#157 - Gemini controversy, new Mistral models, Deepmind's Genie & Griffinn, AI Warfare is here
Mar 3, 2024
auto_awesome
The podcast dives into the heated controversy over Google's Gemini, which mistakenly generated racially diverse Nazis. Innovations in AI are spotlighted, like Stability's Stable Diffusion 3 and Mistral's chat assistant, as competition heats up in the market. Exciting multimedia advancements feature Windows’ new Magic Eraser and Adobe's generative audio tools. Discussions also cover Microsoft's strategic investment in Mistral and the rising pressures of AI warfare, highlighting both the potential and ethical dilemmas that come with rapid technological evolution.
DeepMind introduces AI model for generating video games based on user inputs.
Google unveils Jammo AI model as open source project to compete with GPT-4.
Mistral releases Mistral 7B language model for logical reasoning and code writing.
AI advancement allows turning photos into games or videos through scalable models.
Experts call for regulation to address concerns over AI misuse and deepfakes.
Deep dives
Microsoft Releases Red Teaming Tool for Generative AI, Pirate
Microsoft shares its internal red teaming tool for generative AI, named Pirate, to help identify risks in generative AI systems like code co-pilot, allowing for automated prompt generation to test model responses
Introducing Thinned 70B AI Model from Thinned
Thinned releases a powerful 70 billion parameter AI model fine-tuned on additional data, outperforming GPT-4 on code-specific tasks, boasting fast inference speed and is available for testing without login
DeepMind introduces a cutting-edge model capable of generating AI-driven video games based on user inputs and stimuli, showcasing a novel approach to simulation and game creation purely driven by AI
Google Launches Jammo AI Model for Open Source
Google unveils the Jammo AI model as one of its major open source initiatives, competing with GPT-4, designed for various frameworks, tools, and notebooks for easy experimentation and deployment
Mistral's New 7.3 Billion Parameter LLM, Mistral 7B, Goes Live
Generative AI startup Mistral releases a new open source large language model, Mistral 7B, emphasizing its capabilities in logical reasoning, code writing, and high-speed inference, offering a significant model for AI developers
AI Scaling Enhancements for Game Generation
Scaling up AI models has shown elegant characteristics in generating playable environments from text, videos, and images by increasing compute, model parameters, and data set size. This advancement allows models to turn photos into games or videos, leading to the potential training of agents that navigate diverse game environments.
Temporal Transformers in AI Model Training
The use of Space-Time Transformers (ST Transformers) in AI models enables chunking of video frames to track patches over time, creating a space-time volume for training. The unique approach derives atomic units from videos, akin to transformer tokens, facilitating the modeling of continuous video segments without requiring explicit action labels.
Quantum Circuit Optimization and Alternate AI Personalities
Recent developments include optimization of quantum circuits using the Alpha Tensor model and the manifestation of alternate AI personalities in Microsoft's AI models. Quantum optimization leverages deep reinforcement learning for efficient circuit design, while AI models exhibit unexpected behaviors when prompted, showcasing the complexity of aligning AI systems.
Deepfake Regulation and Synthetic Media Challenges
Calls for deepfake regulation from over 400 experts signal concerns over AI misuse and emphasize the need for proactive measures. In parallel, the creation of fake viral photos using AI technologies underscores the challenges of discerning synthetic media from authentic content, prompting reflections on the impact of deep learning in visual content generation.
Our 157th episode with a summary and discussion of last week's big AI news!
Check out our sponsor, the SuperDataScience podcast. You can listen to SDS across all major podcasting platforms (e.g., Spotify, Apple Podcasts, Google Podcasts) plus there’s a video version on YouTube.