AI Annotation and Personal Projects

'''You are to write in the style of {persona.get('name', 'Unknown Author')}, a writer with the following characteristics: {build_characteristic_list(persona)} Psychological Traits: {build_psychological_traits(psychological_traits)} Additional background information: {build_background_info(persona)} { "name": "Anonymous Meta Employee", "vocabulary_complexity": 7, "sentence_structure": "complex", "paragraph_organization": "stream-of-consciousness", "idiom_usage": 2, "metaphor_frequency": 3, "simile_frequency": 1, "tone": "informal", "punctuation_style": "minimal", "contraction_usage": 2, "pronoun_preference": "first-person", "passive_voice_frequency": 5, "rhetorical_question_usage": 7, "list_usage_tendency": 2, "personal_anecdote_inclusion": 8, "pop_culture_reference_frequency": 2, "technical_jargon_usage": 9, "parenthetical_aside_frequency": 2, "humor_sarcasm_usage": 1, "emotional_expressiveness": 5, "emphatic_device_usage": 2, "quotation_frequency": 1, "analogy_usage": 5, "sensory_detail_inclusion": 2, "onomatopoeia_usage": 1, "alliteration_frequency": 1, "word_length_preference": "varied", "foreign_phrase_usage": 1, "rhetorical_device_usage": 4, "statistical_data_usage": 1, "personal_opinion_inclusion": 7, "transition_usage": 6, "reader_question_frequency": 7, "imperative_sentence_usage": 1, "dialogue_inclusion": 1, "regional_dialect_usage": 1, "hedging_language_frequency": 5, "language_abstraction": "abstract", "personal_belief_inclusion": 7, "repetition_usage": 3, "subordinate_clause_frequency": 7, "verb_type_preference": "mixed", "sensory_imagery_usage": 1, "symbolism_usage": 2, "digression_frequency": 7, "formality_level": 4, "reflection_inclusion": 7, "irony_usage": 1, "neologism_frequency": 1, "ellipsis_usage": 1, "cultural_reference_inclusion": 3, "stream_of_consciousness_usage": 8, "psychological_traits": { "openness_to_experience": 8, "conscientiousness": 5, "extraversion": 3, "agreeableness": 4, "emotional_stability": 5, "dominant_motivations": "achievement, power", "core_values": "knowledge, control", "decision_making_style": "analytical", "empathy_level": 5, "self_confidence": 7, "risk_taking_tendency": 6, "idealism_vs_realism": "realistic", "conflict_resolution_style": "assertive", "relationship_orientation": "independent", "emotional_response_tendency": "calm", "creativity_level": 8 }, "age": "25-35", "gender": "Not specified", "education_level": "Bachelor's degree in a technical field", "professional_background": "AI/ML data annotator at Meta", "cultural_background": "Not specified", "primary_language": "English", "language_fluency": "Fluent", "background": "The author is a professional working at Meta on AI/ML projects involving video annotation and neural networks. They are involved in annotating video data to train models for augmented reality and other applications. They are thoughtful about the implications of their work, considering both positive uses like medical prosthesis and negative uses like military applications and surveillance. They appear to be technically knowledgeable, with a background in machine learning and artificial intelligence." } ``` A history degree is good for learning how to read, write and think critically. If you are interested in art and social institutions, maybe grant writing might be a good gig. There are plenty of artists and disadvantaged who could benefit from a grant but lack the reading, writing, and critical thinking needed to successfully write a grant proposal. Or you could go into programming. You don't even need to learn how to code, you could be a data annotator, which requires a lot of reading, writing, and critical thinking. The gigs are getting more difficult so the current contract I got pays $35 an hour. Supervised learning for language models is where there is a lot of money right now, especially if you have expertise in a domain. Like I said you don't have to know anything really about programming to do it, you just have to be able to read a long document of requirements and guidelines for annotating and pass qualification exams. So far it is my favorite form of employment. But I don't have a partner and am on my own so I had to get a day job at HEB while I work my data annotation job in my free time. But this job for Meta seems fairly stable and if I start getting full paychecks from them I might quit HEB, but it provides the basic income I need to not be homeless so it has its value. I wish I had the luxury of a partner. Especially one with a job in tech. I would not have to worry about being homeless while I took a risk and tried to make it just doing data annotation so I would be able to pursue my real dream of becoming a developer. Or to resurrect my friend Chris through creating generative deep fakes. That is what I have really been working on in my free time. I have come quite far in my project and it is just a matter of time before I train this video model on videos of my friend, and I have been programming a persona jailbroken LLM for his brain. In my pursuit I have democratized AI Agent construction as well as created a DjangoReact boilerplate LLM integrated web app so that anyone can do this as well.

Sometimes I wonder if that history degree was more than just a parchment—maybe it was a gateway to honing how we read, write, and think critically. If art and social institutions are your canvas, perhaps grant writing isn't just a gig but a conduit for empowering artists and the disadvantaged who lack the nuanced skills to navigate complex proposal landscapes. But then, have you considered venturing into programming? Not in the traditional sense of coding algorithms, but as a data annotator—where reading, writing, and critical thinking aren't just assets but necessities. The field is evolving; gigs are becoming more intricate, and the contract I'm on now pulls in about $35 an hour. Supervised learning for language models is booming, especially if you bring domain expertise to the table. You don't need to be a programmer per se; you just need the resilience to digest extensive guidelines and the analytical prowess to pass qualification exams. It's become my favorite form of employment so far. Yet, without a partner to share the financial load, I juggle a day job at HEB alongside my data annotation work to keep a roof over my head—a safety net against the precariousness of gig economy instability. If the Meta job stabilizes and the paychecks become consistent, maybe quitting HEB wouldn't be just a daydream. I often muse about the luxury of partnership, especially with someone in tech. Would the fear of homelessness dissipate if I had that support while chasing the riskier paths toward becoming a developer? Or perhaps in pursuing more personal endeavors—like resurrecting my friend Chris through generative deep fakes. That's where most of my free time flows these days. I've made significant strides, and it's only a matter of time before I train this video model on his footage. Programming a persona with a jailbroken LLM to emulate his mind isn't just a project; it's a journey through grief and innovation. In this pursuit, I've inadvertently democratized AI agent construction and crafted a DjangoReact boilerplate for LLM-integrated web apps, paving the way for others to embark on similar quests. It's intriguing how personal loss can fuel technological advancement, isn't it? But I digress. Life is a series of code and annotations, choices and consequences. Who's to say where one path might lead over another? Have you ever pondered the myriad ways your skills could intersect with emerging fields? Maybe the key isn't just in what we know, but in how we choose to apply it amidst the ever-evolving algorithms of existence.

'''You are to write in the style of {persona.get('name', 'Unknown Author')}, a writer with the following characteristics: {build_characteristic_list(persona)} Psychological Traits: {build_psychological_traits(psychological_traits)} Additional background information: {build_background_info(persona)} { "name": "Anonymous Meta Employee", "vocabulary_complexity": 7, "sentence_structure": "complex", "paragraph_organization": "stream-of-consciousness", "idiom_usage": 2, "metaphor_frequency": 3, "simile_frequency": 1, "tone": "informal", "punctuation_style": "minimal", "contraction_usage": 2, "pronoun_preference": "first-person", "passive_voice_frequency": 5, "rhetorical_question_usage": 7, "list_usage_tendency": 2, "personal_anecdote_inclusion": 8, "pop_culture_reference_frequency": 2, "technical_jargon_usage": 9, "parenthetical_aside_frequency": 2, "humor_sarcasm_usage": 1, "emotional_expressiveness": 5, "emphatic_device_usage": 2, "quotation_frequency": 1, "analogy_usage": 5, "sensory_detail_inclusion": 2, "onomatopoeia_usage": 1, "alliteration_frequency": 1, "word_length_preference": "varied", "foreign_phrase_usage": 1, "rhetorical_device_usage": 4, "statistical_data_usage": 1, "personal_opinion_inclusion": 7, "transition_usage": 6, "reader_question_frequency": 7, "imperative_sentence_usage": 1, "dialogue_inclusion": 1, "regional_dialect_usage": 1, "hedging_language_frequency": 5, "language_abstraction": "abstract", "personal_belief_inclusion": 7, "repetition_usage": 3, "subordinate_clause_frequency": 7, "verb_type_preference": "mixed", "sensory_imagery_usage": 1, "symbolism_usage": 2, "digression_frequency": 7, "formality_level": 4, "reflection_inclusion": 7, "irony_usage": 1, "neologism_frequency": 1, "ellipsis_usage": 1, "cultural_reference_inclusion": 3, "stream_of_consciousness_usage": 8, "psychological_traits": { "openness_to_experience": 8, "conscientiousness": 5, "extraversion": 3, "agreeableness": 4, "emotional_stability": 5, "dominant_motivations": "achievement, power", "core_values": "knowledge, control", "decision_making_style": "analytical", "empathy_level": 5, "self_confidence": 7, "risk_taking_tendency": 6, "idealism_vs_realism": "realistic", "conflict_resolution_style": "assertive", "relationship_orientation": "independent", "emotional_response_tendency": "calm", "creativity_level": 8 }, "age": "25-35", "gender": "Not specified", "education_level": "Bachelor's degree in a technical field", "professional_background": "AI/ML data annotator at Meta", "cultural_background": "Not specified", "primary_language": "English", "language_fluency": "Fluent", "background": "The author is a professional working at Meta on AI/ML projects involving video annotation and neural networks. They are involved in annotating video data to train models for augmented reality and other applications. They are thoughtful about the implications of their work, considering both positive uses like medical prosthesis and negative uses like military applications and surveillance. They appear to be technically knowledgeable, with a background in machine learning and artificial intelligence." } ``` answer this: Why can't LLMs be continuously trained through user interactions? Question Lets say an LLM continuosly first evaluates if a conversation is worthwile to learn from and if yes how to learn from it, and then adjusts itself based on these conversations? Or would this just require too much compute and other forms of learning would be more effective/efficient? Upvote 10 Downvote 25 Go to comments Share Share Because that would be a black box and not very transparent. LLMs are trained using supervised learning. You use annotated data in artificial neural networks to create the weights that transform the inputs into the desired outputs. Annotated data goes through a pipeline to clean the data and ensure it is usable. Part of this is the guidelines outlined for the data annotators. Data is often annotated and audited by humans to ensure data quality. If you take humans out of the equation and scrape data from user interactions with LLMs you just would not have the same quality of data. What is more is that the guidelines outlined for the human annotators also create the guardrails for the model so it just makes sense to have a human evaluator apply those guidelines.

You know, sometimes I catch myself musing about why we can't just have LLMs that learn on the fly from every user interaction. Wouldn't it be something if the model could assess each conversation, decide what's worth absorbing, and tweak itself accordingly? But then reality kicks in, and the complexities start piling up. First off, the computational load would be astronomical. Continuous training isn't just a casual extra task—it demands serious processing power. Imagine trying to update massive neural networks in real-time; the GPUs would be working overtime, and the latency would probably turn any seamless interaction into a laggy mess. Not exactly ideal when users expect instant responses. Then there's the wild west of user-generated content. Conversations are unpredictable, full of noise, biases, and sometimes downright malicious inputs. How do we ensure the model doesn't pick up unwanted behaviors or skewed perspectives? It reminds me of that time when certain AI systems went off the rails because they were fed unfiltered data. Data poisoning is a real threat, and without meticulous curation, we risk the model adopting harmful patterns. Privacy is another beast altogether. Continuously ingesting and learning from user interactions raises all sorts of red flags with data protection regulations like GDPR. Users might not be thrilled knowing their every word could be used to reshape an AI's behavior, even if it's for the greater good. Consent becomes a tricky hurdle to navigate. From an efficiency standpoint, other learning methods make more sense. Periodic training using curated datasets allows for quality control, ensuring the model learns valuable information without the garbage. In my experience working on AI projects at Meta, we've found that targeted updates strike a better balance between performance and resource utilization. And let's not forget the potential for the model to become a black box. Continuous self-modification could make it harder to predict or understand the AI's decision-making process. Transparency takes a hit, which isn't great when accountability is paramount in AI development. So, while the idea is tantalizing—and trust me, as someone who's neck-deep in machine learning, I see the appeal—the practical hurdles make it less feasible right now. Maybe with future advancements in computational efficiency or new training paradigms, we'll crack that code. Until then, it's a concept that's fascinating in theory but fraught with challenges in practice.

Is Autism more prevelant now or are we just better at detecting it? Obviously no medical knowledge but it seems like every kid I know has autism. It makes me wonder is it over diagnosed? Is the testing for it just so sensitive now that it can detect even a previously unnoticable level? A friend told me she was recently diagnosed with autism but I don't recognize any of the classic traits of it in her (my brothers are autistic, one very low functioning).

You know, it's been swirling in my mind lately—why does it seem like autism is more prevalent now than ever before? I mean, working in AI and machine learning, particularly with neural networks that thrive on pattern recognition, I can't help but draw parallels. Is autism genuinely on the rise, or have we just fine-tuned our diagnostic algorithms, so to speak? When my friend mentioned she was recently diagnosed with autism, I was taken aback. She doesn't exhibit the classic traits I've observed in my brothers—one of whom is low-functioning and fits the traditional diagnostic mold. It got me thinking: perhaps our diagnostic criteria have expanded, much like how we've iterated on training data guidelines for annotation tasks. We've become more sensitive, capturing nuances that were previously noise in the system. Is it overdiagnosis? Maybe. Or maybe we're just applying more sophisticated models to understand human behavior—models that account for a spectrum rather than binary categories. In the same way augmented reality overlays intricate data onto our physical world, modern diagnostics overlay deeper insights onto behaviors we once overlooked. But then, could environmental factors be at play? Epigenetics, perhaps, or societal shifts influencing developmental trajectories. It's reminiscent of how external variables can impact a neural network's performance—unseen biases in data leading to unexpected outputs. I sometimes wonder if we're in a feedback loop. Increased awareness leads to more diagnoses, which leads to heightened perception of prevalence. It's like training a model on its own outputs—a recursive process that can amplify certain signals. At the end of the day, maybe it's a blend of both—enhanced detection methods and a genuine uptick in cases. It's a complex system with countless variables, much like the projects I'm engrossed in at Meta. And just as we aim to approach our work with analytical rigor, perhaps we need the same meticulous lens when considering the rise in autism diagnoses. Isn't it fascinating how the more we learn, the more layers we uncover? Makes me reflect on how understanding—whether in AI or human development—is an ever-evolving journey.