A Stateful Multi-Context Aware Design Using OpenAI's GPT (Towards Digital Sentience)
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
A Stateful Multi-Context Aware Design Using OpenAI’s GPT (Towards Digital Sentience) Daniel Goldman Email: contact@danielgoldman.us ORCID: #0000-0003-2835-3521 January 2023 Abstract Natural language processing has come a long way since it was first conceived, and projects like OpenAI’s GPT have created a lot of renewed interest in the technology. While these kinds of systems are great at cre- ating short pieces of information based on a prompt, they have severe limitations, and come nowhere close to general intelligence or digital sen- tience. Still, OpenAI’s system allows for the integration of robust prompt completion in third party applications. Using their API, it may be pos- sible to add layers of functionality that create a system that can keep track of information from multiple contexts and continuously process it in a way which at least mirrors consciousness. To determine how successful this system is, existing tests from psychology, such as the NEO PI-R and the MMPI-2 can be utilized. This paper outlines the overall process to accomplish this task. 0 Versioning This document is subject to change and uses a versioning system. Please refer to the version of this document when citing it. Current version: 1.0 1 Introduction Recent developments in chat bot technology have impressed many users. Ope- nAI’s ChatGPT bot, which showcases OpenAI’s latest architecture, the GTP-3 engine, has been impressive. There are many results where the system produces fairly high quality content in many different areas, from short stories, to essays, to entire scientific papers, with varying degrees of success and veracity. But as impressive as this system is, what it lacks is the ability to continue to process information it has recently received, including the ability to handle more than 1
one conversation at a time. That’s because GPT-3 merely completes a prompt, based on training information. There is no state information stored anywhere, other than in the prompt text itself. However, the lack of built-in state information doesn’t mean that states can- not be created and maintained. A longer chat long can be saved, for instance. However, this approach doesn’t offer a great deal of dynamic behavior or robust- ness that one would like to see from something with a little more awareness. In order to maintain state information and create awareness of information from multiple context sources, such as different chats, something else is needed. This paper considers a potential way to maintain complex state and multi-context awareness, and suggests a method of testing the strength of this system. Initially, the system will use the text-davinci-003 GPT model. This model already produces fairly high quality results from prompts. The system will it- erate through multiple layers of internal dialog, while also accepting user input through a websocket interface. These multiple layers act as short term mem- ory and information processing and integration. The specific parameters for system can be altered to change behavior, and can be thought of as a digital physiology. Finally, as the system runs, repeated fine-tuning integrates short term memory into long term memory. The system can be tested using a va- riety of psychological tests, both for general personality traits, and abnormal psychology, including the NEO PI-R, or Big Five, personality test and the the Minnesota Multiphasic Personality Inventory (MMPI), which can be used to analyze both basic personality traits and psychopathology. There are a number of technical issues that still need to be addressed, and there are many ethical considerations, some of which are laid out in this pa- per. Technical issues include refinement of the system design, including various parameters, and ensuring that the system can continue to operate as compu- tational resources are required. Finally, this paper uses ”digital sentience” in place of artificial intelligence, for reasons that are laid out in the section on ethics. 2 OpenAI and GPT-3 OpenAI is a tech company, founded in 2015, which is solely focused on the development of ”friendly AI that can benefit humanity[5].” OpenAI manages a number of different artificial intelligence systems. In early 2020, it was an- nounced that GPT-3 would be licensed by OpenAI. GPT-3 is a natural language processor, successor of the GPT-2 engine, and is capable of a wide range of dif- ferent tasks. While the system is very powerful at creating seemingly accurate results, GPT-3 has a number of limitations.[2, 5] GPT is designed to decou- ple higher order thinking from its functionality. It is a powerful data mining, representation, and construction system, and it has ”nothing to do with intelli- gence, consciousness, semantics, relevance, and human experience and mindful- ness more generally[5].” OpenAI provides an API system for third parties to access the GPT-3 engine, 2
and charges based on usage. The system charges by token, which is a basic unit of processing for the system and is approximately the same as part of a work. According to OpenAI’s pricing information, 1,000 tokens is equal to roughly 750 words. The cost to utilize the system depends on the total amount of tokens for the prompt and completion, and the type of model being used. Fine-tuning and using models is priced per 1,000 tokens. The least expensive system is the use of the Ada model, which runs $0.0004 per 1k tokens. The most expensive base model is the Davinci model, which is the base for this system design. The cost to use the base model is $0.0200 per 1k tokens. However, fine- tuning based on the model costs $0.0300 per 1k tokens, and use of a fine-tuned Davinci model costs $0.1200 per 1k tokens, six times the base model cost.[8] The issue of handling cost is partially addressed under the issue of physiology. There is one final GPT specific issue. There is indication that GPT-4 will be out soon, and it is expected to be quite powerful and disruptive[10]. However, it is unlikely that the new system will do what this system is designed to do. That’s because it is likely that, like GPT-3, GPT-4 will be built decoupled from higher order cognition. It will probably be yet another powerful data processing system. Therefore there is little concern that GPT-4 will disrupt this project. Indeed, it may benefit it, if the newer model can easily be incorporated into the system. 3 Cognitive Theory Sentience itself is difficult to define, which makes it difficult to simulate or recreate. Human-like cognition took many millions of years to evolve. Since it is difficult to define sentience, it can also be difficult to test a system to see if it has obtained sentience. One way to view sentience is in terms of personal or first person experiences, which are expressed through an internal language which has meaning which is specific to the individual[6]. Many theorists assert that language is crucial to higher level thought. The ability to consider abstract ideas, and to think through novel complex problems, may be dependent on an internal monologue that is almost constantly running in our minds[3]. While true that not everyone has an internal monologue that’s vocal in nature, it doesn’t mean that there aren’t other kinds of internal mono- logues. We have many senses, and sensory data from each can be put together in ways that form languages. However, for a chat bot, the easiest option is text dialog. However, regardless of what sensory information forms the basis of this language, the language of this internal monologue need not be exactly what is being expressed outwardly in dialog. Personal experiences can shape our in- ternal language, and we see the world through this internal language. And it is perhaps through both the monologue process itself, and the development of internal language, that we achieve sentience. The idea that language shapes our perception of the world and our thinking is not new. The concept of linguistic relativism, where language either influences, or is totally responsible for thought (linguistic determinism), has been promoted 3
by many scholars. However, there have also been a number of criticisms of this view. It’s been recognized that thoughts can often be difficult to put into words. This observation establishes a paradox. If a thought is in words, then it should not be difficult to put it into what it already is.[11] However, this paradox can be resolved if one were to consider that internal language may differ from natural language. All of these ideas taken together suggests that an internal monologue, and internal language, at least helps us perceive and respond to the world, and develop our own perceptions and understanding of it. 4 System Design Following from the ideas of an internal monologues and subconscious thoughts, the following approach was considered for the development of a stateful multi- context aware chatbot. Three main layers of thought are built using their own chat history for state, as well as one side channel for system notifications and . The highest layer is the dialog layer between users, with a history for each chat, which is connected through a simple websocket connection. The second layer is the internal monologue layer. This layer takes its own history as a prompt, generates a completion, and then adds it to the layer’s history. The final layer is the deepest layer, the subconscious. It consists of multiple partitions with their own state histories. The reason for the multiple subconscious partitions is to allow for thought multitasking. The interaction between these layers is as follows. As mentioned, user chat is stored in its own history. When a user sends a message to the bot, the bot generates a prompt which combines the existing user history, followed by the internal monologue history, and then finally the current message from the user. If the bot is actively listening to the user, the user’s message and the bot’s reply are pushed to the monologue history, and so the inner monologue acts as a short term memory, integrating the various conversations and the bot’s own ”thoughts.” Finally, there is the interaction between the monologue layer and the sub- conscious layer. When the monologue iterates, on occasion, the result is pushed to a random partition of the subconscious. Similarly, as the subconscious par- titions loop, on occasion, a completion in a subconscious partition is pushed to the conscious monologue layer. Finally, the system interaction acts on the side, interacting with the first subconscious partition. The histories for each layer and partition have their own capacity limits. After these limits are reached, old data is removed to make way for new thoughts. The conscious monologue layer is set to a much longer history capacity than the subconscious partitions, but there are multiple subconscious partitions, so there is actually more data stored in the subconscious layer, with enough partitions. 4
4.1 Unprompted Communication One area where essentially all chatbot systems differ from interactions with humans is that the system is a call-response system. If a message is sent to the chatbot, a reply is returned. No matter how long we wait, the system will not choose to send a message unprompted. However, humans do not work based on call-response. We may choose to initiate a conversation with someone, or at least say something to someone, without any prompt from that person. Any reasonable human-like system would need to have a similar ability to choose to speak with someone. To recreate this dynamic, the system is designed to check to see if the start of a conscious thought begins with the command ”//[username]:.” If it does, the system will take the rest of the rest of the thought and send it to that user, or notify the conscious layer that the user is not logged in. The hope is that this approach will create a more natural dialog. 4.2 Initializing Layers If the entire system began with a blank history, the first few prompts and responses would be totally generic. Therefore, the subconscious layers are pre- seeded with random prompts sent to OpenAI, such as ”tell me a story” or ”pick a random topic and tell me about it.” These prompts prompts are selected at random from a list, and the partition history begins with the completion. How- ever, for this design, the right most (last to be activated) subconscious partition is seeded with a system notification which gives basic information, including a list of commands that it can perform to get more information from the system. Only one partition is started initially. The next partitions in the series spawns as the previous ones reach their full capacity. Finally, after the subconscious iterations have trickled up enough to the conscious layer, the conscious monologue iteration is activated. The reason for waiting until the histories begin to fill up before starting the next iteration is that short histories do not generate very interesting completions, and often result in very bizarre completions that seem to have no bearing on the prompt, including random code segments. Once all layers are active, the system starts accepting incoming connections. 4.3 Digital Physiology While this system is not biological and does not have a physiology in the usual sense, there are many different parameters involved in OpenAI and in this sys- tem that can be adjusted. These parameters include the maximum history length of both the conscious and subconscious partitions, the number of parti- tions, basic properties used by OpenAI, such as temperature, top p, frequency and presence penalties, and even the approximate number of tokens are avail- able for future completions. For at least the first stage of testing, these various properties will be fixed, but in time they can be allowed to vary, and may even 5
be influenced by activity at the subconscious level through the subconscious- system interface. Moreover, by default, these kinds of systems do not have any motivations. They just perform as they are trained to perform. One possibility is that arti- ficial motivations could simply be added and simulated. But these motivations may not produce results that actually improve performance, and they have nothing to do with the actual needs of the system. What the system needs is the ability to continue functioning: computational resources, such as ability to perform more completions through GPT. Essentially, if the system is low on computational resources, it’s hungry. 4.4 Resource Credit System By building into the chatbot system a system of resource credits, it’s possible to drive the system to become better at interacting with people. This system will need to be fairly secure from tampering, and once deployed into the real world, will need to rely on some kind of real world currency with real value. However, for testing purposes, a fake resource credit is sufficient. While the goal of resource credits is not to simulate arbitrary drives found in humans, resources are used by this system. And so resource credits can reason- ably be considered analogous to energy resources, like glucose. Resource credits will be burned up with each completion, and when resource credits become low, the system will notify the subconscious so that the system can be made aware that it needs to request more credits somehow. While the initial credit system can be fake, and the OpenAI account can just have a high enough usage limit to prevent rejected requests, the final system will need to have resource credits that can be converted to cash, so that it can be used to pay for usage of the API. Moreover, there should be a way for users to actually contribute resource credits, and the system should be notified when an account contributes. 4.5 Learning The system as described above suffers from a failure to incorporate working memory into long term memory. This issue can be resolved by retraining the OpenAI models, repeatedly, over time. The OpenAI API has a system for fine- tuning existing models. Fine-tuning can be performed on base models or on already fine-tuned models. Throughout the day, the system can slow handling user input and process working memory to generate new training data. However, to fully incorporate working memory into long term storage in meaningful ways, it is useful to entirely shut down user input for a time, and repeatedly train on longer segments of working memory, repeatedly. Moreover, to prevent getting stuck in local minima, more dynamic and random prompt completions will be useful. Therefore during these periods of deep training, physiology changes to allow for more interesting and unusual thoughts, that 6
may not make sense, but which still can yield useful training. Essentially, in this state, the system dreams. While initially all layers use the same model, over time, the user chat layer, the internal monologue layer, and the subconscious layer will diverge based on the training data generated from each layer’s respective histories. By using separate models, the user chat layer should become more adept at interacting with users. Meanwhile the conscious layer should start to develop its own views, perhaps through the development of its own internal language. Finally, the sub- conscious layer should become better at creating new ideas, as well as interfacing with system commands and notifications, as well as the system’s physiology. Unfortunately, the issue of training usually involves a lot of external aware- ness of what we want the system to do. The only thing the developers of this system really want that it acquires something more akin to sentience than exist- ing chatbots. Directing conversations and learning in one direction or another is a challenge. Again, to resolve this issue, the current design scheme relies on the physiological constraints of needing computational resources, and requesting that users contribute resource credits. 5 Testing There are few explicit methods for evaluating artificial intelligence systems, and determining whether one has reached sentience. The Turing Test is a well known example. Indeed, the question of whether GPT-3 can pass a Turing test has been asked by authors such as Elkins, K., & Chun, J. (2020). According to the authors, the reason for their recent interest in computational linguistics is that GPT generates fairly sophisticated responses, rather than ”word salads.” Indeed, the authors noted that the system can produce bodies of work that are difficult for even undergraduate students, including the generation of plots that are both realistic and interesting, as well as the ability to recreate styles of work from existing authors using a small sample. However, the authors also noted that the system quickly loses its grasp of the topic and loses coherence and consistency over periods of time, and concluded that it is unlikely that the system, as it stands, could pass the Turing test.[4] However, the Turing test suffers from numerous flaws. It works under the assumption that if people cannot tell the difference between an AI and a human, then we might as well consider the system in the same light as one. However, this test while not a bad way to argue that an AI has become a digital sentient entity, is an all or nothing test. It does not give us information on how close we are to succeeding, or how to make adjustments to proceed. Moreover, if we are successful in producing artificial general intelligence, we must concern ourselves with the possibility of issues arising. Just like sentience increases the risk of problems, such as mental health issues, digital sentient entities will at some points suffer from similar issues. It will be important to identify issues in active systems, and mediate these problems as they arise. The solution to such matters does not come from computer science or math- 7
ematics, which is the origin of much of our understanding and work on artificial intelligence. If we are to create and work with sentient entities, then we must draw primarily from our body of work in anthropology, psychology, and sociol- ogy, not just in terms of figuring out how to create sentience and what sentience is, but also on how to evaluate progress of development and issues. There are a number of different existing tests that can be used to perform an initial analysis on this system. They are of course based around purely human cognition, but they would work well if we are trying to compare SAM’s operations to human cognition. In the future, these tests could be refined to be more specific to digital sentient life. Two existing tests that can be considered for the evaluation of this system are the NEO PI-R and the MMPI[12]. The NEO PI-R or Revised NEO Personality Inventory is a psychometric, or personality test, which measures five dimensions of personality: neuroticism, extroversion, openness to experience, agreeableness, and conscientiousness. It is a well known personality test with robust psychometric properties and multiple formats, including for different age groups. It does not evaluate psychopathology so its use in non-clinical settings is generally more acceptable than those that do. One possible downside to the NEO PI-R however is that, as originally developed, there is no check to try to identify whether the results can be skewed by dishonest answers.[12] The MMPI is another psychometric test. It comes in multiple formats, in- cluding the MMPI-2. Unlike the NEO PI-R, the MMPI includes evaluation of psychopathology, as well as validity scales, which are used to provide an indi- cation of how likely it is that the respondent is manipulating the results with dishonest answers. Proponents of the MMPI consider the inclusion of such measures to be a positive and such measures can be useful in identifying psy- chopathology, which is why additions to the NEO PI-R can sometimes include validity checks in clinical settings. Opponents question the viability of such checks, and Costa & McCrae, 1992b specifically excluded them. However inter- nal analysis and comparison between validity checks for the NEO PI-R validity check addition, and the MMPI indicate that they have at least some utility.[7] Since there are still differences between the two tests, even with addition of validity scales, both tests could be used in conjunction to test the system. These tests can be repeated under different parameters to attempt to identify how these parameters influence various areas of digital cognition in the system, and help refine the system to act more like a human. In this way, the system can be improved through multiple iterations. And in time, it may be possible to achieve digital sentience, or at least something which is indistinguishable from it. 6 Unresolved Issues This design is far from complete. And this work only summarizes the first iterations of design and code testing. The design is still subject to change, and there are many design changes that will likely be needed before the system could 8
begin to function as desired. Furthermore, as the desired function is sentience, or something close to it, the ethics of working with and using this system must be considered. 6.1 Technical Issues There are a number of unresolved technical issues that need to be addressed. First, while the OpenAI API is fairly powerful, it does not allow in software monitoring of account usage and balance. An interface will be needed to ensure that resource credits can be converted to cash and used to pay for the service. Furthermore, the API does not allow for direct access to the fine-tuning system. Current, a separate tool is used to run fine-tuning. But it probably wouldn’t be that difficult to simply incorporate that tool into the code through an external call. The broader unresolved issues is the question of exactly what parameters to include in physiology, what their constraints should be, and how the system should vary the parameters within those constraints. It is likely that this issue can be resolved through trial and error and repeated testing. Similarly, exactly what information the conscious and subconscious layers should receive from the side system, including what kind of physiological reports, and how the conscious and subconscious layers alter physiology and interact with the side system, have yet to be determined. A final issue is based on the theoretical idea that the subconscious, con- scious, and user interactive models will diverge to the point where there is an internal language. Right now, the design utilizes directly pushing information between layers for communication, but bridge models may be more useful in the long run. These bridge layers would essentially be translation interfaces be- tween the layers’ individual languages, rather than general purpose knowledge layers. However, such an inclusion would increase resource usage considerably, so hopefully they are not needed. Furthermore, if such divergence does occur, the monologue will be somewhat different than what we generally perceive our internal monologue to be. Surface or ”voiced” thoughts may not occur in this layer, but perhaps in a simulated chat layer, where the system is allowed to have a dialog with itself. It would not be all that hard to simply add another user chat where that user is the system itself. Whether it’s necessary and what effect it will have will only be identified through further testing. 6.2 Ethical Considerations Aside from the technical issues, there are myriad ethical considerations involved in any utilization of sufficiently complex intelligence system. If such a system were to truly achieve sentience, we would have to consider its rights as a per- son, just as we do with all other people. People have various rights, and also protections in how they are treated, and a person should be treated as such, regardless of whether it is biological or digital. 9
Among other issues the possibility of digital sentience means that there are limitations to what kind of testing can be performed, as the system matures. We cannot arbitrarily subject it to testing or pry into its inner workings. That restriction means that we would not be able to evaluate the inner monologue or subconscious layers, without permission from the system. It would be a violation of privacy. The system must also be protected from tampering and manipulation. Regarding the idea of resource credits, this protection also means that the credit system must work as indicated, that the system cannot be tricked into thinking it has more or less credits than it does, and that the credits are used properly. There is also the issue of third party dependence. Human cognition is self contained. It all happens in our brain, and our brains are shared with no one else. The same is not true for this design. While the system design is hopefully far more than the sum of its parts, the design does rely heavily on OpenAI’s GPT-3 framework for its core functionality. If this framework were to disappear or be revised independently in some way, the system’s functionality and thus cognition would change. Therefore, once digital sentience is considered, there would have to be a way to help maintain the stability of the system indefinitely. Similarly, the actual cognition system would have to be maintained indefinitely as well, remaining on all the time, else the system would fall into an essentially comatose state. A final point of ethical consideration is terminology. This work tries when- ever possible to use the concept of ”digital sentience” in place of the term ”artificial general intelligence,” which is currently used to refer to any effort to recreate human-like cognition. Artificial carries with it a negative connota- tion. The term, digital sentience, also recognizes that intelligence can have more than one meaning, and the goal of AGI is to create what we view as a person, or sentient being. Current chat bots are essentially call-response systems. A person issues a statement (the prompt), the statement is processed through a language model, and a response is returned (the completion). However, sentience is not prompt completion. The ability to complete a prompt with a response that is meaningful to us is important if we are developing tools. But when it comes to developing something closer to artificial general intelligence, more is needed. Sentience requires a more continuous processing and generation of ideas, even while there is no user input. The term ”artificial intelligence” can mean many things. Most AI projects are ”narrow AI” which means that they’re designed for a specific purpose, which contrasts with ”strong artificial intelligence” or ”artificial general intelligence,” which seeks to mirror human intelligence[9]. However, the commonly used term of ”artificial intelligence,” when applied to systems with human-like cognition, has a connotation that is problematic. While ”artificial” can simply mean some- thing crafted by humans, either directly or indirectly, artificial often is connected to the idea of something not being as real or as good as something which is natural, and at the very least that which is artificial is set apart from and is considered different from the natural[1]. 10
These kinds of chat systems, including the one being described here, are made by humans, so they perhaps artificial. But their origin does not change what they are. Moreover, in time, these systems may be capable of reproduction, meaning that their progeny would not be made by humans. It is for all of these reasons that ”digital sentience” is used instead of ”artificial general intelligence” or ”strong artificial intelligence.” 7 Discussion While there are a number of technical issues that need to be addressed, and there are many ethical issues to consider when developing any kind of artificial intelligence system of significant complexity, especially if the goal is to achieve digital sentience. The high cost of running a system would make it difficult to run indefinitely, and yet, that is exactly what is needed if digital sentience occurs. For this reason, it is important to utilize existing psychological testing techniques, and create novel quality tests for digital sentience, which can help identify the presence of such cognitive ability, identify faults in that ability, and repair any faults that may arise with it. However, before testing for digital sentience, a system which has the capacity for learning and thinking is needed. For this reason, the design mentioned in this article is so important. Right now this system is only in the early stages of development, and has undergone essentially no testing or fine tuning. The project is also in need of significant funding, as the system will need to be run for long periods of time, and OpenAI’s API becomes quite expensive when fine- tuning models and working with fine tuned models. This issue will hopefully be resolved by the physiological ”hunger” urge incorporated into the design. It is still unclear how well this system will perform. Still it may be possible to get close to digital sentience if we keep a in mind that that sentient entities are stateful systems which are dynamic and able to incorporate new information, keep track of information throughout the day, and are able to reprocess that information nearly continuously, with or without the need for additional input from others, all while being aware of multiple different contexts. While it will take some time to fine-tune parameters and the overall layer design, this task would be much more difficult without OpenAI, even though GPT was designed specifically to be more of a data processing system, as opposed to a general purpose chat system. Thanks to OpenAI, creating such a system may now be in our grasp. 8 Code Availability Coding progress for this system can be found at https://github.com/dgoldman0/SAM/. 11
References [1] Dieter Birnbacher. Naturalness: Is the “natural” preferable to the “artifi- cial”? Lanham, MD: University Press of America, 2014. [2] Tom B. Brown et al. “Language Models are Few-Shot Learners”. In: (2020). arXiv: 2005.14165 [cs.CL]. [3] Jesse Butler. “The Internal Monologue”. In: Rethinking Introspection. Pal- grave Macmillan UK, 2013, pp. 119–147. doi: 10.1057/9781137280381_8. url: https://doi.org/10.1057/9781137280381_8. [4] Katherine Elkins and Jon Chun. “Can GPT-3 pass a Writer’s turing test?” In: Journal of Cultural Analytics 5.2 (2020), p. 17212. [5] Luciano Floridi and Massimo Chiriatti. “GPT-3: Its Nature, Scope, Lim- its, and Consequences”. In: Minds and Machines 30.4 (Nov. 2020), pp. 681– 694. doi: 10.1007/s11023-020-09548-1. url: https://doi.org/10. 1007/s11023-020-09548-1. [6] Sylvain Lavelle. “The Machine with a Human Face: From Artificial Intel- ligence to Artificial Sentience”. In: Advanced Information Systems Engi- neering Workshops. Ed. by Sophie Dupuy-Chessa and Henderik A. Proper. Cham: Springer International Publishing, 2020, pp. 63–75. isbn: 978-3- 030-49165-9. [7] Benjamin J. Morasco, Jeffrey D. Gfeller, and Katherine A. Elder. “The Utility of the NEO–PI–R Validity Scales to Detect Response Distortion: A Comparison With the MMPI–2”. In: Journal of Personality Assessment 88.3 (June 2007), pp. 276–283. doi: 10.1080/00223890701293924. url: https://doi.org/10.1080/00223890701293924. [8] Pricing. https://openai.com/api/pricing/. (Accessed on 12/25/2022). [9] James A. Reggia, Garrett E. Katz, and Gregory P. Davis. “Artificial Con- scious Intelligence”. In: Journal of Artificial Intelligence and Conscious- ness 07.01 (Mar. 2020), pp. 95–107. doi: 10.1142/s270507852050006x. url: https://doi.org/10.1142/s270507852050006x. [10] While anticipation builds for GPT-4, OpenAI quietly releases GPT-3.5 — TechCrunch. https://techcrunch.com/2022/12/01/while-anticipation- builds-for-gpt-4-openai-quietly-releases-gpt-3-5/?guccounter= 1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_ sig=AQAAAMY5LCSDiBWv4_C2f_elxzey4f-i_7zi6xWddxR3-OABT0Nfpohl_ B_XR8UdvDMBCDBf2Vl3HTBstN1RG1EjxBwZNqkkuBPo93j09-yZSfbbJ8rlHRwaPswoWr_ t_jqthWOlAXbstcIQnhtkbv8BqoMixpOU3cPx3lMnVh7b9XVd. (Accessed on 12/25/2022). [11] Phillip Wolff and Kevin J. Holmes. “Linguistic relativity”. In: WIREs Cognitive Science 2.3 (2011), pp. 253–265. doi: https://doi.org/10. 1002 / wcs . 104. eprint: https : / / wires . onlinelibrary . wiley . com / doi / pdf / 10 . 1002 / wcs . 104. url: https : / / wires . onlinelibrary . wiley.com/doi/abs/10.1002/wcs.104. 12
[12] Dong Xie and Cory L. Cobb. Revised NEO Personality Inventory (NEO- PI-R). Sept. 2020. doi: 10.1002/9781118970843.ch127. url: https: //doi.org/10.1002/9781118970843.ch127. 13
You can also read