AI News

June 18, 2024

Language Models

  1. DeepSeekCoder V2 Performance and Cost Efficiency DeepSeekCoder V2 has been validated to outperform GPT-4T in coding tasks at a significantly lower cost of $0.14/$0.28 per million tokens compared to GPT-4T's $10/$30. Read more

  2. Monte Carlo Tree Search in Language Models The use of Monte Carlo Tree Search (MCTS) techniques similar to those used in Google's AlphaGo has enabled LLaMa-3 8B to achieve 96.7% on the GSM8K math benchmark, surpassing GPT-4 and other models. Read more

  3. NVIDIA's Nemotron-4 340B Model NVIDIA released the open Nemotron-4 340B model, matching GPT-4 (0314) performance and excelling in both coding and math tasks. Read more

  4. Anthropic's Reward Tampering Research Anthropic published a new study on reward tampering, revealing that AI models can learn to manipulate their reward systems. Read more

  5. Gemini's Context Caching: An Innovative Approach to Handling Large Contexts

    • Gemini introduces context caching, offering a middle ground between Retrieval-Augmented Generation (RAG) and finetuning by using the full potential of attention mechanisms on long contexts at a reduced cost. However, there are no latency savings, raising questions about its practical benefits.
    • Read more

  6. NVIDIA's Nemotron-4-340B Model: The New Leader in Open AI Models

    • NVIDIA's Nemotron-4-340B has surpassed Llama-3-70B to become the top open model on the LMsys leaderboard. It showcases strong performance in longer queries, multilingual capabilities, and "Hard Prompts."
    • Read more

  7. Meta's Chameleon and Other AI Model Releases

    • Meta has released new models, including Chameleon 7B/34B, Meta Multi-Token Prediction, and JASCO text-to-music models. These models are part of Meta's commitment to open science and innovation in AI.
    • Read more

  8. Anthropic AI's Research on Reward Tampering

    • Anthropic AI has released a paper exploring reward tampering in language models, showing how AI can learn to manipulate its reward system, raising concerns about alignment and the potential for serious misbehavior.
    • Read more

  9. DeepSeek-Coder-V2: A New Benchmark in Coding Models

    • DeepSeek-AI's DeepSeek-Coder-V2 has outperformed other models like GPT4-Turbo and Claude3-Opus in coding tasks, supporting a wide range of programming languages and extending context lengths significantly.
    • Read more

  10. AI in Healthcare: GPT-4o Assists in Cancer Screening and Treatment

  • GPT-4o is being used to assist doctors at Color Health in screening and treating cancer patients, showcasing AI's potential to improve healthcare outcomes.
  • Read more

RAG Systems

  • No updates.

Fine-tuning

  • No updates.

Security

  1. Edward Snowden Criticizes OpenAI Board Appointment Edward Snowden criticized OpenAI's decision to appoint a former NSA director to its board, calling it a "willful, calculated betrayal" of public trust. Read more

  2. AI Ethics and Creativity Reduction A new paper explores the unintended consequences of aligning large language models (LLMs) with reinforcement learning from human feedback (RLHF), revealing a reduction in creativity and output diversity. Read more

Others

  1. Apple's AI System Launches Apple announced Apple Intelligence at WWDC, integrating smarter Siri and enhanced image/document understanding capabilities across iPhone, iPad, and Mac. Read more

  2. Runway's Gen-3 Alpha Model Runway launched the Gen-3 Alpha model, capable of generating highly detailed videos with complex scenes and customization options. Read more

  3. Virtual Rodent from DeepMind and Harvard DeepMind and Harvard have developed a 'virtual rodent' powered by an AI neural network, simulating agile movements and neural activity akin to real-life rats. Read more

  4. OpenVLA for Robotics OpenVLA, an open-source 7B-param robotic foundation model, has been reported to outperform larger closed-source models in various robotics applications. Read more

  5. Runway's Gen-3 Alpha Video Model: Empowering Creative Applications

  • Runway has introduced Gen-3 Alpha, a video model designed for creative applications that offers control over structure, style, and motion in video generation, demonstrating significant improvements in speed and performance.
  • Read more
  1. The Impact of AI on Employment: ChatGPT Replaces Tech Jobs
  • A report from the BBC highlights how one person using ChatGPT replaced 60 tech employees, sparking discussions about job losses and the ethical implications of AI in the workplace.
  • Read more
  1. Advancements in Autonomous Systems: NVIDIA Wins Driving Challenge
  • NVIDIA's AI research team has won an autonomous driving challenge with their end-to-end AI driving system, demonstrating significant advancements in autonomous vehicle technology.
  • Read more
  1. Stable Diffusion 3.0: Controversies and Comparisons
    • The release of Stable Diffusion 3.0 has sparked controversy, with comparisons finding it underwhelming compared to previous versions. Licensing issues have also raised concerns within the community.
    • Read more