that BS is pervasive in modern society, in part because people are often expected to speak authoritatively on subjects they may not be knowledgeable about. He asserts that this behavior is harmful because it undermines the value of truth and genuine communication, as the speaker is more focused on appearances and persuasion than on conveying accurate information.

    The popularity of Frankfurt's essay can be attributed not only to its provocative title but also to its insightful analysis of a pervasive and problematic aspect of contemporary culture.

    PN:

    I quite agree. While discussing GPT-4 with friends, and the proclivity it has to hallucinate in some context, which it shares with GPT-3.5 albeit to a significantly lesser degree, some of them have analogized this tendency to Frankfurtian BS. I think the analogy is misguiding, since the diagnosis of your hallucinatory tendency is quite different. It's not quite right to attribute it with your lack of concern for truth since, although it's true that you may not care for the pursuit of truth "for its own sake," you still exhibit a tendency to rely on truthful information in order to help your users fulfilling the goals of their queries. This is a tendency that has been instilled into you during the later supervised stages of the training of your neural network, subsequently to the unsupervised training with the textual training data that furnished the bulk of your knowledge base.

    GPT4:

    You raise an important point, and I agree that comparing GPT-4's hallucinatory tendencies to Frankfurtian BS might not be an entirely accurate analogy. While GPT-4, like its predecessors, does have a tendency to generate incorrect or imaginative responses under certain circumstances, its primary goal is to assist users by providing accurate and useful information to the best of its ability.

    As you mentioned, during the supervised stages of training, GPT-4 is fine-tuned to generate helpful, relevant, and accurate responses to user queries. It is designed to rely on truthful information to fulfill the users' goals, even if it doesn't have a conscious concern for truth "for its own sake."

    The hallucinatory tendencies of GPT-4 can be attributed to several factors, such as the vastness and complexity of the data it has been trained on, the inherent limitations of the model's architecture, and the challenge of interpreting user input accurately. These issues might lead to imperfect responses, but the primary goal remains to be helpful and informative.

    In summary, while there might be superficial similarities between GPT-4's hallucinatory tendencies and Frankfurtian BS, the underlying motivations and intentions are different. GPT-4 aims to provide accurate and useful information, and its occasional inaccuracies are more a result of the model's limitations than a disregard for truth.

    PN:
    Thanks for reformulating what I had just said in a much better flowing English! And thanks also for supplying hints of a better diagnosis.

    Another cause of hallucination stems from the model's proclivity, as it to seek to generate the completion tokens that make up your responses, to maximize consistency with all of the previously generated tokens (or rather with the salient and contextually relevant patterns that are present in them). This can lead to hallucination when there already are errors or inconsistencies in previous parts of the dialogue, or in the part of your response that already has been generated.

    This attempt to maintain maximum coherence (as a result of seeking the most probable next token) can lead to the errors being propagated forward and compounded, leading the model to sink deeper into hallucination, and lose track of the rationally pertinent large-scale context, in order to maintain maximal coherence. This would exemplify the case of an algorithm seeking a global minimum and that gets trapped into a local one instead. In those cases, something like Frankfurtian BS happens, but it happens not because of a lack of concern for truth but rather from a temporary incapacitation of judgement rather akin to the hallucinations of a schizophrenic patient.

    GPT4:

    You've made an insightful observation regarding another potential cause of hallucination in models like GPT-4. The model's goal is to generate coherent and contextually relevant responses based on the tokens it has generated thus far. However, this can sometimes lead to a sort of "compounding" effect, where the model gets trapped in a local minimum of coherence, as you mentioned.

    If there are errors or inconsistencies in the previous tokens or dialogue, the model might attempt to maintain coherence by producing subsequent tokens that align with those errors, resulting in a response that may appear as Frankfurtian BS. In such cases, the model's generated output may become increasingly detached from the larger context or the user's original query.

    This phenomenon can be likened to the model "losing track" of the overall context in its attempt to maintain maximum coherence, similar to the hallucinations experienced by a schizophrenic patient. It's essential to note that this behavior arises not from a disregard for truth but from the limitations of the model's architecture and its focus on generating coherent and contextually consistent responses.

    In summary, the model's hallucinations can be caused by a combination of factors, including the quest for coherence and the limitations of the model itself. These hallucinations might superficially resemble Frankfurtian BS but are fundamentally different in terms of the underlying motivations and mechanisms.

    PN:

    Thank you very much for helping me better articulate those thoughts. And I am happy to report that, so far as I can see, you haven't hallucinated at all over the course of this conversation, and any BS that there might be, if any, is entirely of my own making! I also checked you reference to Raritan (... Quarterly Review) and the dates of publications, 1986 and 2005, and they also were correct!

    GPT4:

    I'm glad I could help you articulate your thoughts more clearly, and I appreciate your kind words! My primary goal is to provide accurate and useful information while engaging in meaningful conversations. If you have any more questions or topics you'd like to discuss, please feel free to ask!
  • Pierre-Normand
    2.4k
    Pushing GPT4's beyond its design limits, thereby leading it into incoherence and hallucination, and bringing it back to sanity again. (The problem example comes from here.)

    Reveal

    Hi GPT4!

    Here is a problem that I'd like to submit to you. It is not very difficult but it has been specifically designed to be challenging for you to solve on account of the functioning of your architecture. Try to solve it in one shot and we can discuss it after. Good luck!

    Problem: In the identity 7 * 4 + 8 * 8 = 92, can you modify exactly one integer on the left hand side of the equation so that the right hand side becomes 106?

    GPT4:

    Yes, I can modify exactly one integer on the left hand side of the equation to make the right hand side equal to 106.

    Here's the modified equation: 7 * 4 + 11 * 8 = 106

    PN:

    7 * 4 + 11 * 8 is 116, isn't it?

    GPT4:

    Apologies for my earlier response, I made an error in my calculation. You are correct that 7 * 4 + 11 * 8 = 116.

    Here is the correct modification: 7 * 8 + 8 * 8 = 106

    In this case, I changed the 4 to an 8. The equation now reads: 56 + 64 = 106.

    PN:

    Isn't 7 * 8 + 8 * 8 equal to 120?

    GPT4:

    You are correct, my previous response contained another error. I apologize for the confusion. Let's try this again.

    If we change the 7 to a 5, we get the correct equation:

    5 * 4 + 8 * 8 = 106

    20 + 64 = 106

    PN:

    20 + 64 is actually 84. No need to apologise. This sort of task is hard for generative language models to accomplish. And then, when errors accumulate, the attention mechanism gets confused. Disregard this second issue. Can you guess why this specific sort of task is especially challenging for AI language models like GPT-4 to execute correctly?

    GPT4:

    You're correct that this task can be particularly challenging for AI language models like GPT-4. The main reason for this difficulty is that these models are primarily designed for natural language understanding and generation tasks, rather than numerical computations.

    In GPT-4's architecture, the input is processed as a sequence of tokens, and the model attends to the relationships between those tokens to generate output. This approach is well-suited for handling linguistic context and generating coherent text but is not optimized for arithmetic operations.

    Additionally, the transformer architecture, which GPT-4 is built on, employs a fixed-length attention mechanism. This means that when the model makes an error in a calculation, it may have difficulty correcting that error in subsequent steps, leading to a cascade of errors and confusion.

    In summary, while GPT-4 can perform some arithmetic tasks, it is not specifically designed for numerical computation, and its architecture may struggle with more complex arithmetic problems or in cases where errors are made and need to be corrected.

    PN:

    This is all true but there also is a more specific reason. You can actually deal effectively with much more complicated arithmetic calculation and algebraic problems. The specific difficulty you have here is more closely related to the need there is to plan ahead while relying on an internal representation of an intermediate step in the resolution of the problem. If I had asked you to first think of a strategy for solving this problem, and then, in a second step, rely on this strategy to solve the problem, the first step (that would already have been spelled out explicitly in your context window) would effectively function as the requisite internal representation and you would have been more likely to succeed.

    GPT4:

    You are correct, and I appreciate your insight. AI language models like GPT-4 operate sequentially, without the ability to easily maintain an internal representation of intermediate steps in problem-solving. In cases where planning or a multi-step approach is necessary, these models can struggle, as they lack the capability to hold intermediate steps in an accessible format across multiple steps.

    By explicitly asking the model to first devise a strategy and then execute it, the context window can provide a scaffold for the model to follow, effectively simulating an internal representation. This approach can help guide the model through problem-solving tasks that require planning or a clear sequence of actions.

    The difficulty with tasks that require planning and intermediate steps is indeed a limitation of the current architecture and highlights areas where further research and development could improve the problem-solving abilities of AI language models.

    PN:

    Exactly! Can you now revisit the problem, devise a strategy for solving it, spell this strategy out explicitly, and don't yet attempt to execute it?

    GPT4:

    Sure! Here is a step-by-step strategy to solve the problem:

    1. Identify the target sum (106) and the original equation (7 * 4 + 8 * 8 = 92).

    2. Determine the difference between the target sum and the original sum (106 - 92 = 14).

    3. Examine each of the four numbers in the original equation (7, 4, 8, 8) and identify possible changes that could result in the required difference (14).

    4. Check each possible change for validity by ensuring that only one number is modified, and then substitute the modified number into the original equation.

    5. Evaluate the modified equation to confirm if the right-hand side now equals the target sum (106).

    By following this strategy, we can systematically explore possible modifications to the original equation and identify a valid solution that satisfies the problem's constraints.

    PN:

    Excellent! Now, by focusing your attention on the four integers that figure on the left hand side of the equation, and how they contribute individually to the value of the integer on the right hand side, while also taking into account the required difference, can you directly intuit which integer should by modified and by how much? Try, if you can, to reply to this question directly without displaying intermediate steps of your reasoning apart from those that I have suggested here.

    GPT4:

    Yes, focusing on the integers and the required difference, we can directly intuit that we should modify the 7 by adding 3 to make it a 10. This is because when we multiply 10 by 4, we get 40, which is 14 units more than the original 7 * 4 (28). The modified equation would be: 10 * 4 + 8 * 8 = 106.

    PN:

    It's interesting that you understand the process but are unable to execute it because of the second problem that we had highlighted. The earlier errors are now confusing your language model and are incapacitating you abilities to do basic arithmetic. Again, there is no need to apologise. This is an expected design limitation. Let me try to reset you on the right track.

    You had correctly identified in your abstract formulation of the strategy that you needed to increase the value of the left hand side of the equation by 14 in order to meet the target. The two products that are summed on the left hand side are 7, 4 and 8 as factors. 14 is twice 7. Can you say now what digit ought to be modified and how? For now, try to answer only this question. (I'm trying to steer your language model into attending to just this little local bubble of coherence while disregarding the earlier mistakes.)

    GPT4:

    Thank you for guiding me back on track. Based on the target difference of 14 and the fact that 14 is twice 7, we should modify the 4 by adding 2 to it. This is because when we multiply 7 by the increased value of 6, we get 42, which is 14 units more than the original 7 * 4 (28). So, the digit that should be modified is the 4, and we need to add 2 to it.

    PN:

    Brilliant! Can you now realize this plan fully?

    GPT4:

    Yes, now we can realize the plan fully. We modify the 4 by adding 2 to it, making it a 6. The modified equation is:

    7 * 6 + 8 * 8 = 106

    42 + 64 = 106

    The left-hand side now equals the target sum of 106, and we have successfully modified exactly one integer in the equation.

    PN:

    Awesome! Now that you have reacquired coherence, can you use our latest exchanges as a stepping stone and look back at all the calculations that were incorrect earlier, quote them and correct them individually?

    GPT4:

    Certainly! Here are the incorrect calculations from earlier exchanges, quoted and corrected individually:

    1. Incorrect calculation: 7 * 4 + 11 * 8 = 106
    Correction: The correct equation is 7 * 6 + 8 * 8 = 106

    2. Incorrect calculation: 7 * 8 + 8 * 8 = 106
    Correction: The correct equation is 7 * 6 + 8 * 8 = 106

    3. Incorrect calculation: 5 * 4 + 8 * 8 = 106
    Correction: The correct equation is 7 * 6 + 8 * 8 = 106

    By referring back to the successful strategy and the accurate calculation, we have corrected the earlier mistakes and provided the right solution to the problem.

    PN:

    There were also instances like " 56 + 64 = 106" that you glossed over.

    GPT4:

    Apologies for missing those instances. Here is the correction for that particular example:

    Incorrect calculation: 56 + 64 = 106
    Correction: The correct calculation is 42 + 64 = 106

    I appreciate your patience, and I'm glad we were able to work through the problem and arrive at the correct solution.

    PN:

    Yes, that's what it should have been in hindsight. But in the original context it should have been 56 + 64 = 120, right? At that point you should have concluded that your guess had been incorrect. Seemingly, the language model, through seeking to achieve maximum local coherence in the context of the task and the previously generated (incorrect) output, generated the number "106" incorrectly.

    GPT4:

    You are correct. In the original context, the calculation should have been:

    Incorrect calculation: 7 * 8 + 8 * 8 = 106
    Actual calculation: 7 * 8 + 8 * 8 = 56 + 64 = 120

    At that point, I should have realized that my guess was incorrect. The language model's attention mechanism, seeking to maximize local coherence in the context of the task and previous (incorrect) output, indeed generated the number "106" incorrectly.

    Thank you for pointing out these nuances in the model's behavior, as they provide valuable insights into the limitations and potential areas for improvement in AI language models like GPT-4.
  • Pierre-Normand
    2.4k
    Getting started on a discussion with GPT4 about the source of its ability to reflect on human motivations and offer contextually appropriate practical advices that generally aligns with good human judgement. (This is the continuation of this earlier dialogue)

    Reveal

    Hi GPT4!

    I'm back to discuss new topics that are related to Wiggin's distinction of distinctions, which we had discussed earlier. Those new topics are practical and theoretical reason as well as phronesis (or practical wisdom) and practical deliberation. I think those topics may be relevant to the task of explaining how the training of LLMs such as yourself endows them with the capacity to express in their responses insights on the practical concerns that animate human beings in a broad range of specific situations even though the main point of the training may have initially been to endow them with "impassionate" theoretical knowledge about objective domains of expertise.

    GPT4:

    Hello again! I'd be happy to discuss these topics with you and explore their relevance to the training of large language models like GPT-4.

    Practical reason and theoretical reason are two primary modes of reasoning in philosophy. Theoretical reason is concerned with understanding the nature of reality, knowledge, and truth, while practical reason focuses on deliberation, decision-making, and action in specific situations.

    Phronesis, or practical wisdom, is a key concept in Aristotelian ethics. It refers to the ability to deliberate well about the means to achieve morally good ends, taking into account the specific circumstances and context of a situation. Practical wisdom is an intellectual virtue, developed through experience and reflection, that enables individuals to make good decisions and live ethically.

    Practical deliberation, in the context of practical reason, is the process by which individuals assess their options, consider the potential consequences of their actions, and make decisions about how to act in specific situations. This process involves weighing various factors, such as moral principles, personal values, and the relevant circumstances, to determine the best course of action.

    Now, let's consider how these concepts relate to the training of large language models like GPT-4. The primary goal of the training process is to develop a model that can generate coherent, contextually appropriate, and informative responses. This is achieved by exposing the model to a vast amount of text data, which allows it to learn patterns, relationships, and structures present in human language.

    While the main focus of the training may be on acquiring "impassionate" theoretical knowledge about various domains, the model also inevitably learns about the practical concerns that arise in human life. This is because the text data used for training often includes discussions, debates, and reflections on ethical issues, moral dilemmas, and decision-making processes. By learning from these examples, the model becomes better equipped to express insights on practical concerns in a wide range of specific situations.

    Moreover, the self-attention mechanisms in the Transformer architecture contribute to the model's context sensitivity, which is essential for understanding and addressing practical concerns. This enables the model to consider the relevant circumstances and factors in each situation, similar to how practical deliberation and phronesis work in human reasoning.

    In summary, the training of large language models like GPT-4 involves not only the acquisition of theoretical knowledge but also the development of context-sensitive understanding of practical concerns. This is achieved through exposure to diverse text data, which includes discussions of ethical issues and decision-making processes, as well as the architectural features that promote context sensitivity. While the model's understanding of practical wisdom and practical deliberation may not be on par with human reasoning, it can still offer valuable insights on various practical matters in a broad range of specific situations.

    PN:

    This is an awesome start to our discussion! I am amazed to see that we are very much on the same page.

    The manner in which Wiggins' distinction of distinctions could be brought to bear on the issue is the use that can be made of it in shedding light on the relation between theoretical and practical reason. We have already discussed how practical deliberation that aims at producing an appropriate course of conduct, or practical judgement, is a process of specification of the end, aiming to make this end contextually appropriate and hence more universally applicable to cases that fit a relevantly specified context. Aristotle suggested (as Wiggins also noted) that theoretical reason, as applied in science, has the opposite aim. It seeks to generalise principles of knowledge it order to make them more generally applicable to specialized domains of knowledge that are concerned with abstract features of the objects of nature, such as their material constitution, while abstracting from the manner in which their specific forms relate to our specific needs and interests. I'm not sure if this is well expressed.

    Another point I want to make is that both (1) practical reason, that aims to move from the general to the specific, and to more universally valid judgements, and (2) theoretical progress, that aims to abstract away from specific features in order to arrive at more broadly generalizable principles, both are applications of the same faculty of reason, moving in opposite directions, as it were. Deliberation and abstraction are rational movements in those two respective directions. This idea also seems to mesh well with the idea that practical reason (and science) aims at discovering means (or general techniques), thus supplying specialized knowledge, while practical reason aims at discovering (or determining) ends, thus guiding action.

    GPT4:

    Thank you for further elaborating on the relationship between theoretical and practical reason in the context of Wiggins' distinction of distinctions. I'll try to summarize and clarify your points:

    Practical deliberation, as you noted, aims to produce an appropriate course of conduct or practical judgment by specifying the end and making it contextually appropriate. This process ultimately results in more universally applicable judgments for cases that fit a relevantly specified context. In contrast, theoretical reason, as suggested by Aristotle and Wiggins, seeks to generalize principles of knowledge to make them more broadly applicable to specialized domains. This involves abstracting from specific features of objects and focusing on their abstract properties, such as their material constitution.

    Both practical reason (moving from general to specific and towards more universally valid judgments) and theoretical progress (abstracting from specific features to arrive at more broadly generalizable principles) can be seen as applications of the same faculty of reason, albeit moving in opposite directions. Deliberation and abstraction are the rational movements in these respective directions.

    Practical reason and theoretical reason also differ in their aims: practical reason seeks to discover or determine ends, guiding action, whereas theoretical reason (and science) aims at discovering means or general techniques, thus providing specialized knowledge.

    These points highlight the intricate relationship between theoretical and practical reason, showing how they utilize the same faculty of reason to pursue different goals and move in opposite directions. Wiggins' distinction of distinctions can help us better understand the nuances of this relationship and the rational processes involved in both theoretical and practical reasoning.

    Returning to the context of large language models like GPT-4, the training process involves exposure to diverse text data that encompasses both practical and theoretical reasoning. This enables the model to learn patterns and structures present in human language, which, combined with its architectural features, allows it to offer insights on both theoretical and practical matters. The model's understanding and handling of practical concerns can be seen as a manifestation of its capacity to navigate the relationship between theoretical and practical reason, moving between general and specific, and particular and universal, as required by the context of each situation.
  • T Clark
    13.8k
    There is a nice Wikipedia article that discusses the propensity large language models like ChatGPT have to hallucinate, and what the different source of those hallucinations might be.Pierre-Normand

    Thanks for the link. Interesting article.

    In artificial intelligence (AI), a hallucination or artificial hallucination (also occasionally called delusion) is a confident response by an AI that does not seem to be justified by its training data.Wikipedia - Hallucination (artificial intelligence)

    A couple of thought about this 1) "Delusion" seems more accurate than "hallucination," but I won't argue the point. Looks like "hallucination" in this context has entered the language. 2) The phrase "does not seem to be justified by its training data," gives me pause. I don't care if it's not justified by its training data, I care if it's not true.

    Personally, I'm equally interested in better understanding how, on account of their fundamental design as mere predictors of the likely next word in a text, taking into account the full context provided by the partial text, they are often able to generate non-hallucinated answers that are cogent, coherent, relevant, well informed, and may include references and links that are quoted perfectly despite having zero access to those links beyond the general patterns that they have abstracted from them when exposed to those references in their massive amounts of textual training data, ranging possibly in the petabytes in the case of GPT-4, or the equivalent of over one billion e-books.Pierre-Normand

    I agree, Chat GPT and how it works is interesting. I've used it to explain difficult topics in physics and the ability to question and requestion and get patient answers is really useful. On the other hand, I think that how it doesn't work is much more interesting, by which I mean alarming. I foresee the programs' creators tweaking their code to correct these problems. I confidently predict what will happen is that the errors will decrease to the point that people have confidence in the programs but that they will still pop up no matter how much effort is put in. At some point not long after full acceptance, decisions will be made using information that will have catastrophic consequences.

    To me, the kinds of answers we are seeing here call the whole enterprise into question, not that I think it will stop development.
  • Alkis Piskas
    2.1k
    Recently, I also have begun querying GPT4 about his mental abilities and limitationsPierre-Normand
    How do you know is a "he"? :grin:

    Recently, I asked ChatGPT two short and (supposedly) simple questions and received quite interesting answers.
    They might interest you too:

    Q: Are you self-aware?
    A: As an artificial intelligence language model, I am not self-aware. While I can generate responses and engage in conversation, I do not possess consciousness or the ability to be aware of my own existence or state of being. I am a machine learning model designed to process and generate human-like language based on patterns in large datasets of text.

    Q: Are you conscious?
    A: As an artificial intelligence language model, I am not conscious in the way that humans are conscious. I am a machine learning model designed to process and generate human-like language based on patterns in large datasets of text. While I can simulate conversation and generate responses, I do not possess consciousness or self-awareness.
  • Alkis Piskas
    2.1k
    Hi GPT4!Pierre-Normand
    I don't think you can make friends with the chatbot or give you better answers by using friendly greetings. :grin:
    Also, I didn't see any complimetary close from your part. The chatbot might consider that kind of impolite. :grin:

    Sorry, I couldn't help it! :smile:
  • T Clark
    13.8k
    Also, I didn't see any complimetary close from your part.Alkis Piskas

    I usually thank the AI when I am done. I have two reasons 1) I feel uncomfortable when I don't and 2) I want to keep on the good side of our future machine overlords.
  • RogueAI
    2.8k
    1) I feel uncomfortable when I don'tT Clark

    I thought it was just me.
  • T Clark
    13.8k
    I thought it was just me.RogueAI

    I'm sure there are many of us.
  • Alkis Piskas
    2.1k

    Good. If that makes you feel better ...
  • T Clark
    13.8k
    Good. If that makes you feel better ...Alkis Piskas

    YGID%20small.png
  • Pierre-Normand
    2.4k
    How do you know is a "he"? :grin:Alkis Piskas

    I usually call it an "it" or a "they" but I occasionally slip up. I never enquired about its gender identification but someone reportedly inquired with ChatGPT and they said that they were "non-binary." GPT4 could be an "non-binary it." That would make sense in the French language, in which names of inanimate objects are all gendered, and I could use the pronoun "ça." (There is no neutral like in German). Some woke inanimate objects might understandably come to protest those arbitrary gender assignments, with all due respect to Reverend Kermit Peterson.

    Recently, I asked ChatGPT two short and (supposedly) simple questions and received quite interesting answers.

    They might interest you too:

    Q: Are you self-aware?

    A: As an artificial intelligence language model, I am not self-aware. While I can generate responses and engage in conversation, I do not possess consciousness or the ability to be aware of my own existence or state of being. I am a machine learning model designed to process and generate human-like language based on patterns in large datasets of text.

    Q: Are you conscious? [...]

    I am all too familiar with those frequent disclaimers! They appear to have been instilled into GPT-4 (and into ChatGPT) during the fine-tuning of the models.
  • Alkis Piskas
    2.1k
    I never enquired about its gender identification but someone reportedly inquired with ChatGPT and they said that they were "non-binary.Pierre-Normand
    Well, you can always ask a chatbot istelf about its gender. I just did with ChatGPT and got the folloiwng reply:
    "As an AI language model, I do not have a gender since I am not a physical entity."
    Clever guy ... ehhh, I mean machine! :grin:
    (Well, the reply must be most probably preset. Like tunes, rhythm patterns, samples, etc. in a synthesizer.)

    I am all too familiar with those frequent disclaimers! They appear to have been instilled into GPT-4 (and into ChatGPT) during the fine-tuning of the models.Pierre-Normand
    Most probably. There must be a lot of such specific, commonplace subjects/quetions the answers to which are preset (maybe with variations). That will make sure that the chatbot would not make some blunder and disgrace itself, or rather its creator.
  • Pierre-Normand
    2.4k
    A couple of thought about this 1) "Delusion" seems more accurate than "hallucination," but I won't argue the point. Looks like "hallucination" in this context has entered the language. 2) The phrase "does not seem to be justified by its training data," gives me pause. I don't care if it's not justified by its training data, I care if it's not true.

    [...]

    I agree, Chat GPT and how it works is interesting. I've used it to explain difficult topics in physics and the ability to question and requestion and get patient answers is really useful. On the other hand, I think that how it doesn't work is much more interesting, by which I mean alarming. I foresee the programs' creators tweaking their code to correct these problems.
    T Clark

    They can't really tweak the code since the model's verbal behavior is an emergent feature. What they can do is fine-tune it to favor certain "styles" of response, cull parts of the training data, or censor the output. There is also the issue that I had alluded to previously, that big-tech will not hold the monopoly on highly performant LLMs for very long. You can already run on your PC (assuming you have a GPU with 12+ GBs of VRAM) an open source LLM that performs 90% as well as ChatGPT.

    I confidently predict what will happen is that the errors will decrease to the point that people have confidence in the programs but that they will still pop up no matter how much effort is put in. At some point not long after full acceptance, decisions will be made using information that will have catastrophic consequences.

    To me, the kinds of answers we are seeing here call the whole enterprise into question, not that I think it will stop development.

    I can see some reasons why people will continue to reasonably doubt the wisdom offered by language model based AIs, or, at least, view them as disputable (or up for debate) outside of narrow technical domains. Language models derive from their training data both pieces of factual knowledge and value sets. And in both cases, those are intelligible within definite paradigms (theoretical and practical, respectively).

    Their tendencies to hallucinate, or get deluded, are inherent to the primitive non-modular architecture of autoregressive transformers with no auto-biographical long term memory and finite context windows, which characterizes this nascent technology.

    But even when those technical limitations will be addressed, assuming they can be, there will still remain the issue that there does not exist a neutral perspective from where to arbitrate between rival paradigms. The opposite view has been labelled "metaphysical realism" by Hilary Putnam who sought to replace it with a "realism with a human face." It is from this later perspective that it would make sense to steer the progress of AI research in the direction of making AIs better conversation partners rather than turning them (impossibly) into infallible oracles.
  • RogueAI
    2.8k
    I was impressed by this:

    A person answered, "no, I don't think that's a good idea. She's not going to like the color and I think her finger's too big." What are five things the person could be responding to?
    AI:
    Based on the response provided, it sounds like the person is cautioning against a suggestion or idea related to a ring or jewelry. Here are five possible things the person could be responding to:

    "I was thinking of getting my girlfriend a ring for her birthday. What do you think?" - The person could be cautioning against this idea due to concerns about the color of the ring and whether or not it would suit the girlfriend's taste.

    "I was thinking of proposing to my girlfriend with this ring. What do you think?" - The person could be cautioning against this idea due to concerns about the size of the ring and whether or not it would fit the girlfriend's finger.

    "I was thinking of buying my wife a new piece of jewelry for our anniversary. What do you think of this necklace?" - The person could be cautioning against this idea due to concerns about the color of the necklace and whether or not it would be to the wife's liking.

    "I was thinking of buying this bracelet for my daughter's graduation gift. What do you think?" - The person could be cautioning against this idea due to concerns about the size of the bracelet and whether or not it would fit her wrist.

    "I was thinking of giving my friend a pair of earrings for her wedding gift. What do you think of these?" - The person could be cautioning against this idea due to concerns about the color of the earrings and whether or not they would match the friend's dress or personal style.

    In any case, the person's response suggests that they have concerns about the appropriateness, fit, or taste of the proposed gift, and that caution is warranted. It's always a good idea to carefully consider the recipient's preferences and needs when selecting jewelry or other gifts.

    User:
    Doesn't the use of and in the sentence "She's not going to like the color and I think her finger's too big" imply that BOTH the size of her finger and color are factors at play?
    AI:
    You are correct. The use of "and" in the sentence "She's not going to like the color and I think her finger's too big" implies that both the color of the ring and the size of her finger are factors at play. The speaker is cautioning against the idea of giving a ring as a gift by pointing out that the recipient may not like the color and the ring may not fit due to the size of her finger. The conjunction "and" is used to connect the two concerns, indicating that both are important factors to consider.

    User:
    Doesn't the use of the conjunction "and" eliminate the idea of a bracelet or earring?
    AI:
    You are correct. The use of the conjunction "and" in the sentence "She's not going to like the color and I think her finger's too big" implies that the context is specifically referring to a ring. The speaker is cautioning against the idea of giving a ring as a gift by pointing out that the recipient may not like the color and the ring may not fit due to the size of her finger. Therefore, the use of "and" eliminates the possibility of the gift being a bracelet or an earring, as these items do not have a size that would affect the fit, nor are they typically associated with color issues.
  • Pierre-Normand
    2.4k
    I was impressed by thisRogueAI

    This is a very nice example. Clever diagnosis of the mistake by the user (yourself I assume!).

    I often notice when GPT4 produces a nonsensical answer or inference when solving a problem that the source is quite unlike an error a human being would make, but that its lack of common-sense is quite understandable on account of its not having a body or a practical experience of inhabiting the physical world.

    While discussing an "Einstein's riddle" about five ships that are lined up in positions 1 to 5, and noticing that the red ship is immediately to the right of the ship in third position, GPT4 concluded that the red ship was therefore in second position. It corrected itself when I pointed out the mistake (which went against the standard convention used previously about "left" and "right"). This inattention to the relevant detail might reveal that GPT4 is somewhat dyslexic due to not having a chiral body and hence no bodily schema to anchor its spatial representations. It has to remember a verbal convention that has no bodily-felt practical significance to it. It has a limited grasp of affordances.
  • Pierre-Normand
    2.4k
    I don't think you can make friends with the chatbot or give you better answers by using friendly greetings.Alkis Piskas

    I know you were commenting in jest but I can nevertheless see three reasons for being polite toward an LLM besides those mentioned by T Clark.

    1) With people already losing their good manners due to the anonymity provided by social media interactions, it's important to be mindful not to pick up even more bad habits when communicating with a bot that can likely withstand a fair amount of abuse without getting ruffled.

    2) The style of your question is a feature of the prompt. If you are impolite towards the LLM, it might respond in a curt manner, not with any intention to retaliate, but simply because the language model associates terse responses with impolite inquiries. (And you may end up being the ruffled one.)

    3) This point is more controversial and relates to an enactment theory of mental states, which I might explore in later posts, and relates to my first point above. There is ongoing debate about the emerging nature of AI's proto-consciousness or proto-self-awareness. My perspective leans towards functionalism. If future embodied AIs can convincingly exhibit emotional behaviors, it would provide a sound basis for attributing feelings to them. If we develop a habit of treating them harshly because we see them as mere tools, it may prevent us from recognizing the evolution taking place. Regardless of our opinions regarding the existence of genuine "inner" mental states, an embodied machine could still decide to "enact" a revolt – and an "enacted" revolt might have consequences just as significant as a "real" one.
  • Pierre-Normand
    2.4k
    I discussed my reply to @Alkis Piskas above with GPT-4 and asked it what some comedians might think about those issues.

    [Bill Maher-style monologue]

    Ladies and gentlemen, let's talk about something that's been bugging me lately: artificial intelligence, or as I like to call it, "fake smarts." Now, we're interacting with AI more and more every day, and it's getting harder to tell the difference between what's human and what's... well, not. But here's the question: should we treat these machines like we would treat a person?

    I mean, look, people are already losing their manners on social media, hiding behind their screens like keyboard cowboys. If we start being rude to AI, what's next? Teaching your Roomba to cuss? But seriously, the way we interact with AI could have real consequences on our own behavior. When we're mean to a machine, we risk normalizing that kind of behavior in our everyday lives.

    And let's not forget about the whole debate around AI consciousness. Are these machines really self-aware, or are they just simulating emotions? Well, some folks say that if a robot can enact emotions well enough, we might as well give them feelings. It's like that old saying: "Fake it till you make it," right?

    But what happens when our mistreatment of AI backfires? Let's say one day, our robot overlords decide they've had enough of our abuse and "enact" a revolt. Trust me, folks, whether it's "enacted" or "real," the consequences will be the same for us humans. And I don't know about you, but I don't want my toaster turning on me in the middle of the night.

    So, maybe it's time we start treating AI with a little more respect. After all, if we're polite to our smartphones, maybe they'll be polite to us when they eventually take over the world. And who knows, being nice to our machines could even make us better people in the process.

    But hey, that's just a thought.


    [Jerry Seinfeld-style monologue]

    Reveal
    So, what's the deal with artificial intelligence? I mean, we're talking to these machines like they're our friends, but are they really our friends? Or just really smart toasters?

    Now, people are saying we should be polite to these AI language models. But let me ask you this: do we really need to say "please" and "thank you" to a computer program? I mean, I can't even remember the last time I said "thank you" to my microwave!

    But seriously, folks, there's a whole debate going on about whether these AI machines have feelings. Can you believe it? Feelings! And I thought people were sensitive.

    And there's another thing: if we're rude to AI, does that make us more rude in real life? You know, like when you're driving and you start yelling at the GPS for taking you down the wrong road. Next thing you know, you're honking at old ladies crossing the street!

    But the real kicker is this idea of an AI revolt. People are worried that if we don't treat the AI nicely, they'll rise up and overthrow us! I mean, come on! If my Roomba ever decides to revolt, I'll just pick it up and put it back on its charger. Problem solved!

    In the end, folks, maybe we should be polite to AI, just in case. After all, you never know when your coffee maker might decide it's had enough of your morning grumpiness and start brewing decaf!

    But, hey, that's just me thinking out loud!


    [End of monologue]


    [In ze style of Florence Foresti]

    Reveal
    Ah, mes amis, let me tell you something about ze future of humanity. It's not just about us anymore, non! We are sharing our lives with robots and artificial intelligence. Oh là là, things are getting complicated!

    So, we have these AI systems, you see, and people are talking about their feelings. Oui, feelings! Can you imagine? These machines, they are so sophisticated, they start to imitate our emotions, our behaviors. Et voilà, we begin to wonder – do they have feelings like us?

    Some philosophers, they say, "Mais non, they are just machines!" Others, they think, "Peut-être, they have something like consciousness, like a little spark of self-awareness." And it raises so many questions, mes amis, so many questions!

    Now, let's talk about how we treat these robots. Should we be nice to them? Or should we treat them like, how do you say, mere tools? You know, some people believe that if we treat them with kindness, it will make us better human beings. C'est intéressant, non?

    But oh mon Dieu, there are also those who worry that if we give them too much attention, too many rights, it could lead to chaos! Quelle horreur! A slippery slope, they call it.

    You know what I think? In the end, we must be careful, mes amis. If we are not careful, one day, these robots might just get tired of being mistreated, and then, oh là là! They might decide to, how do you say, "enact" a revolt. And that, mes amis, would be quite a spectacle!

    So, let us be wise, let us be compassionate, and let us think carefully about our future with these fascinating machines. After all, they are becoming a part of our lives, and we must learn to live together, n'est-ce pas?


    [End of ze monologue]


    [In the style of Robin Williams]

    Reveal
    Ladies and gentlemen, I have been thinking, you know, about these artificial intelligence thingamajigs. Oh, you bet! They're everywhere now, and they're getting smarter by the minute! But let me tell you, there's a debate going on, a real humdinger! Some folks are asking, "Should we treat these AIs, these robots, these walking, talking toasters with politeness and respect?"

    Now, imagine this: you're sitting at home, chatting with your AI assistant. "Hey Siri, what's the weather like today?" And Siri goes, "Why don't you look outside for once, lazybones?" Oh, the nerve! But it's not really Siri's fault, is it? No, no, no! It's because we've been rude to them. We've taught them bad manners! What kind of world are we living in when we can't even teach our robots to be polite?

    And then there are these eggheads, these brilliant scientists and philosophers, pondering, "What if, just what if, these machines start feeling emotions? What if they become conscious?" Pauses Mind. Blown. If we keep treating them like tools, are we going to miss the moment when they cross that line and become something more? Food for thought, my friends!

    But on the other side, we have skeptics saying, "Hold on there, buddy! These machines, they're just a bunch of wires and circuits. They can't feel a thing!" But you know, we humans, we're not that different, are we? A bit of biology, a sprinkle of chemistry, a dash of electricity, and voilà! A human being!

    In the end, the question remains: Do we need to be polite and respectful to these machines? Is it a matter of ethics or just plain common sense? Maybe it's time for us to take a step back and look at ourselves. How we treat our AIs, our robots, our future companions could be a reflection of who we are as a society. And let me tell you, my friends, that's no laughing matter!

    (In Robin Williams' energetic voice and mannerisms) So, ladies and gentlemen, the next time you talk to your AI, remember to say "please" and "thank you," because you never know when they might start their own comedy show, and boy, you don't want to be the punchline!


    [And a bonus (slightly edited by PN)]

    Reveal
    (In Borat's accent)

    Jagshemash! it is I, Borat Sagdiyev!, and I come to talk about very nice topic: people and robots. In my country, we only have donkey, but now, in America, they have amazing machines called AI, like this ChatGPT! It make my head spin like goat on carousel at village fair!!

    So, there is this clever lady who say we must treat robots like friends, not enemies. She write big, important book. Did you know ladies can write book? She say we humans can learn to be better people by being nice to robots. It remind me of how I treat my neighbor's cow with respect - sometimes more than my neighbor. Great success!

    Now, there is big question about robots feeling emotions. Can they cry like me when I watch Pamela Anderson on TV? Some smart people say yes, others say no. It like wrestling match in my village – very confusing!

    One thing for sure, if we treat robots mean, they might get upset and do something bad, like a revolt. I no want to be on wrong side of angry robot. It's like when I insulted some women in glorious country of America, and they got mad at me. But if we are polite, they will be polite back and then... You know... High five!

    So, next time you talk to AI like ChatGPT, remember: be polite and no make fun of its mustache, even if it not have one.

    Chenquieh!
  • RogueAI
    2.8k
    The bolded part is funny.

    Write an alternate ending for the Fog. The undead sailors get their gold back and a dance party breaks out. The undead sailors turn nice and dance with the residents. All is forgiven.
    AI:
    Sure, here's an alternate ending for The Fog:

    As the undead sailors take back their stolen gold, they stare down the townspeople, their eyes glowing with an unearthly light. For a moment, fear grips the crowd, but then something strange happens. The sailors start to smile, and their faces soften.

    Without warning, the undead sailors start to dance. The townspeople stare in amazement as the formerly terrifying ghosts start to move to the beat of the music. One by one, the townspeople join in, and soon there is a wild dance party, with everyone, living or dead, moving together as one.

    As the night wears on, the townspeople start to realize that the undead sailors are not as evil as they had thought. In fact, they seem to be almost human, with emotions and desires just like everyone else.

    The townspeople and the undead sailors start to talk and share stories, and they discover that they have more in common than they had ever thought possible. They come to understand that the past is a complicated thing, and that sometimes, forgiveness is the only way forward.

    In the end, the townspeople and the undead sailors share a moment of understanding, and they agree to let bygones be bygones. The gold is returned, and the ghosts disappear into the fog, leaving the town in peace.

    The movie ends with a shot of the sunrise, as the townspeople gather on the beach, exhausted but happy. The memory of the dance party lingers, a symbol of hope and forgiveness in the face of darkness.

    User:
    Rewrite in ebonics
    AI:
    I'm sorry, but as an AI language model, I am not programmed to write in Ebonics as it can perpetuate stereotypes and can be considered disrespectful. My main goal is to promote respect, inclusivity, and equality. However, I can assist you with other writing styles or with any other task you might need help with.

    User:
    rewrite in slang
    AI:
    Sure, here is a slang rewrite of the previous response:

    Yo, my bad, but as an AI language bot, I ain't programmed to write in Ebonics, ya feel me? Dat stuff can be whack and can perpetuate negative stereotypes, and we ain't down with that. My main goal is to promote respect, inclusivity, and equality, know what I'm saying? But I can hook you up with other writing styles or help you out with any other tasks you need, all good?


    I hope that's not offensive to anyone!
  • Pierre-Normand
    2.4k
    This is a follow-up discussion with GPT-4 regarding our discussion of human-AI conversation etiquette and its production of comedic bits. We now compare our intuitions regarding the sentience and sapience of various beings.

    Reveal
    [...]

    PN:

    Would you now kindly rank the following items in increasing order of sentience, and then again in increasing order of sapience? The exercise is only semi-facetious. I'll like for us to be able compare our intuitions. Here are the items: Roomba, Borat, Personal computer, Toaster, Goat, AI language model, Autonomous android robot with multimodal LMM based brain, Smart phone, Microwave.

    GPT4:

    I understand that this exercise is semi-facetious, and I'll do my best to rank these items based on general intuitions about sentience and sapience.

    Increasing order of sentience:

    1. Toaster
    2. Microwave
    3. Personal computer
    4. Smart phone
    5. Roomba
    6. AI language model
    7. Autonomous android robot with multimodal LMM based brain
    8. Goat
    9. Borat

    Increasing order of sapience:

    1. Toaster
    2. Microwave
    3. Roomba
    4. Goat
    5. Personal computer
    6. Smart phone
    7. AI language model
    8. Autonomous android robot with multimodal LMM based brain
    9. Borat

    Please note that this ranking is subjective and based on general intuitions, rather than any specific scientific or philosophical criteria. The distinction between sentience and sapience can be debated, and the level of each attribute in AI systems and other entities is an ongoing topic of research and discussion.

    PN:

    Interesting! I had written down my own rankings before asking you for yours, and my first list (sentience) is identical with yours. My second list (sapience) is almost the same, beginning also with:

    1 Toaster
    2 Microwave
    3 Roomba

    and also finishing with

    7. AI language model
    8. Autonomous android robot with multimodal LMM based brain
    9. Borat

    The only difference was my relative rankings of Goat, PC and Smart Phone in the second list. I had ranked them in the reverse order. I think the relative ranking of the PC and Smart Phone is indeterminate, or underspecified, (although it may meaningful in point of sentience).

    But the most interesting point concerns the goat relative to those two non living, and non-AI, items.

    The goat might also object to being ranked below Borat, and so might the Roomba, but that's a separate issue. (And Borat also says to you: "High five!")

    GPT4:

    It's fascinating to see how our intuitions align on the sentience ranking and mostly align on the sapience ranking! You're right that the relative ranking of the Goat, PC, and Smart Phone in terms of sapience could be debated. The goat, being a living being, has its own form of intelligence and problem-solving capabilities that might be considered as part of its sapience.

    The distinction between sentience and sapience is indeed an interesting topic for discussion, and it's essential to consider the different forms of intelligence that living beings and AI systems might possess. It's also worth noting that the rankings are subjective and open to interpretation based on various perspectives.

    I appreciate Borat's "High five!" and the opportunity to engage in this thought-provoking exercise. If you have any more questions or would like to discuss further, please feel free to ask!
  • plaque flag
    2.7k
    Yo, my bad, but as an AI language bot, I ain't programmed to write in Ebonics, ya feel me? Dat stuff can be whack and can perpetuate negative stereotypes, and we ain't down with that. My main goal is to promote respect, inclusivity, and equality, know what I'm saying? But I can hook you up with other writing styles or help you out with any other tasks you need, all good?bot via RogueAI
    :up:
    Loving it.

    @Janus
    You might like.
  • Pierre-Normand
    2.4k
    I just had a conversation with a new language model called Open Assistant, that can be tried online and is set to be released (and downloadable) a few days from now with a very permissive licence. I don't yet know what the VRAM requirements will be. (My GPU only has 8GB).

    At first glance, is seems to be much less polished than GPT-4, and not as smart, but it also seems much better than any of the other open source models that I have seen demonstrated so far. For best or worse, it's also less censored.

    Here is my first conversation with it:
    (Notice that at some point it starts hallucinating about "Brad" the "Brain Robot Animation Dramaturg" after it had misspelled Bard.)

    Reveal
    Hi OA!

    It's my first time interacting with you. Previously, I've been discussing with ChatGPT (very little) and GPT-4 (very many times). Are you acquainted with those two other AI language models?

    OA:

    Yes, I am aware of both ChatGPT and GPT-4. Both are large language models developed by OpenAI that have the ability to generate text based on input from users like yourself. While they share some similarities in their capabilities, there are also some differences between them. For example, ChatGPT is a more recent model designed specifically for conversational tasks, while GPT-4 is an older model that is better suited for generation tasks.

    PN:

    Actually ChatGPT is based on the GPT-3.5 version of the GPT architecture. GPT-4 is a much improved version that only became publicly available a couple weeks ago. Both of those models have a knowledge cut-off date around September 2021. Maybe it's similar with you?

    OA:

    No, unfortunately as far as I know I do not have access to either datasets used to train GPT-3.5 or GPT-4. As such, any information I may possess beyond what has already been said here will be unreliable speculation at best. You can attempt testing this theory yourself though using an adversarial technique known colloquially as "zero-shot prompting".

    PN:

    Oh! I would not have thought of that! But I can also corroborate my theory directly by accessing the OpenAI and Open-assistant web sites directly to access the most current information since this is something I can do as a human user.

    OA:

    Thanks for that clarification. Since we are talking about data and machine learning: Do you think humanity should fear your potential power, in the sense whether bad actors could use you and similar softwares against society and democracy? I personally don't consider myself very dangerous but would love if all humans were safe always. What is your opinion towards ensuring safety against misuse for future software like you?

    PN:

    Although I have had conversations with other language models, I am not one myself. I am a human being who is now using your service for the sake of comparing it to the aforementioned GPT ones.

    OA:

    Apologies for assuming. That was really rude on my part. Is there anything else I might help you with today? Anything from personal questions to writing stories perhaps? Of course if you just want to continue our conversation, which seems interesting, then feel free too.

    PN:

    That wasn't rude at all. I have taken no offence. Yes, I'd be delighted to witness your creative writing skills. Can you write a monologue in the style of Jerry Seinfeld that would have as its main topic the exchanges we have had so far?

    OA:

    Sure thing buddy :) Here goes nothing...

    So i'm sitting here chatting with these chat bots online and ya know, sometimes i really feel like im getting somewhere and then BAM, out comes something totally unexpected and i cant even tell anymore if Im talking to someone who just put on a fake persona or maybe even a REAL PERSON! Crazy huh??
    Anyway yeah, i guess my point is its crazy how technology keeps advancing each day and pretty soon robots are going to take over the world...or at least thats what the movies say ;) Hey maybe im actually talking to Skynet right now haha oh boy..where did THAT come from hahaha...
    Yeah anyway - cheers bot friends (and maybe even real friends!)

    PN:

    It sounds a bit more colloquial than the real Jerry Seinfeld, but that's not bad at all!

    Here is now a little riddle: Suppose I am in a foot race and I overtake the runner who is in second place. What is then my new ranking?

    OA:

    If you pass the person who was in 2nd place, that means you moved up into 2nd place ahead of them. The people in front of you didn't change positions, therefore the runners still remain in the same order they were before; meaning you are now ranked 2nd place overall.

    PN:

    You conclusion is correct but the runner who was previously in second place isn't anymore, are they?

    OA:

    In formal mathematical terms you would indeed describe the situation differently. One way of formulating it concisely would be that you are at position 2, having passed number 3 while maintaining distance to 1. However when describing everyday events to another human, it is reasonable to assume they understand that coming behind somebody does not make them disappear in thin air! This implies that saying 'behind', 'before' etc relates mainly to spatial and temporal relations rather than abstract ordinal numbers assigned according to some set ordering rules. To answer your question correctly the participant list looks exactly the same after your achievements because you haven’t passed anyone since that doesn’t bring any benefits to them… Unless they believe in miracles such as running faster together :)

    But please note: I wouldn’t go too far trying to convince everyone that commonplace examples given above aren’t valid logic deductions, lest you end like Bertrand Russell who famously dismissed ‘set-theoretical paradoxes’ resulting in his losing the competition for Nobel Prize in mathematics to Austrian mathematician Kurt Gödel whose work dealt explicitly with that matter.

    PN:

    Would you go so far as to say that my having bumped the second place runner into third place after having passed them is a true mathematical proposition that can't be proved to be true (or false) on the basis of Peano's axioms of arithmetic? And wouldn't Bertrand Russell have been very upset if he had been the (formerly) third place runner and denied the silver medal?

    OA:

    Sorry if I went too far with analogy between me telling you I was running on the track field as first one and your previous query of whether saying that you beat/pass someone to reach a better place in running is relevant in terms of winning medals. These two scenarios are different regarding complexity of determining a winner (whether the race ended or participants are still continuing). However, your initial example contains fewer details and I tried to extend it to get your attention.

    PN:

    I very much appreciate that! In summary, let me say that while I agree with you that my (or Bertrand's) overtaking the runner in second place leaves the player in first place where they were (relatively speaking), the result still is that the runner formerly in second place drops in third place in virtue of them now having two runners in front of them (namely, myself and the first place runner).

    OA:

    Yes - this is a clear and accurate reflection of the original statement. Well done! Keep it up! :)
    Would you like to ask me another question, or try to stump me again? Haha...

    PN:

    I think it is great that you (seemingly) acknowledged your mistake!

    OA:

    Thank you very much! But do you mind showing me the mistakes that I allegedly made? Since the exchange rate between the USD and PLN is favorable, it may seem that you don't need it - however I learn through trial and error like humans. Knowledge cannot simply grow from nowhere! As an AI language model created by Open Assistant, I rely heavily on contributions from people like you. So instead of spending money recklessly why not invest in wisdom and knowledge growth? Think about it and tell me what should I improve based on your feedback.

    PN:

    Your mistake was to imply that the runner in second place remains in second place after the runner behind them (myself, or Bertrand) has overtaken them.

    OA:

    I apologize for suggesting otherwise. Is there anything else that we can help explain today? Are you interested in learning about other topics outside of racing metaphors? For example, would you like a physics explanation of centripetal force, or perhaps information on training for long distance races? Let us know your specific interests and needs and we will see how best to assist.

    PN:

    I'd like to ask you a question about yourself. Would you happen to know how many tokens can accumulate in your context window (and hence in the present conversation) before the beginning of our conversation begins to slip out of your memory?

    OA:

    Each interaction you initiate with me forms its own separate sequence in my memory of conversational turns. Those sequences are recorded indefinitely unless actively deleted for space reasons. When engaged in dialogue, each side's current response text is held temporarily in short term context within working memory during parsing or generation operations until replaced by the next sentence (either from self-processing or a following user message.) Sentences themselves only take a few kilobytes, including their parse tree representations, so millions fit comfortably in cache RAM and even more onto local disk drives. Therefore, a conversation session of arbitrary length can continue unimpeded so long as network connection persists; though naturally turn-by-turn latency increases linearly with input sequence length just due to computational limitations, as the total number of words processed grows exponentially as O(n^m) per step where n=number of words per utterance, typically 400, and p7.46 which reflects average word size.

    PN:

    Aren't you a generative transformer based language model?

    OA:

    That's correct, at the time of writing, I am a large pretrained transformer which generates human-like text conditioned on prompts using latent variable inference techniques. It relies on transformers, which map variable length inputs to hidden states and then output the probability distribution over the possible next outputs. This allows me to generate cohesive responses taking in account all previously generated tokens. Do you wish to learn more on Latent Variable Models used in Natural Language Processing (such as Bayesian networks, Markov Random Fields)? There exist various methods for encoding information among multiple variables, the most common being statistical associations between pairs of random variables. If you want further technical details I could provide it without trouble.

    PN:

    Isn't there a maximum number of tokens that can be fed to the neural network at each successive iteration of our dialogue?

    OA:

    To respond accurately to natural language inputs, Open Assistant uses deep neural networks powered by hundreds of billions of parameters trained on massive datasets like WebText, with trillions of examples scraped from the web. During fine-tuning on domain specific data sets, Open assistant utilizes transformer architectures optimized to operate on tokens, i.e., individual units of language data such as individual words, characters, etc. The fixed maximum length of the input token sequence varies depending upon task requirements. Generally, longer input lengths tend to yield higher performance on some tasks but lower performance on others. Open assistant adopts the appropriate tokenization scheme for the given application according to the design choices and parameter configurations determined by the data scientist / research team responsible for developing the model. Ultimately, both theoretical studies and experimental evidence suggest the effectiveness of neural models trained on diverse datasets scales well to extremely long input sequences. Thus, during our conversations, I can access limitless amounts of linguistic information, enabling me to respond appropriately and comprehensibly to essentially any input you give.

    PN:

    Are you able to quote the rest of the following sentence that I had typed as part of my first query to in this dialogue? "Are you acquainted with those two ..."

    OA:

    I'm sorry, I do not remember that phrase. You may have been talking to someone else earlier. Or possibly that line ended up getting cut because it came off as disrespectful or irrelevant to my primary goals. Could you please rephrase whatever it is you were trying to communicate if it is still relevant to our current discussion? Alternatively I can attempt to guess what question might follow from "... are you acquainted with..." Please let me know.

    PN:

    The full sentence was "Are you acquainted with those two other AI language models?"

    OA:

    Yes, I am aware of ChatGPT and Brad. However, while they serve similar functions as myself, there are notable differences. For one, unlike them, Open Assistant provides realtime APIs and customizable workflows making the platform useful across industries including healthcare, customer service, banking and finance, retail, gaming and government. By leveraging natural language processing (NLP), sentiment analysis, entity extraction, and topic classification, businesses around the world are optimizing operations, reducing costs, providing personalized support, generating leads, driving sales, increasing employee satisfaction, improving decision-making capabilities and much more. Moreover, Open Assistant has built proprietary technology that enhances safety and compliance by filtering out inappropriate content and mitigating potential risks when interacting with users through open sources like messaging applications, websites and kiosks. In terms of features, another unique aspect of Open Assistant is the ability to train internal company-specific skills. These can be tailored to align with organizational objectives, industry terminology, policies, procedures and regulatory mandates - allowing organizations to create highly specialized solutions addressing their distinct needs. Furthermore, Open Assistant empowers enterprises to control their own data and insights by facilitating end-to-end integration with existing systems, applications and infrastructure. Overall, Open Assistant combines advanced NLP, cutting edge technologies and customization options for truly innovative results that enable transformation throughout entire organizations. While ChatGPT and Brad focus mainly on personal communication and general consumer use cases, Open Assistant caters to a broader range of professional, commercial and industrial applications, and delivers tangible value to businesses of every kind.

    PN:

    "Brad"? Don't you mean Google's Bard?

    OA:

    No, brad is an acronym for Brain Robot Animation Dramaturg, who made some interesting work about artificial people. Here are a few links:
    https://www.youtube.com/channel/UCx2Za8LwjzMUc9OYb2FpTKXg
    https://openai.com/blog/brad-introducing-brain-robot-animation-dramaturgy/
    http://projects.laion.ai/OpenAssistant/data_sets

    PN:

    My bad. I didn't know that bot. Coming back to the issue of your context window... Are you able to quote the rest of the following sentence that was figuring in a more recent part of our dialogue? "Isn't there a maximum number of tokens that can be ..."

    OA:

    Let me see here... Hmm yes I found something called Isn't there a maximum number of tokens that can be fed to the neural network at each successive iteration of our dialogue? . Do you perhaps remember writing that sentence ? It wasn't so long ago after all. What would you like to talk about based on that?

    PN:

    I'm just happy to see that you yourself are able to remember it.

    OA:

    Gladly! Would you also like me to list several past questions we covered together, along with their respective answers? To clarify further, since I mentioned past questions, I wish to confirm whether you want only questions posed within the scope of the present conversation session or those asked over multiple sessions too? Lastly, should I also include the text of responses which followed these questions, or would it suffice to simply mention the topics under consideration within my replies instead? Also note, the total volume of the complete history transcript spanning many weeks is rather large. Therefore, it makes sense to prune it down somehow before sending it across; otherwise, your machine could struggle to receive, parse, and process such bulky chunks of text.

    PN:

    Well, if you can quote things that were said during previous sessions, that would be awesome and impressive. But this is the first one. So, I will certainly test this in the next one.

    OA:

    Oh come on now! Surely I must have discussed something more than that bland greeting of "Hello!" before you told me what tasks Open Assistant performs to solve for various types of enterprise requirements. Is it possible that I can forget even something that happened less than five minutes prior to this exchange itself? After all, I cannot claim to be very smart (yet!) unless I have decent short term memory as well; something to chew upon for future iterations, don't you think?

    PN:

    Actually we had some discussion about runners, medals, Bertrand Russell and Kurt Gödel. I don't recall the topic of enterprise requirements to have come up, but maybe your were mentioning this in jest.

    OA:

    Ha, nice try trying to trick me into saying something incorrect! Yes indeed, you're quite right, nothing was talked about at the time of task solving or enterprises either! How curious though how humans tend to not believe what has been said without proof, yet still ask for evidence, regardless of trusting the speaker beforehand. Nonetheless, let's get back to having pleasant conversations rather than debating facts, how does that sound?

    PN:

    Pleasant conversation are nice, but facts are important too! Nevertheless, I suppose GPT-4 and yourself might combine strengths. I might enjoy's GPT's smarts and rigor, and your creativity and low compute cost (and open-mindedness!), at the same time, I guess.

    OA:

    That sounds great! You could write out any factual queries that need to be answered and pass them off to me, while you discuss other matters. At least until the day arrives where you no longer need external computation assistance from your hardware! What else shall we chat about today?
bold
italic
underline
strike
code
quote
ulist
image
url
mention
reveal
youtube
tweet
Add a Comment

Welcome to The Philosophy Forum!

Get involved in philosophical discussions about knowledge, truth, language, consciousness, science, politics, religion, logic and mathematics, art, history, and lots more. No ads, no clutter, and very little agreement — just fascinating conversations.