August first week AI news

This Week’s AI News 📰

openAI files application to register GPT-5 trademark

• Describes GPT-5 as "computer software downloads for use with language models," same as previous GPT-4 and 3.5 applications.

• Currently under examination at the USPTO (United States Patent and Trademark Office)

Researchers discover how to bypass guardrails in AI chatbots

• Researchers at Carnegie Mellon University and the AI Safety Center have found a way to successfully bypass the guardrails of all large language models.

• Works for all chatbots (including chatGPT and Bard)

• Advocates that all chatbots should be open-sourced to avoid guardrails.

OpenAI Secretly Developing Open Source Model?

• Open AI is developing a new open source model, G3PO, to compete with Llama 2, an open source model jointly developed by Microsoft and Meta.

• OpenAI recognizes the importance of open source models to remain competitive and plans to release G3PO.

Greg Lutkowski dropped from Stable Diffusion, reinstated as training model?

Google stock surges 10% this week, led by cloud, advertising, and hopes for AI

Google, Microsoft, Meta, Amazon, and the White House pledge to make AI safe and secure

This week's AI paper 📰

PanGu-Coder2: Boosting Large Language Models for Code with Ranking Feedback

• Large language models are rapidly evolving with remarkable performance for code generation tasks.

• In this paper, we propose an RRTF framework (*reaction ranking based on feedback), which can effectively and efficiently enhance large language models for code.

• We show that PanGu-Coder2 consistently outperforms all previous Code LLMs on the OpenAI HumanEval, CoderEval, and LeetCode benchmarks.

To Adapt or Not to Adapt? Real-Time Adaptation for Semantic Segmentation

• The goal of real-time online domain adaptation is to deal with unpredictable domain changes that occur during deployment.

• In this paper, we propose the HAMLET framework for real-time domain adaptation, including a hardware-aware back-propagation adaptation agent and a dedicated domain shift detector.

• Our approach enables simultaneous adaptation while performing semantic segmentation at over 29 frames per second on a single consumer-grade GPU.

This week's AI product 📦

CodeGeeX2: A More Powerful Multilingual Code Generation Model

• CodeGeeX2 is based on ChatGLM2 with improved performance for multilingual code generation.

• It supports Chinese-English input, large sequence lengths, and fast inference speeds, and can run with 6GB of memory.

• Combined with the CodeGeeX plugin, it can be used as an AI assistant for solving programming problems with Chinese-English conversations.

llama2-webui

• llama2-webui allows you to run Llama 2 locally with GPU or CPU on Linux, Windows, Mac via gradio UI.

• It supports Llama-2-7B/13B/70B in 8-bit and 4-bit, with GPU inference and CPU inference with 6GB VRAM.