ChatGPT User’s Account Compromised, Exposing Pharmacy Data
A ChatGPT user who complained of seeing someone else’s data in their account reportedly had their account compromised, according to OpenAI officials. They explained, “From what we discovered, we consider it an account take over in that it’s consistent with [the] activity we see where someone is contributing to a ‘pool’ of identities that an external community or proxy server uses to distribute free access.” The official did not mention OpenAI’s plans to address the lack of two-factor authentication (2FA) or IP location tracking for current logins, features commonly found on other major platforms. The leaked conversations included usernames and passwords related to a pharmacy prescription drug portal’s support system.
Via Ars Technica
LASER: Microsoft’s Technique for Sharpening the Accuracy of Language Models
Microsoft researchers have introduced a technique named Layer-Selective Rank Reduction (LASER) to boost the precision of big language models (LLMs). LASER works by swapping out a chunk of the LLM with a smaller version, keeping the important stuff intact. The surprising part is that this doesn’t mess up the model; it actually makes it better by reducing errors and improving accuracy. Tests on popular models like RoBERTa and Eleuther’s GPT-J showed a 20-30% accuracy boost. It’s a nice helpful upgrade, fixing mistakes AI sometimes makes – like hallucinations or just getting stuff wrong.
Via The Verge
Nvidia Uses AI to in Chip Design
Nvidia, the chipmaker of choice for AI, has created a custom AI model called ChipNeMo to accelerate the chip design process. Building GPUs is labor-intensive and involves close to 1,000 people who must understand various aspects of the design process. ChipNeMo, which runs on a large language model based on Meta’s Llama 2, can answer questions related to chip design, including GPU architecture and chip design code generation. It remains to be seen whether ChipNeMo has achieved its goal since its unveiling in October, but it has shown promise by helping to train junior engineers and summarizing notes across 100 teams.
Via Business Insider
Open Source AI Chatbot Builder Takes on ChatGPT with Customization and Free Access
Hugging Face is the latest company to create a ChatGPT rival: Hugging Chat Assistants. This free product allows users to create customized AI chatbots, similar to custom GPTs in the paid ‘Plus’ version of OpenAI’s offering. Users can select from various open-source LLMs, including Mistral’s Mixtral and Meta’s Llama 2, to power their AI Assistant’s intelligence. Some people in the open-source AI community view Hugging Chat Assistants as superior to GPTs due to their greater customizability and the fact that they are free to use. However, custom GPTs still excel in web search, retrieval augmented generation (RAG), and logo generation using DALL-E 3.
Via Venture Beat
Amazon Launches Product to Assist People in Spending More Money on Amazon
Rufus is Amazon’s new generative AI-powered shopping assistant integrated into its mobile app, currently only available for U.S. customers. Trained on a massive dataset of Amazon’s products, reviews, and web info, Rufus is like having a personal shopping consultant in your pocket, aiding your greatest consumerist desires. The chatbot answers various shopping-related questions, including product comparisons, recommendations, and specific product inquiries. So now you have more help in your quest to join the air fryer cult.
Via Amazon
AI Illuminates Ancient Wisdom Enshrined in Charred Roman Scroll
Scholars, aided by computer-savvy students, have harnessed the power of AI to partially decipher a charred Roman scroll buried by the cataclysmic Vesuvius eruption in 79 AD. This breakthrough offers a glimpse into ancient beliefs about food and pleasure. The Herculaneum Papyri, unearthed in Southern Italy, bore the scars of time, their charred remains defying conventional methods of deciphering. However, Professor Brent Seales and his team employed X-ray technology and machine learning to uncover meanings from minute differences in the rolled-up documents. The project aims to decode more scrolls, potentially rewriting the history of key periods in the ancient world. The team’s ambitious goal is to unlock 85% of the scroll by the end of the year and help reshape our understanding of the past.
Via Silicon Angle