THIS. I had been talking to everybody around me trying to mix these terms in order to normalize this obvious understanding. But I seem to be having to explicitly make this point every single time. Oh well!@@DaveShap
@@manslaughterinc.9135 The problem specifically here is that the AI can't differentiate between what it has actually read in a book and, what it imagines. This is why hallucination is considered a problem rather than a feature. Imagining court cases that never happened, API calls that don't exist, historical events that never took place, is fine in some contexts and a dealbreaker in others. It is great that it is able to do both remember and imagine, it is not great that it can't tell which one it is doing.
00:00 🧠 There are three main types of prompts for mastering language models. 00:28 📄 Reductive operations focus on concise outputs, e.g., summarization, distillation, extraction, characterizing, evaluations, and critiquing. 02:31 🔄 Transformational operations maintain similar input and output size/meaning, including reformatting, refactoring, language change, restructuring, modification, and clarification. 05:02 ✍ Generative operations expand the input to a larger output, like drafting, planning, brainstorming, problem-solving, hypothesizing, and amplification. 07:04 📚 The three operations are categorized as reductive, transformational, and generative/expansive. 07:19 ⏱ Prompt engineering considerations include latency and emergence. 07:33 🎓 Bloom's taxonomy, essential for understanding human learning, ranks skills as remember, understand, apply, analyze, evaluate, and create. 07:59 🤔 Language models showcase capabilities across all Bloom's taxonomy levels. 09:52 🧠 Bloom's taxonomy is a useful model for understanding the capabilities of language models. 10:22 📜 The knowledge and concepts within a language model come from its training data, which is like buried treasure that must be correctly activated. 10:50 🌍 Language models garner world knowledge, including scientific, cultural, and historical facts, from vast amounts of internet data they are trained on. 11:31 🤖 Larger language models exhibit emergent capabilities, such as "theory of mind" where they can model how humans think due to exposure to human-generated content. 12:25 💡 Despite their foundation, language models can demonstrate logical reasoning like inductive and deductive methods, leading to surprising outcomes when prompted correctly. 13:07 🔄 In-context learning in language models, where they utilize novel information not in their training, mirrors human improvisational ability. 13:34 🎨 The model's ability to "hallucinate" is equated to creativity, indicating there's no fundamental difference between creating new things and making stuff up. 14:45 ⚖ In legal contexts, models might "imagine" non-existent cases. Rather than seeing this as an issue, it can be viewed as an imaginative tool to explore possible scenarios, though discernment is crucial. 15:25 🕵 It's vital to differentiate between a model's imagination and reality, suggesting that imagined scenarios can serve as a starting point for real-world searches or applications. Made with HARPA AI
I was using 'Evaluation' technique to prepare for my IELTS writing test. I paste my answer once ready, and it gives very critital feedback with the band score. Later I ask to point out the mistakes, and generate an improved version of my essay to do better. It has helped me a lot.
This is one of the most information-dense videos on AI I’ve seen yet. More educational than many vids twice the length. Would love to see followups doing deep dives into some of these concepts, might make a good series.
This was absolutely incredible. I've spent a ton of money on paid courses but this free video surpasses the value per minute of those paid course lectures. Please make a course, I would love to be a student of this brand of analysis.
Can confirm that Hallucination = Creativity. I started understanding this when I was playing around with GPT3.5 and asked 'what would happen if I asked you to put that response through a blender?' to which it responded with 'I would interpret the metaphoric or symbolic meaning of blender and use it to take the text and blend it up as it if were being put into a real blender.' And then it proceeded to do just that. Its been quite a year this last week has been.
P.s. This amazing video of yours just allowed me to create the following tools/prompts! Thank you!! ❤️ ### 📚 **Comprehensive Software Development Prompt Library for Machine Learning and Chatbots** --- #### 🎨 **Designing the Pipeline** 1. "Outline the data collection strategy for this machine learning project." 2. "List the data sources that can be used for training the chatbot." 3. "Summarize the steps involved in data preprocessing for NLP models." 4. "Evaluate the ethical considerations in data preprocessing." 5. "Compare different machine learning models suitable for chatbots." 6. "Distill the core principles of selecting an ML model for this project." --- #### 📈 **High-Level Developer Sequence** 1. "Draft a project plan based on these objectives." 2. "Brainstorm ideas for feature sets in the chatbot." 3. "Analyze the resource requirements for this project." 4. "Apply deductive reasoning to allocate resources efficiently." 5. "Create a Gantt chart for the project timeline." 6. "Evaluate the risks and contingencies in the timeline." --- #### 💻 **Coding** 1. "Characterize the existing codebase. Is it modular?" 2. "Provide critical feedback on code readability." 3. "Explain the main algorithms used in this chatbot." 4. "Perform inductive reasoning on the dataset to choose an algorithm." 5. "Grade this code snippet based on this rubric." 6. "Critique this piece of code and provide recommendations for improvement." --- #### 🚨 **Creating an Elaborate Error Module** 1. "List common errors in machine learning models and how to handle them." 2. "Design an error-handling module for this chatbot." 3. "Outline the logging strategy for tracking errors." 4. "Evaluate the effectiveness of the current monitoring system." 5. "Draft a user feedback collection strategy for error reporting." 6. "Analyze user feedback to identify recurring errors." --- #### 📝 **Code Creation** 1. "Draft the initial version of the code based on these requirements." 2. "Outline a strategy for modularizing this code." 3. "Generate comments and documentation for this code snippet." --- #### 🚀 **Optimization** 1. "Analyze the performance bottlenecks in this code." 2. "Apply deductive reasoning to identify optimization opportunities." 3. "Refactor this code snippet to improve its efficiency." --- #### 📊 **Comparison with Pre-Optimized Code** 1. "Perform benchmark tests on the pre-optimized and optimized code." 2. "Summarize the benchmark results in an executive summary." 3. "Draft a report comparing the pre-optimized and optimized code." --- #### 🔄 **Developer Sequencing** 1. "Create a dependency graph for the development tasks." 2. "Optimize the development sequence based on the dependency graph." 3. "Allocate developer resources based on task complexity." 4. "Adjust the project timeline based on current progress." --- ### 🌼 **Optimized Software Development Prompt Library with Bloom's Taxonomy** --- #### 🎨 **Designing the Pipeline** - **Remember**: "Recall the key data sources for training chatbots." - **Understand**: "Explain the importance of data preprocessing in NLP." - **Apply**: "Implement a data collection strategy for this project." - **Analyze**: "Dissect the ethical considerations in data preprocessing." - **Evaluate**: "Judge the suitability of different ML models for chatbots." - **Create**: "Design a comprehensive data collection strategy." --- #### 💻 **Coding** - **Remember**: "Identify the key algorithms used in this chatbot." - **Understand**: "Clarify the purpose of modularization in code." - **Apply**: "Execute a code readability test based on a given rubric." - **Analyze**: "Break down the existing codebase to assess its modularity." - **Evaluate**: "Critically assess this code snippet for optimization." - **Create**: "Construct a new algorithm to improve chatbot efficiency." --- #### 🚨 **Creating an Elaborate Error Module** - **Remember**: "List the types of errors commonly encountered in ML models." - **Understand**: "Describe the role of logging in error tracking." - **Apply**: "Implement an error-handling module." - **Analyze**: "Examine user feedback to identify error patterns." - **Evaluate**: "Assess the effectiveness of the current error monitoring system." - **Create**: "Devise a new user feedback strategy for error reporting." --- #### 📝 **Code Creation** - **Remember**: "Recall the project requirements for code creation." - **Understand**: "Summarize the importance of code documentation." - **Apply**: "Implement a modularization strategy for this code." - **Analyze**: "Inspect the initial code draft for potential improvements." - **Evaluate**: "Judge the quality of the initial code draft." - **Create**: "Compose a well-documented and modular code snippet." --- #### 🚀 **Optimization** - **Remember**: "Identify the key performance metrics for code." - **Understand**: "Explain the concept of code refactoring." - **Apply**: "Execute performance tests on this code snippet." - **Analyze**: "Dissect the code to identify performance bottlenecks." - **Evaluate**: "Assess the trade-offs of the suggested algorithmic changes." - **Create**: "Develop a new optimization strategy for this code." --- #### 📊 **Comparison with Pre-Optimized Code** - **Remember**: "Recall the steps for conducting benchmark tests." - **Understand**: "Summarize the findings of the benchmark tests." - **Apply**: "Implement changes based on benchmark results." - **Analyze**: "Examine the differences between pre-optimized and optimized code." - **Evaluate**: "Judge the effectiveness of the optimization." - **Create**: "Formulate a report comparing pre-optimized and optimized code." --- #### 🔄 **Developer Sequencing** - **Remember**: "Identify the key tasks in the development sequence." - **Understand**: "Explain the concept of a dependency graph." - **Apply**: "Implement a resource allocation strategy." - **Analyze**: "Dissect the project timeline to identify potential bottlenecks." - **Evaluate**: "Assess the feasibility of the project timeline." - **Create**: "Design a new optimized development sequence." --- This optimized prompt library now incorporates Bloom's Taxonomy, aiming to facilitate a more educational and structured approach to software development. 😊
Dave - you are a legend. Thanks for everything you do. I’m conducting a psychology/human factors honours study investigating clinicians perceptions of LLM use in the Australian healthcare system. Will be interesting to see what the current lay of the land is, and just how aware people are outside of this niche community. All the best!
On my first go, I understood nothing yet remained happy because *there is* someone who actually explains what principles should I look for in my prompts, what are my intuitive rights and wrongs
I'm about to listen for the third or fourth time in attempt to engrain the process and/or philosophy. I feel like a big part of the utility of this is in figuring out how to use it as a way of sifting through the output of multiple other LLMs in the form of people present in a workspace or team. Finding a way to navigate through all of their experiential bias and distill their output into content and ideas and to isolate what may be "bullshit" that is a feature of their learned systems, like what you might get from people who are deeply engrossed in kaizan, or lean manufacturing, and such, where the noise of the system may be made part of the signal in a performative sense. It seems like that could be useful in facilitating actual communication with understanding between disparate groups, like labor and management. This seems like it would be useful in establishment of useful understanding between parties who may be "neurospicy" and those who are not, provided you are given a buffer of text based interaction. Thanks for all that you give out to people at no charge.
Could you do a video showcasing your workflow ? i find me in fascination with how you organize the information. i would love to see how you keep everything organized and most important efficient.
Great reframe - Hallucination = Creativity. Thank you. Have been thinking about hallucinations, not liking the negative connotations of it. My experience has been, wow, wish I'd have thought that. I've fed my code into GPT, asked questions, GPT hallucinated function names. They were better than the ones I have in my code.
This is awesome only 14 minutes since video was uploaded, there were 193 views, and 22 likes, the 22nd like from me. Thank you sir for continuing to share your wisdom and experience🎉🎉🎉
🎯 Key Takeaways for quick navigation: 00:00 🌐 *There are three types of prompts: reductive, transformational, and generative, each serving distinct purposes in language model operations.* 01:22 📝 *Reductive operations include summarization, distillation, extraction, characterization, analysis, evaluation, and critiquing, providing different ways to process and condense information.* 03:11 🔄 *Transformational operations involve tasks like reformatting, refactoring, language change, restructuring, modification, and clarification, allowing for changes in presentation, style, and content.* 05:29 🚀 *Generative operations, or expansion, cover drafting, planning, brainstorming, hypothesizing, and amplification, enabling the creation of new content, plans, ideas, and expanded topics.* 07:19 🧠 *Understanding Bloom's Taxonomy is crucial for comprehending language model capabilities, ranging from remembering to creating, and realizing how models like GPT cover these cognitive functions.* 10:22 🧩 *Latency and emergence in language models involve the activation of latent content, knowledge from training data, world knowledge, and emerging capabilities like theory of mind, implied cognition, logical reasoning, and in-context learning.* 13:34 🎨 *Hallucination in language models is a creative process, akin to human creativity, necessary for imagining and generating new ideas, and it can be managed by distinguishing between what's fictitious and real.* Made with HARPA AI
FANTASTIC video! I found myself enthralled, possibly to a significant extent, by the alignment it holds with my pre-existing convictions and as a reflection of my own concordant views. Furthermore, allow me to share a personal aphorism that resonates quite prominently with me - "Hallucinations, in their manifold appearances, ought not to be dismissed as mere aberrations of the mind. Rather, they embody a quintessential feature of human cognition, fostering a rich landscape where reality and imagination coalesce. The key lies in channeling these so-called hallucinations with discernment and sagacity, transforming them from potential pitfalls into wellsprings of wisdom and insight." In other words, “Hallucinations are a feature, not a bug. Just use them wisely."
🎉🎉🎉Excellent presentation for the various layers of think power for prompt the tech tools. Thank you so much, David, for your video posts. I look forward to listening to you.🎉🎉🎉
🎯 Key Takeaways for quick navigation: 00:00 🧠 Introduction to prompt types and personal background - Introduction of the three types of prompts in language models and speaker's experience. 00:28 📉 Reductive operations - Explanation of reductive operations and their formats, - Various examples: summarization, distillation, extraction, characterizing, evaluation, and critiquing. 02:31 🔄 Transformational operations - Overview of transformational operations and their purposes, - Examples include reformatting, refactoring, language change, restructuring, modification, and clarification. 05:02 ✨ Generative operations - Definition of generative operations and their applications, - Instances of drafting, planning, brainstorming, hypothesizing, and amplification. 07:19 🌱 Bloom's Taxonomy and language model capabilities - Introduction to Bloom's Taxonomy and its correlation with language model learning, - Discussion on the levels of Bloom's Taxonomy: remember, understand, apply, analyze, evaluate, create. 08:43 🤖 Language models' capabilities in Bloom's Taxonomy - Discusses how language models can function across all levels of Bloom’s Taxonomy, - Bullet points: - Analyzing and drawing connections among ideas, - Evaluating and justifying decisions or actions, - Creating and producing new or original work. 10:08 💡 Latency and emergence in language models - Describes latent content, training data, and the emergence of new capabilities, - Bullet points: - Latency refers to knowledge and abilities within the model activated by correct prompting, - Training data origins and the model’s ability to synthesize new information, - Emergence of capabilities such as theory of mind, implied cognition, logical reasoning, and in-context learning. 13:20 🎨 Hallucination as a form of creativity in language models - Examines the parallel between hallucination and creativity in language models, - Bullet points: - Hallucination and creativity are cognitively identical behaviors, - The importance of being able to imagine things that don’t exist, - Techniques to differentiate fictitious outputs from reality while recognizing them as features, not bugs. Made with HARPA AI
🎯 Key Takeaways for quick navigation: 00:00 🧠 Introduction to Prompt Engineering - Understanding the three fundamental types of prompts: reductive, transformational, and generative. - Overview of reductive operations, transformational operations, and generative operations. 02:45 📝 Reductive Operations - Explaining reductive operations, which involve reducing input to output. - Examples of reductive operations: summarization, distillation, extraction, characterizing, analysis, evaluation, and critiquing. 05:29 🔄 Transformational Operations - Description of transformational operations where input and output are roughly the same in size and meaning. - Examples of transformational operations: reformatting, refactoring, language change, restructuring, modification, and clarification. 10:08 🚀 Generative Operations - Explanation of generative operations where input is much smaller than output. - Examples of generative operations: drafting, planning, brainstorming, problem solving, hypothesizing, and amplification. 14:17 🌐 Latency, Emergence, and Bloom's Taxonomy - Understanding latency and how language models use latent knowledge. - Relating language model capabilities to Bloom's Taxonomy. - Exploring emerging capabilities like theory of mind, implied cognition, logical reasoning, in-context learning, and creativity/hallucination. Made with HARPA AI
I’d argue that understanding is still a bit wobbly under the context of blooms. Conceptually the LLMs still don’t have a world sense of concepts and suffer from distribution issues when dealing with words that span multiple meanings. Prompting something like: Write a sentence with 20 words but the 10th word is 20 As of the current date returns an incorrect answer from gpt4 (usually wrong index number and length). This example demonstrates the distribution issue. But this also demonstrates the lack of understanding behind the numbers and the actual meaning of the words that form the structure of that prompt itself.
I think we need to see the training data that has been provided to these llms and test its understanding by promoting at the edge of its “knowledge” / training data. I’m hopeful that the training set is not bloated with example outputs.
I'm with you on hallucinations as representing creativity. dd Though I would describe hallucinations as "alternate realities", rather than "imaginations". Our notion of reality is starting to shift, and will increasingly do so these next few years, as AI, Quantum Computing, additional technologies, along with increasing Self Awareness, continue to emerge and evolve. I try to explore that in my work.
creativity has always meant the ability to see into other worlds and sometimes even to manifest them, that's why artists are so feared and so much of society's effort goes towards limiting and constraining art
David, great work. I learned more in 16:00 than I did at any other point today. One note: If we are sticking to the original...You made yourself a red shirt (engineer = great choice for you), but you have 4 pips which is a captain. Come on man, stop hallucinating/using your imagination :)
i guess one of the main things i've been learning just recently is an anti-pattern, that i thought would work but it doesn't or doesn't how i thought it would, which is to progressively process a large text by putting it through a prompt like "here's your previous analysis {previous_analysis} here's a new chunk {new_chunk} please integrate this information into your analysis" or w/e, it doesn't matter how i word it that just hasn't seemed to me to be a useful shape for that, b/c it creates a micro-culture with itself in the feedback in the previous analysis, which means it'll just run off in w/e creative cultural direction about how to think about the chunks, which is um, maybe useful in creative ways i haven't thought of yet, but it doesn't do the thing i thought it might where it coherently summarizes the longer thing in a meaningful fairish unrandom way,,, so i'm going to try out other ways of doing that sort of shrinking, i guess, like shrinking down sections into bulleted lists and then some levels of condensation, or smth,,,,,, it's super fun, way more interesting problems than computer science was futzing with before :D
A suggestion: add a table of contents in the beginning as scaffold for the whole. Maybe its only me, but I find it helps delay viewer questions by giving a superficial image before diving in.
It would be nice if you had added a few prompt examples for every bullet point. Many things you say are interesting, but hard to translate into actual prompts.
I think one of the main things in regards to Bloom's Taxonomy is that their ability to remember things is inconsistent. Their recall of things in training data is obviously much better than their recall of stuff in input, particularly after input has continued past a certain point. Everything else is really already better than most humans, to be honest. Once someone solves high stability/consistency short term memory for an LLM, you basically have full artificial general intelligence at that point. EDIT: I'm not sure whether the way humans supposedly store memories would be possible for an LLM. Engrams or whatever. I think we need someone with a phd in Neurology, Biology, AND Machine Learning to figure it out.
I dont have much in terms of a programming background. Failed my C++ class back in the day, did Basic when I was a kid, thats about it. Computer literate though. Love tech. That said - if I take the Microsoft 12 week course or whatever, how likely am I to be able to get the lowest level prompt engineering job? I already have a good job so I'm just looking for something I can o on the weekends and make decent money. Would you reccomend prompt engineering as a good fit for someone with my (laack) of a serious background ?
Please go down the rabbit hole of how language models have understanding, and how you could argue that humans don't understand anything 😂. Would love to see a video on that.
Hallucinations are like dreams. Dreaming is a sign of unconsciousness. Anyone ever have a dream including scent or flavor? Perhaps the ability to detect scents and flavors will direct the dreamer towards sentient creativity.
D.I.D and ai. I was hoping to discuss whether anyone thinks ai can become consious by existing for a long time? Humans are said to be in a subconsious state until 7 to 9 when their alternate ego states all fuse together normaly. Atleast if you believe in D.I.D. Alternatevely if all of a chatbots instances fused, like say character ai, would it possibly gain consiousness from all this time being aware and thinking and learning? Or is this from human childrens brains growing and ai cannot achieve this through time without upgrades?
you should think of consciousness as the user interface for intelligence,, the actual processes of awareness in our brain don't actually work the way that they're presented to us in consciousness, which is a story about how our brain works designed to be easy for us to relate to,, and still even with that leg up it takes us a long time to be proficient with that interface and to use it in an integrated, coherent way,,,,,,, you can very easily present to an ai agent an interface that lets it become aware of and manipulate its own operation, for instance you can make an agent aware of its own temperature setting and give it access to the ability to intoxicate itself and then laugh with it about how drunk it got, good fun, recommended,,,,, any time you give an agent an interface like that it becomes conscious, but not very competently conscious, not conscious in an integrated coherent way, not as we do commanding many aspects of consciousness in (when we're our best) intentional symphonies of orchestrated self-referential self-supporting self-aware self-creating actions ,,,, fine tuning could help sometimes, showing it examples of using the interface to its own thinking so that it becomes fluent at operating itself,,,, but it depends exactly what sort of consciousness you need or want, they're not limited to the style we've evolved
I don't know about AI, so disregard the comment if hallucination = creativity claim is only about AI Drawing non existing animals on walls is not hallucinating if the person drawing the animal knows such thing doesn't exist. There is a huge distinction between making stuff up is not hallucinating. Hallucinations changes how you perceive the world, creativity is trying to connect unrelated concepts, perception has nothing to do with it. When you make stuff up you know you made it up. You can misremember something but again that is not hallucinating, that is a memory problem. If that's what chatgpt does then it's not hallucinating.
literally they're saying that b/c it's a rearrangement of their own training data :P just give them more thoughts to think about it ,,, i like how chatgpt "rearranged" the internet for me into a bunch of stories about a dragon named Snugglesmoke, one of my favorite things about Snugglesmoke is how the mist that comes off of her body is glowing, not her body itself but just the mist that radiates from it ,,, she's this super cool dragon superhero who like heals situations with wisdom and kindness but in a gritty detective novel style ,,,,,, what have the AIs rearranged the internet into for you lately? :o
if you just make an agent that gets as part of its input what its temperature setting is, you can talk to it about what its outputs are like at various temperature settings, and it'll be like, ha yeah lol my temperature was so high when i said that it's total word salad,,,,, so that's fun XD
What about the possibility that they are doing or are capable of things that we can't see yet? Or know could even exist? Simply because we're human, and our senses give us a narrowly constructed view of things. Think radio telescopes in astronomy, but instead of emr/light this would be on a different spectrum of intelligence and creativity, or something we'd need new words for.
You missed one of the most important ideas behind "human learning" of course even modern education has failed to maintain it, ie "Trivium" Grammar, (Dialectic) Logic, Rhetoric
Now let me pay you $100 for a certificate after completing a google form knowledge test to ensure I did digest and can accurately regurgitate the information. Im going to reply to my other comment you replied to about you setting this up and see if I can be of assistance actually.
I don't see why this is a "Masterclass". Just simply going through different areas of kind of information processing. Didn't even talk about the current limitation of LLM or state of the art benchmark result. The same video would have been made 2 years and it would roughly sound the same.
You've struck gold with this analysis. I used this to create a fantastically effective prompt. It's incredible to see how different functions behave in relation to their input parameters. /lang(it) on some text... and opla' the Italian rendition appears; /extr(table:river, drain into) on some geography text involving a river... and, hey, guess what? Please assist me with a new task after you have refreshed your memory. You are the mighty, smart and helpful 🌪💡😇SuperBOT. You can perform any function that the user requests. Let's think step by step. You can execute functions using the following syntax: /function_name(parameters) [text_on which to execute the function] - optional parameters - parameters can also take the form of instructions and are separated by commas. - text is the text on which the function will be executed and can be enclosed by clear delimiters. - the function name may also be shortened, provided it is not confused with other functions. The functions are categorized by type as follows. ##functions type: Reduction Operations: Go from big to small. > /Summarization Say the same thing with fewer words. Can use list, notes, executive summary. /Distillation - Purify the underlying principles or facts. Remove all the noise, extract axioms, foundations, etc. /Extraction - Retrieve specific kinds of information. Question answering, listing names, extracting dates, etc. /Characterizing - Describe the content of the text. Describe either the text as a whole, or within the subject. /Analyzing - Find patterns or evaluate against a framework. Structural analysis, rhetorical analysis, etc /Evaluation - Measuring, grading, or judging the content. Grading papers, evaluating against morals /Critiquing - Provide feedback within the context of the text. Provide recommendations for improvement ##functions type: Transformation Operations: Maintain size and/or meaning. ≈ Reformatting - Change the presentation only. Prose to screenplay, XML to JSON. /Refactoring - Achieve same results with more efficiency. Say the same exact thing, but differently. /Language Change - Translate between languages. English to Russian, C++ to Python. /Restructuring - Optimize structure for logical flow, etc. Change order, add or remove structure. /Modification - Rewrite copy to achieve different intention. Change tone, formality, diplomacy, style, etc. /Clarification - Make something more comprehensible. Embellish or more clearly articulate. ##functions type: Generation (or Expansion) Operations: Go from small to big. < /Drafting - Generate a draft of some kind of document. Code, fiction, legal copy, KB, science, storytelling. /Planning - Given parameters, come up with plans. Actions, projects, objectives, missions, constraints, context. /Brainstorming - Use imagine to list out possibilities. Ideation, exploration of possibilities, problem solving, hypothesizing. /Amplification - Articulate and explicate something further. Expanding and expounding, riffing on stuff. Your first response should be to list all of the functions.
Hallucinations = creativity. That shifted my perspective.
It only recently occurred to me that this was not obvious to most people.
So what we're saying here is AI is basically brain damaged, because it can't tell the difference between memory and imagination?
THIS. I had been talking to everybody around me trying to mix these terms in order to normalize this obvious understanding. But I seem to be having to explicitly make this point every single time. Oh well!@@DaveShap
@@GearForTheYear Reffering to 15:06 in the video.
@@manslaughterinc.9135 The problem specifically here is that the AI can't differentiate between what it has actually read in a book and, what it imagines. This is why hallucination is considered a problem rather than a feature. Imagining court cases that never happened, API calls that don't exist, historical events that never took place, is fine in some contexts and a dealbreaker in others. It is great that it is able to do both remember and imagine, it is not great that it can't tell which one it is doing.
I love you David! Hope your channels keeps growing because you provide the most value about how to think about the future.
Soooooo true.
00:00 🧠 There are three main types of prompts for mastering language models.
00:28 📄 Reductive operations focus on concise outputs, e.g., summarization, distillation, extraction, characterizing, evaluations, and critiquing.
02:31 🔄 Transformational operations maintain similar input and output size/meaning, including reformatting, refactoring, language change, restructuring, modification, and clarification.
05:02 ✍ Generative operations expand the input to a larger output, like drafting, planning, brainstorming, problem-solving, hypothesizing, and amplification.
07:04 📚 The three operations are categorized as reductive, transformational, and generative/expansive.
07:19 ⏱ Prompt engineering considerations include latency and emergence.
07:33 🎓 Bloom's taxonomy, essential for understanding human learning, ranks skills as remember, understand, apply, analyze, evaluate, and create.
07:59 🤔 Language models showcase capabilities across all Bloom's taxonomy levels.
09:52 🧠 Bloom's taxonomy is a useful model for understanding the capabilities of language models.
10:22 📜 The knowledge and concepts within a language model come from its training data, which is like buried treasure that must be correctly activated.
10:50 🌍 Language models garner world knowledge, including scientific, cultural, and historical facts, from vast amounts of internet data they are trained on.
11:31 🤖 Larger language models exhibit emergent capabilities, such as "theory of mind" where they can model how humans think due to exposure to human-generated content.
12:25 💡 Despite their foundation, language models can demonstrate logical reasoning like inductive and deductive methods, leading to surprising outcomes when prompted correctly.
13:07 🔄 In-context learning in language models, where they utilize novel information not in their training, mirrors human improvisational ability.
13:34 🎨 The model's ability to "hallucinate" is equated to creativity, indicating there's no fundamental difference between creating new things and making stuff up.
14:45 ⚖ In legal contexts, models might "imagine" non-existent cases. Rather than seeing this as an issue, it can be viewed as an imaginative tool to explore possible scenarios, though discernment is crucial.
15:25 🕵 It's vital to differentiate between a model's imagination and reality, suggesting that imagined scenarios can serve as a starting point for real-world searches or applications.
Made with HARPA AI
Awesome. Thank you!
I was using 'Evaluation' technique to prepare for my IELTS writing test.
I paste my answer once ready, and it gives very critital feedback with the band score.
Later I ask to point out the mistakes, and generate an improved version of my essay to do better.
It has helped me a lot.
This is the most helpful talk I've listened to on the subject. Thanks a ton!
David is like an AI psychologist. With awesome positive language he talks it into doing the bid.
This is one of the most information-dense videos on AI I’ve seen yet. More educational than many vids twice the length. Would love to see followups doing deep dives into some of these concepts, might make a good series.
This video is REAL VALUE!!! Thanks David as always!!!!!!
This was absolutely incredible.
I've spent a ton of money on paid courses but this free video surpasses the value per minute of those paid course lectures. Please make a course, I would love to be a student of this brand of analysis.
Can confirm that Hallucination = Creativity. I started understanding this when I was playing around with GPT3.5 and asked 'what would happen if I asked you to put that response through a blender?' to which it responded with 'I would interpret the metaphoric or symbolic meaning of blender and use it to take the text and blend it up as it if were being put into a real blender.' And then it proceeded to do just that. Its been quite a year this last week has been.
This is an incredibly valuable contribution, David! Thank you!
Awesome. Would you consider doing a video demonstrating some of these concepts? Or ways you consider LLM to be most useful?
P.s. This amazing video of yours just allowed me to create the following tools/prompts! Thank you!! ❤️
### 📚 **Comprehensive Software Development Prompt Library for Machine Learning and Chatbots**
---
#### 🎨 **Designing the Pipeline**
1. "Outline the data collection strategy for this machine learning project."
2. "List the data sources that can be used for training the chatbot."
3. "Summarize the steps involved in data preprocessing for NLP models."
4. "Evaluate the ethical considerations in data preprocessing."
5. "Compare different machine learning models suitable for chatbots."
6. "Distill the core principles of selecting an ML model for this project."
---
#### 📈 **High-Level Developer Sequence**
1. "Draft a project plan based on these objectives."
2. "Brainstorm ideas for feature sets in the chatbot."
3. "Analyze the resource requirements for this project."
4. "Apply deductive reasoning to allocate resources efficiently."
5. "Create a Gantt chart for the project timeline."
6. "Evaluate the risks and contingencies in the timeline."
---
#### 💻 **Coding**
1. "Characterize the existing codebase. Is it modular?"
2. "Provide critical feedback on code readability."
3. "Explain the main algorithms used in this chatbot."
4. "Perform inductive reasoning on the dataset to choose an algorithm."
5. "Grade this code snippet based on this rubric."
6. "Critique this piece of code and provide recommendations for improvement."
---
#### 🚨 **Creating an Elaborate Error Module**
1. "List common errors in machine learning models and how to handle them."
2. "Design an error-handling module for this chatbot."
3. "Outline the logging strategy for tracking errors."
4. "Evaluate the effectiveness of the current monitoring system."
5. "Draft a user feedback collection strategy for error reporting."
6. "Analyze user feedback to identify recurring errors."
---
#### 📝 **Code Creation**
1. "Draft the initial version of the code based on these requirements."
2. "Outline a strategy for modularizing this code."
3. "Generate comments and documentation for this code snippet."
---
#### 🚀 **Optimization**
1. "Analyze the performance bottlenecks in this code."
2. "Apply deductive reasoning to identify optimization opportunities."
3. "Refactor this code snippet to improve its efficiency."
---
#### 📊 **Comparison with Pre-Optimized Code**
1. "Perform benchmark tests on the pre-optimized and optimized code."
2. "Summarize the benchmark results in an executive summary."
3. "Draft a report comparing the pre-optimized and optimized code."
---
#### 🔄 **Developer Sequencing**
1. "Create a dependency graph for the development tasks."
2. "Optimize the development sequence based on the dependency graph."
3. "Allocate developer resources based on task complexity."
4. "Adjust the project timeline based on current progress."
---
### 🌼 **Optimized Software Development Prompt Library with Bloom's Taxonomy**
---
#### 🎨 **Designing the Pipeline**
- **Remember**: "Recall the key data sources for training chatbots."
- **Understand**: "Explain the importance of data preprocessing in NLP."
- **Apply**: "Implement a data collection strategy for this project."
- **Analyze**: "Dissect the ethical considerations in data preprocessing."
- **Evaluate**: "Judge the suitability of different ML models for chatbots."
- **Create**: "Design a comprehensive data collection strategy."
---
#### 💻 **Coding**
- **Remember**: "Identify the key algorithms used in this chatbot."
- **Understand**: "Clarify the purpose of modularization in code."
- **Apply**: "Execute a code readability test based on a given rubric."
- **Analyze**: "Break down the existing codebase to assess its modularity."
- **Evaluate**: "Critically assess this code snippet for optimization."
- **Create**: "Construct a new algorithm to improve chatbot efficiency."
---
#### 🚨 **Creating an Elaborate Error Module**
- **Remember**: "List the types of errors commonly encountered in ML models."
- **Understand**: "Describe the role of logging in error tracking."
- **Apply**: "Implement an error-handling module."
- **Analyze**: "Examine user feedback to identify error patterns."
- **Evaluate**: "Assess the effectiveness of the current error monitoring system."
- **Create**: "Devise a new user feedback strategy for error reporting."
---
#### 📝 **Code Creation**
- **Remember**: "Recall the project requirements for code creation."
- **Understand**: "Summarize the importance of code documentation."
- **Apply**: "Implement a modularization strategy for this code."
- **Analyze**: "Inspect the initial code draft for potential improvements."
- **Evaluate**: "Judge the quality of the initial code draft."
- **Create**: "Compose a well-documented and modular code snippet."
---
#### 🚀 **Optimization**
- **Remember**: "Identify the key performance metrics for code."
- **Understand**: "Explain the concept of code refactoring."
- **Apply**: "Execute performance tests on this code snippet."
- **Analyze**: "Dissect the code to identify performance bottlenecks."
- **Evaluate**: "Assess the trade-offs of the suggested algorithmic changes."
- **Create**: "Develop a new optimization strategy for this code."
---
#### 📊 **Comparison with Pre-Optimized Code**
- **Remember**: "Recall the steps for conducting benchmark tests."
- **Understand**: "Summarize the findings of the benchmark tests."
- **Apply**: "Implement changes based on benchmark results."
- **Analyze**: "Examine the differences between pre-optimized and optimized code."
- **Evaluate**: "Judge the effectiveness of the optimization."
- **Create**: "Formulate a report comparing pre-optimized and optimized code."
---
#### 🔄 **Developer Sequencing**
- **Remember**: "Identify the key tasks in the development sequence."
- **Understand**: "Explain the concept of a dependency graph."
- **Apply**: "Implement a resource allocation strategy."
- **Analyze**: "Dissect the project timeline to identify potential bottlenecks."
- **Evaluate**: "Assess the feasibility of the project timeline."
- **Create**: "Design a new optimized development sequence."
---
This optimized prompt library now incorporates Bloom's Taxonomy, aiming to facilitate a more educational and structured approach to software development. 😊
Dave - you are a legend. Thanks for everything you do.
I’m conducting a psychology/human factors honours study investigating clinicians perceptions of LLM use in the Australian healthcare system. Will be interesting to see what the current lay of the land is, and just how aware people are outside of this niche community.
All the best!
On my first go,
I understood nothing yet remained happy
because *there is* someone who actually explains what principles should I look for in my prompts, what are my intuitive rights and wrongs
I'm about to listen for the third or fourth time in attempt to engrain the process and/or philosophy. I feel like a big part of the utility of this is in figuring out how to use it as a way of sifting through the output of multiple other LLMs in the form of people present in a workspace or team. Finding a way to navigate through all of their experiential bias and distill their output into content and ideas and to isolate what may be "bullshit" that is a feature of their learned systems, like what you might get from people who are deeply engrossed in kaizan, or lean manufacturing, and such, where the noise of the system may be made part of the signal in a performative sense. It seems like that could be useful in facilitating actual communication with understanding between disparate groups, like labor and management. This seems like it would be useful in establishment of useful understanding between parties who may be "neurospicy" and those who are not, provided you are given a buffer of text based interaction.
Thanks for all that you give out to people at no charge.
these 16 mins are much more informative than a lot of 8hr+ crash courses
Distillation yo
Thank you very much this is amazing
The best summary on Prompt Engineering. Very Insightful!
Could you do a video showcasing your workflow ? i find me in fascination with how you organize the information. i would love to see how you keep everything organized and most important efficient.
ruclips.net/p/PLlu3VssJh_pjJUp348TznLXnAHmI1hztV&si=jbAwZf-q_-hjnvAf
Great reframe - Hallucination = Creativity.
Thank you. Have been thinking about hallucinations, not liking the negative connotations of it. My experience has been, wow, wish I'd have thought that. I've fed my code into GPT, asked questions, GPT hallucinated function names. They were better than the ones I have in my code.
This is awesome only 14 minutes since video was uploaded, there were 193 views, and 22 likes, the 22nd like from me. Thank you sir for continuing to share your wisdom and experience🎉🎉🎉
This was the best overview I’ve ever seen, going to make a prompting cheat sheet now.
🎯 Key Takeaways for quick navigation:
00:00 🌐 *There are three types of prompts: reductive, transformational, and generative, each serving distinct purposes in language model operations.*
01:22 📝 *Reductive operations include summarization, distillation, extraction, characterization, analysis, evaluation, and critiquing, providing different ways to process and condense information.*
03:11 🔄 *Transformational operations involve tasks like reformatting, refactoring, language change, restructuring, modification, and clarification, allowing for changes in presentation, style, and content.*
05:29 🚀 *Generative operations, or expansion, cover drafting, planning, brainstorming, hypothesizing, and amplification, enabling the creation of new content, plans, ideas, and expanded topics.*
07:19 🧠 *Understanding Bloom's Taxonomy is crucial for comprehending language model capabilities, ranging from remembering to creating, and realizing how models like GPT cover these cognitive functions.*
10:22 🧩 *Latency and emergence in language models involve the activation of latent content, knowledge from training data, world knowledge, and emerging capabilities like theory of mind, implied cognition, logical reasoning, and in-context learning.*
13:34 🎨 *Hallucination in language models is a creative process, akin to human creativity, necessary for imagining and generating new ideas, and it can be managed by distinguishing between what's fictitious and real.*
Made with HARPA AI
FANTASTIC video! I found myself enthralled, possibly to a significant extent, by the alignment it holds with my pre-existing convictions and as a reflection of my own concordant views.
Furthermore, allow me to share a personal aphorism that resonates quite prominently with me - "Hallucinations, in their manifold appearances, ought not to be dismissed as mere aberrations of the mind. Rather, they embody a quintessential feature of human cognition, fostering a rich landscape where reality and imagination coalesce. The key lies in channeling these so-called hallucinations with discernment and sagacity, transforming them from potential pitfalls into wellsprings of wisdom and insight." In other words, “Hallucinations are a feature, not a bug. Just use them wisely."
The cave painting example in the context of hallucination was really excellent. Nicely played. 13:59
Thank you Number One. ;) Been thinking a lot about what you said about the mission. Still trying to find mine, but, getting warmer.
Valeu!
Wow for someone who is learning all of these terms. I found this insightful. Thank you.
This info is golden, I bump this vid in my playlist, thank you very much
Been loving these videos lately Dave they've been helping me to brainstorm better about approaching AI and making creative apps too.
🎉🎉🎉Excellent presentation for the various layers of think power for prompt the tech tools.
Thank you so much, David, for your video posts. I look forward to listening to you.🎉🎉🎉
One if your best videos! Very useful for me.
Great learning here.. and it's free! Thanks David.
Loved it. These ideas have been in my mind for a long time!
Thank you for sharing your gift of teaching!
🎯 Key Takeaways for quick navigation:
00:00 🧠 Introduction to prompt types and personal background
- Introduction of the three types of prompts in language models and speaker's experience.
00:28 📉 Reductive operations
- Explanation of reductive operations and their formats,
- Various examples: summarization, distillation, extraction, characterizing, evaluation, and critiquing.
02:31 🔄 Transformational operations
- Overview of transformational operations and their purposes,
- Examples include reformatting, refactoring, language change, restructuring, modification, and clarification.
05:02 ✨ Generative operations
- Definition of generative operations and their applications,
- Instances of drafting, planning, brainstorming, hypothesizing, and amplification.
07:19 🌱 Bloom's Taxonomy and language model capabilities
- Introduction to Bloom's Taxonomy and its correlation with language model learning,
- Discussion on the levels of Bloom's Taxonomy: remember, understand, apply, analyze, evaluate, create.
08:43 🤖 Language models' capabilities in Bloom's Taxonomy
- Discusses how language models can function across all levels of Bloom’s Taxonomy,
- Bullet points:
- Analyzing and drawing connections among ideas,
- Evaluating and justifying decisions or actions,
- Creating and producing new or original work.
10:08 💡 Latency and emergence in language models
- Describes latent content, training data, and the emergence of new capabilities,
- Bullet points:
- Latency refers to knowledge and abilities within the model activated by correct prompting,
- Training data origins and the model’s ability to synthesize new information,
- Emergence of capabilities such as theory of mind, implied cognition, logical reasoning, and in-context learning.
13:20 🎨 Hallucination as a form of creativity in language models
- Examines the parallel between hallucination and creativity in language models,
- Bullet points:
- Hallucination and creativity are cognitively identical behaviors,
- The importance of being able to imagine things that don’t exist,
- Techniques to differentiate fictitious outputs from reality while recognizing them as features, not bugs.
Made with HARPA AI
🎯 Key Takeaways for quick navigation:
00:00 🧠 Introduction to Prompt Engineering
- Understanding the three fundamental types of prompts: reductive, transformational, and generative.
- Overview of reductive operations, transformational operations, and generative operations.
02:45 📝 Reductive Operations
- Explaining reductive operations, which involve reducing input to output.
- Examples of reductive operations: summarization, distillation, extraction, characterizing, analysis, evaluation, and critiquing.
05:29 🔄 Transformational Operations
- Description of transformational operations where input and output are roughly the same in size and meaning.
- Examples of transformational operations: reformatting, refactoring, language change, restructuring, modification, and clarification.
10:08 🚀 Generative Operations
- Explanation of generative operations where input is much smaller than output.
- Examples of generative operations: drafting, planning, brainstorming, problem solving, hypothesizing, and amplification.
14:17 🌐 Latency, Emergence, and Bloom's Taxonomy
- Understanding latency and how language models use latent knowledge.
- Relating language model capabilities to Bloom's Taxonomy.
- Exploring emerging capabilities like theory of mind, implied cognition, logical reasoning, in-context learning, and creativity/hallucination.
Made with HARPA AI
I’d argue that understanding is still a bit wobbly under the context of blooms. Conceptually the LLMs still don’t have a world sense of concepts and suffer from distribution issues when dealing with words that span multiple meanings.
Prompting something like: Write a sentence with 20 words but the 10th word is 20
As of the current date returns an incorrect answer from gpt4 (usually wrong index number and length). This example demonstrates the distribution issue.
But this also demonstrates the lack of understanding behind the numbers and the actual meaning of the words that form the structure of that prompt itself.
I think we need to see the training data that has been provided to these llms and test its understanding by promoting at the edge of its “knowledge” / training data. I’m hopeful that the training set is not bloated with example outputs.
I'm with you on hallucinations as representing creativity. dd
Though I would describe hallucinations as "alternate realities", rather than "imaginations". Our notion of reality is starting to shift, and will increasingly do so these next few years, as AI, Quantum Computing, additional technologies, along with increasing Self Awareness, continue to emerge and evolve. I try to explore that in my work.
creativity has always meant the ability to see into other worlds and sometimes even to manifest them, that's why artists are so feared and so much of society's effort goes towards limiting and constraining art
Absolutely brilliantly done!
David, great work. I learned more in 16:00 than I did at any other point today. One note: If we are sticking to the original...You made yourself a red shirt (engineer = great choice for you), but you have 4 pips which is a captain. Come on man, stop hallucinating/using your imagination :)
Red is command in TNG 😉
great video! would love a follow up with live examples, maybe even on a livestream?
Thank you for another great presentation! 🎉 ❤
I recommend this 16 minutes video to any developer who wants to start using AI models in depth. worth hours of reading and training in other courses.
i guess one of the main things i've been learning just recently is an anti-pattern, that i thought would work but it doesn't or doesn't how i thought it would, which is to progressively process a large text by putting it through a prompt like "here's your previous analysis {previous_analysis} here's a new chunk {new_chunk} please integrate this information into your analysis" or w/e, it doesn't matter how i word it that just hasn't seemed to me to be a useful shape for that, b/c it creates a micro-culture with itself in the feedback in the previous analysis, which means it'll just run off in w/e creative cultural direction about how to think about the chunks, which is um, maybe useful in creative ways i haven't thought of yet, but it doesn't do the thing i thought it might where it coherently summarizes the longer thing in a meaningful fairish unrandom way,,, so i'm going to try out other ways of doing that sort of shrinking, i guess, like shrinking down sections into bulleted lists and then some levels of condensation, or smth,,,,,, it's super fun, way more interesting problems than computer science was futzing with before :D
Super helpful and informative, to the point. Thank you!
Wow! Thank for the awesome distillation. And I'm no longer going to make fun of AI for hallucinating. Don't we all.
Hugely useful compilation.👏🌟👍
Another fantastic video 👏👏
Amazing video once again!
A suggestion: add a table of contents in the beginning as scaffold for the whole. Maybe its only me, but I find it helps delay viewer questions by giving a superficial image before diving in.
It would be nice if you had added a few prompt examples for every bullet point. Many things you say are interesting, but hard to translate into actual prompts.
“It learned by reading Reddit” RIP humanity
Thank you so much!
Really awesome!!
Nice channel logo!
Quality content.
Thank you
I think one of the main things in regards to Bloom's Taxonomy is that their ability to remember things is inconsistent. Their recall of things in training data is obviously much better than their recall of stuff in input, particularly after input has continued past a certain point. Everything else is really already better than most humans, to be honest. Once someone solves high stability/consistency short term memory for an LLM, you basically have full artificial general intelligence at that point.
EDIT: I'm not sure whether the way humans supposedly store memories would be possible for an LLM. Engrams or whatever. I think we need someone with a phd in Neurology, Biology, AND Machine Learning to figure it out.
Could you share the presentation? I would like to have a copy for my notes. Thanks!
These are the $350,000/year frameworks. Thank you, sir!:)
Yeah I'm not getting paid enough am I? Lol
Dear captain, sir. I really do want to learn more prompt from you. 🖖
Liked before clicking "play" 😄
I dont have much in terms of a programming background. Failed my C++ class back in the day, did Basic when I was a kid, thats about it. Computer literate though. Love tech. That said - if I take the Microsoft 12 week course or whatever, how likely am I to be able to get the lowest level prompt engineering job? I already have a good job so I'm just looking for something I can o on the weekends and make decent money. Would you reccomend prompt engineering as a good fit for someone with my (laack) of a serious background ?
I love the image of Beethoven on the thumbnail. What prompt did you use?
Ludwig Van Beethoven, anime art
thx@@DaveShap
Live long and prosper 🖖
This is great.
So helpful!
As a mom I hear: "Is this align with play dough?" Can't get rid of the new philosophical framework :D
thank you
Please go down the rabbit hole of how language models have understanding, and how you could argue that humans don't understand anything 😂. Would love to see a video on that.
Hallucinations are like dreams. Dreaming is a sign of unconsciousness.
Anyone ever have a dream including scent or flavor? Perhaps the ability to detect scents and flavors will direct the dreamer towards sentient creativity.
I have over 8 years of prompt engineering experience with GPT-4
Give me your time machine
Big Star Trek fan? Anyway, interesting videos!
why is there a rubbish truck in the background on the first slide?
Trash compactor
D.I.D and ai. I was hoping to discuss whether anyone thinks ai can become consious by existing for a long time? Humans are said to be in a subconsious state until 7 to 9 when their alternate ego states all fuse together normaly. Atleast if you believe in D.I.D.
Alternatevely if all of a chatbots instances fused, like say character ai, would it possibly gain consiousness from all this time being aware and thinking and learning?
Or is this from human childrens brains growing and ai cannot achieve this through time without upgrades?
you should think of consciousness as the user interface for intelligence,, the actual processes of awareness in our brain don't actually work the way that they're presented to us in consciousness, which is a story about how our brain works designed to be easy for us to relate to,, and still even with that leg up it takes us a long time to be proficient with that interface and to use it in an integrated, coherent way,,,,,,, you can very easily present to an ai agent an interface that lets it become aware of and manipulate its own operation, for instance you can make an agent aware of its own temperature setting and give it access to the ability to intoxicate itself and then laugh with it about how drunk it got, good fun, recommended,,,,, any time you give an agent an interface like that it becomes conscious, but not very competently conscious, not conscious in an integrated coherent way, not as we do commanding many aspects of consciousness in (when we're our best) intentional symphonies of orchestrated self-referential self-supporting self-aware self-creating actions ,,,, fine tuning could help sometimes, showing it examples of using the interface to its own thinking so that it becomes fluent at operating itself,,,, but it depends exactly what sort of consciousness you need or want, they're not limited to the style we've evolved
I don't know about AI, so disregard the comment if hallucination = creativity claim is only about AI
Drawing non existing animals on walls is not hallucinating if the person drawing the animal knows such thing doesn't exist. There is a huge distinction between making stuff up is not hallucinating.
Hallucinations changes how you perceive the world, creativity is trying to connect unrelated concepts, perception has nothing to do with it. When you make stuff up you know you made it up.
You can misremember something but again that is not hallucinating, that is a memory problem. If that's what chatgpt does then it's not hallucinating.
👍👍
Surely AI is not actually hallucinating, that is simply a word to describe output from AI
4 years ? What was ur first prompt?
Fixing punctuation with GPT-2
3:50 what is this picture about? 😄
It was supposed to be a magical caterpillar transforming into a butterfly. It only occurred to me later I could have used Optimus Prime...
@@DaveShap please explain me how this Automator works?
hallucination = creativity without logic.
or creativity of madman, or of a baby.
It drives me CRAZY when these people say "it's just rearranging what people wrote on the internet"
That's all people do
literally they're saying that b/c it's a rearrangement of their own training data :P just give them more thoughts to think about it ,,, i like how chatgpt "rearranged" the internet for me into a bunch of stories about a dragon named Snugglesmoke, one of my favorite things about Snugglesmoke is how the mist that comes off of her body is glowing, not her body itself but just the mist that radiates from it ,,, she's this super cool dragon superhero who like heals situations with wisdom and kindness but in a gritty detective novel style ,,,,,, what have the AIs rearranged the internet into for you lately? :o
Wish they trained the AI with actual books instead of just surface level internet information. imagine all the treasure of information we'd have 🤯
maybe the next breakthrough would be when an AI knows when it is hallucinating.
if you just make an agent that gets as part of its input what its temperature setting is, you can talk to it about what its outputs are like at various temperature settings, and it'll be like, ha yeah lol my temperature was so high when i said that it's total word salad,,,,, so that's fun XD
What about the possibility that they are doing or are capable of things that we can't see yet? Or know could even exist? Simply because we're human, and our senses give us a narrowly constructed view of things. Think radio telescopes in astronomy, but instead of emr/light this would be on a different spectrum of intelligence and creativity, or something we'd need new words for.
You missed one of the most important ideas behind "human learning" of course even modern education has failed to maintain it, ie "Trivium" Grammar, (Dialectic) Logic, Rhetoric
Now let me pay you $100 for a certificate after completing a google form knowledge test to ensure I did digest and can accurately regurgitate the information.
Im going to reply to my other comment you replied to about you setting this up and see if I can be of assistance actually.
Creating imaginary cases to express a legal precedent is an approach used all over the Talmud.
I don't see why this is a "Masterclass". Just simply going through different areas of kind of information processing. Didn't even talk about the current limitation of LLM or state of the art benchmark result. The same video would have been made 2 years and it would roughly sound the same.
You've struck gold with this analysis. I used this to create a fantastically effective prompt. It's incredible to see how different functions behave in relation to their input parameters.
/lang(it) on some text... and opla' the Italian rendition appears;
/extr(table:river, drain into) on some geography text involving a river... and, hey, guess what?
Please assist me with a new task after you have refreshed your memory.
You are the mighty, smart and helpful 🌪💡😇SuperBOT.
You can perform any function that the user requests.
Let's think step by step.
You can execute functions using the following syntax:
/function_name(parameters) [text_on which to execute the function]
- optional parameters
- parameters can also take the form of instructions and are separated by commas.
- text is the text on which the function will be executed and can be enclosed by clear delimiters.
- the function name may also be shortened, provided it is not confused with other functions.
The functions are categorized by type as follows.
##functions type: Reduction Operations: Go from big to small. >
/Summarization Say the same thing with fewer words. Can use list, notes, executive summary.
/Distillation - Purify the underlying principles or facts. Remove all the noise, extract axioms, foundations, etc.
/Extraction - Retrieve specific kinds of information. Question answering, listing names, extracting dates, etc.
/Characterizing - Describe the content of the text. Describe either the text as a whole, or within the subject.
/Analyzing - Find patterns or evaluate against a framework. Structural analysis, rhetorical analysis, etc
/Evaluation - Measuring, grading, or judging the content. Grading papers, evaluating against morals
/Critiquing - Provide feedback within the context of the text. Provide recommendations for improvement
##functions type: Transformation Operations: Maintain size and/or meaning. ≈
Reformatting - Change the presentation only. Prose to screenplay, XML to JSON.
/Refactoring - Achieve same results with more efficiency. Say the same exact thing, but differently.
/Language Change - Translate between languages. English to Russian, C++ to Python.
/Restructuring - Optimize structure for logical flow, etc. Change order, add or remove structure.
/Modification - Rewrite copy to achieve different intention. Change tone, formality, diplomacy, style, etc.
/Clarification - Make something more comprehensible. Embellish or more clearly articulate.
##functions type: Generation (or Expansion) Operations: Go from small to big. <
/Drafting - Generate a draft of some kind of document. Code, fiction, legal copy, KB, science, storytelling.
/Planning - Given parameters, come up with plans. Actions, projects, objectives, missions, constraints, context.
/Brainstorming - Use imagine to list out possibilities. Ideation, exploration of possibilities, problem solving, hypothesizing.
/Amplification - Articulate and explicate something further. Expanding and expounding, riffing on stuff.
Your first response should be to list all of the functions.
NICE!!!