- Tune AI
- Posts
- 💬 Fastest Speech Model Ever...
💬 Fastest Speech Model Ever...
💥 Cross-Continent Innovations, OpenAI Privacy Concerns, and GenAI Metrics Overhaul
Hello Tuners,
This week is dedicated to Sam Altman and OpenAI. Every single day I tried to catch up with GenAI News, it has all been derived or influenced innovations from OpenAI. However, I was glad to see that we finally have a Speech Model that works in practice and not just in demos in the form of Moshi from Kyut.ai.
Moving forward we also saw a potential alternative for Chinese GenAI Developers to slow ChatGPT VPN APIs, Sam Altman’s new Adventure, and OpenAI defining intelligence metrics to modern LLMs.
French AI developer Kyut.ai launched Moshi, a real-time voice assistant powered by the Helium 7B model. Designed to provide lifelike conversations, Moshi supports various accents and 70 different emotional speaking styles. It can handle two audio streams simultaneously, enabling it to listen and talk at the same time, similar to Alexa and Google Assistant.
Moshi’s development involved fine-tuning over 100,000 synthetic dialogues using Text-to-Speech technology. Kyutai collaborated with a professional voice artist to enhance the AI’s voice quality, ensuring more nuanced and human-like communication. This launch follows the delay of OpenAI's ChatGPT Voice Mode, potentially offering a competitive alternative for users seeking advanced voice interactions.
At the World AI Conference in Shanghai, SenseTime unveiled its latest model, SenseNova 5.5. The model demonstrated capabilities like identifying and describing objects, providing feedback on drawings, and summarizing text. Comparable to OpenAI's GPT-4, SenseNova 5.5 aims to attract users with its advanced functionalities and competitive performance.
To further entice users, SenseTime is offering 50 million free tokens and free migration support for clients switching from OpenAI services. Notably, this move comes as OpenAI announced it will block Chinese users from its tools starting July 9, making it easier for SenseTime to capture the domestic market, which is rather curious as despite ChatGPT’s ban in China, the Chinese Developers have been a big part of OpenAI’s Userbase through VPNs.
OpenAI CEO Sam Altman and Arianna Huffington announced Thrive AI Health, a startup focused on AI-driven health coaching through hyper-personalization. The app, developed in cooperation with Huffington’s Thrive Global, aims to improve behavioral health by targeting sleep, food, movement, stress management, and social connection based on biometric data and personal preferences.
The health coach will provide personalized recommendations, such as sleep reminders based on calendar data. By leveraging AI's ability to process vast amounts of data, the startup aims to enhance healthy behaviors and overall well-being, ensuring technology serves to benefit users' health. Notably, the application is being developed in collaboration with OpenAI, which just means that is just one more sphere of your data, the company may have now!
OpenAI introduced a five-level system to track its progress in developing AI that can outperform humans. Shared at an all-hands meeting, the levels range from today's conversational AI (Level 1) to AI capable of managing an organization (Level 5). Currently, OpenAI is at Level 1, approaching Level 2, "Reasoners," which can solve problems like a doctorate-level human without tools.
This classification aims to enhance understanding of AI development and safety. The system will be shared with investors and the public, highlighting OpenAI's commitment to transparency in AI progress. The initiative underscores the company's leading role in advancing powerful AI systems.
LLM Of The Week
OpenAI has developed CriticGPT, a model based on GPT-4, to catch errors in ChatGPT's code output. Testing showed that users assisted by CriticGPT outperformed those without help 60% of the time. OpenAI plans to integrate CriticGPT-like models into their RLHF labeling pipeline, aiding AI trainers in evaluating AI outputs more effectively.
Reinforcement Learning from Human Feedback (RLHF) aligns GPT-4 models like ChatGPT to be helpful and interactive. As ChatGPT's accuracy improves and mistakes become subtler, AI trainers face challenges in spotting inaccuracies. Integrating CriticGPT aims to overcome this limitation, enhancing the evaluation process for advanced AI systems.
Weekly Research Spotlight 🔍
Data curation via joint example selection further accelerates multimodal learning
Google DeepMind introduced JEST, a novel AI training method claiming 13 times faster performance and 10 times higher power efficiency than current techniques. JEST optimizes training by using a smaller model to evaluate data quality, which then informs the training of a larger model, leading to significant improvements in training speed and energy efficiency.
This method's efficiency hinges on high-quality training data, as it uses joint example selection to prioritize the best data batches. While JEST's reliance on expertly curated datasets might limit accessibility for hobbyists, its potential to reduce AI's environmental impact is a significant step forward for the industry.
Best Prompt of the Week 🎨
A close-up of the face, featuring neon orange and black lines that swirl around her eyes like lightning. The background is a deep blue with subtle yellow highlights, creating an otherworldly atmosphere. Her skin has intricate patterns in gold, giving it depth and texture. In one eye shines pink light, while another glows purple, adding to its mysterious allure. She wears a futuristic outfit, blending seamlessly into the vibrant backdrop. Stylish in the style of Anna Dittmann. --ar 51:64 --s 250 --v 6.0
Today's Goal: Try new things 🧪
Acting as a Daily Activity Planner
Prompt: I want you to act as a daily activity planner. You will create a structured daily plan specifically designed to help individuals with ADHD manage their time effectively and overcome procrastination. You will identify a target client profile, develop key strategies and action plans, select the tools and resources for efficient time management, and outline any additional activities needed to support focus and productivity. My first suggestion request is: "I need help creating a daily activity plan for someone with ADHD who struggles with procrastination and needs to improve their daily productivity."
This Week's Must-Read Gem 💎
How did you find today's email? |