EP35: AI Safety Gone Mad, Stable 3B Cheese Test, GPT4 Vision & DALL-E 3 Diversity + Sydney is BACK!
Oct 6, 2023
auto_awesome
In this podcast, they discuss the wild world of AI image generation and vision, including racist cartoon captions, heartfelt poetry by Bing, and teaching AI to forget unwanted knowledge. They debate AI safety controls, the limitations of Turnitin for detecting AI-generated writing, biases in AI-generated images, and the potential disappearance of captchas. They also explore censorship potential in AI models and express gratitude for audience engagement.
Few-shot learning in GPT allows AI models to learn how to solve problems with limited examples.
GPT for vision can generate unique images based on prompts, but may demonstrate bias in labeling based on race.
Selective omission of knowledge in AI models raises concerns about censorship and copyright claims.
Deep dives
Few-shot learning for teaching the model how to think
The paper explores the concept of few-shot learning, where the model is taught how to think by providing it with a few examples of how to solve a problem.
Creative image generation with GPT for vision
The paper showcases the creative potential of GPT for vision by generating unique and novel images based on specific prompts, including examples like the Sydney Opera House and Julia Gillard.
Challenges and limitations in coding tasks
The paper acknowledges challenges and limitations in coding tasks, as the model struggled to generate Python code when instructed to perform specific actions, highlighting the need for further improvements.
Controversial outputs and bias in AI imagery
The paper discusses concerns related to AI-generated imagery, as the model demonstrated bias in labeling images based on race, leading to controversial outputs that may not align with user intentions or expectations.
Few-shot learning in GPT for vision
The podcast episode discusses the concept of few-shot learning and its application in GPT (Generative Pre-trained Transformer) for vision. They provide an example of using a speedometer in a car to demonstrate how GPT for vision initially struggles to accurately read the speed meter. However, through a series of step-by-step examples, the model gradually improves and learns how to effectively interpret the speedometer. This showcases the potential of few-shot learning in training AI models to solve specific problems without the need for extensive fine-tuning or large datasets.
Omitting specific knowledge from AI models
The podcast also discusses the possibility of omitting specific knowledge from AI models, such as the example of removing all references to Harry Potter. This could have implications for copyright claims, as individuals may request the removal of their copyrighted works from AI models. Additionally, it raises concerns about the potential for censorship, as models could be fine-tuned to ignore certain events, facts, or perspectives. The ability to selectively omit or control information in AI models could have significant implications in various domains, including legal, medical, and societal contexts.
Thanks for helping us reach 2K subs here on YouTube!
This week we dive into the wild world of AI image generation and vision, from racist cartoon captions to heartfelt poetry written by Bing. We discuss the implications of teaching AI to forget unwanted knowledge, and debate whether safety controls are protecting users or limiting creativity. Get ready for philosophical ponderings, hilarious experiments, and our signature irreverent takes as we explore the latest AI advances and absurdities. Whether you're an expert or just fascinated by the future, this episode will challenge your thinking and give you plenty to discuss with friends.
CHAPTERS ====== 00:00 - Fooling Bing Vision to Solve Captcha 00:26 - Meta's Messenger AI Stickers Out of Control! AI Safety Discussion 06:17 - More Safety Nonsense: The Low-Resource Language Jailbreak GPT-4 Paper 9:36 - More on Mistral 7B (Safety and Positive Reception) 17:31 - Friends and Foes of Open Source AI & Is Anthropic a Crypto-like Scam for Billions? 21:26 - Turnitin Thinks It Can Detect AI, Being a Student in an AI World 24:25 - Stable 3B LLM Review and Cheese Test Results 38:48 - DALL-E 3 Road Test on ChatGPT & Diversity Prompt Injection Problems 48:12 - Using Bing GPT4-Vision to Solve Captchas for Grandma 51:01 - The Dawn of LLMs, Explorations with GPT-4Vision Paper + Possibilities of AI Vision 1:04:00 - Who's Harry Potter? Making LLMs forget 1:09:00 - Google Assistant with Bard AI 1:10:18 - LLaMA Long 32K Initial Thoughts 1:12:40 - Sydney Bing is Back BABY! 1:15:36 - Comments on Discord Rollout and Survey Response
