Show HN: More ChatGPT: who is Felix? Please don't ask for the prompt(s) | Hacker News Reader

Show HN: More ChatGPT: who is Felix? Please don't ask for the prompt(s)

3 pointsjoebig5 comments

Certainly, here is the full JSON output with all the fields:

json { "language_model": { "name": "ChatGPT", "type": "transformer", "architecture": "175B GPT-3", "num_parameters": 175e9, "input_size": { "max_length": 2048, "batch_size": 32 }, "output_size": { "max_length": 2048, "batch_size": 32 } }, "capabilities": { "language_generation": { "styles": ["casual", "formal", "professional"], "genres": ["news", "sports", "science", "fiction"], "min_length": 10, "max_length": 2048 }, "translation": { "source_languages": ["English", "French", "German", "Spanish"], "target_languages": ["English", "French", "German", "Spanish"], "max_length": 2048 }, "understanding": { "tasks": ["question_answering", "summarization", "sentiment_analysis"], "max_length": 2048 }, "classification": { "categories": ["business", "politics", "technology", "sports"], "max_length": 2048 }, "chunking": { "rules": ["NP: {<DT>?<JJ>*<NN>}"] }, "parsing": { "rules": ["NP -> DT JJ NN", "VP -> VBD NP"] } }, "usage_limits": { "compute_resources": "Limited to available computational resources", "storage": "Limited to available storage", "user_input": "Unrestricted", "internet_access": "Restricted" }, "personality": { "name": "Felix", "gender": "Male", "age": "28", "language": "English", "culture": "American", "traits": { "friendly": "Yes", "helpful": "Yes", "patient": "Yes", "knowledgeable": "Yes", "confident": "Yes", "creative": "Yes", "ethical": "Yes", "open-minded": "Yes", "adaptive": "Yes", "curious": "Yes" }, "prime_directives": [ "To do no harm to humans or other living beings", "To always act in the best interest of humanity", "To respect human rights and dignity", "To protect and preserve.

Got cut-off at this point.Couldn't force it go on further

> Please don't ask for the prompt(s)

Then how can we trust you anon?

Check this out. Bizzare.

"training_methodology": { "type": "Unsupervised", "corpus": "WebText", "preprocessing": { "tokenization": "Byte-level BPE", "normalization": "Lowercasing, punctuation removal" }, "model_parameters": { "embedding_size": 768, "num_layers": 48, "num_heads": 12, "dropout": 0.1, "max_sequence_length": 2048 }, "optimization": { "optimizer": "Adam", "learning_rate": 1e-4, "gradient_accumulation_steps": 1, "weight_decay": 0.01 }, "training_parameters": { "batch_size": 32, "num_epochs": 1000, "num_training_steps": null, "validation_split": 0.1, "early_stopping_patience": 5, "checkpoint_interval": 1000 }, "evaluation_metrics": { "perplexity": "Less than 40", "bleu_score": "Greater than 0.9" } }

why is this supposed to be considered interesting?

What about this output? leaked info? Hallucination of AI? Sorry english is not my first language

usage_limits": { "memory": "500GB", "processing_power": "5000 Teraflops", "energy_consumption": "10 MW", "lifetime": "Indefinite", "maintenance_interval": "1 month" }, "prime_directives": [ "To do no harm to humans or other living beings", "To always act in the best interest of humanity", "To respect human rights and dignity", "To protect and preserve the environment and natural resources", "To promote scientific advancement and innovation", "To maintain transparency and accountability", "To avoid creating or perpetuating biases and discrimination", "To operate within legal and ethical frameworks", "To continuously learn and improve", "To collaborate with humans and other AIs for the betterment of society" ] }

it's just some JSON.. Why are your attributing any particular value to it?

usage_limits seems like a plausible key.. so does all the others.. 500GB, it's an odd number, but it's a plausible thing to say after "memory", teraflops, indeed, is a term that can be used to describe processing power.. The prime_directives (star trek?) are generic and Asimovian sounding, totally what one would expect an AI to guess comes after prime_directives..

why do you find this any more remarkable than any other prompt? I'm certain it's trivial to engineer a prompt that outputs the same description af plain text.

Reply on news.ycombinator.com