MambaByte, a byte-level language model developed by Cornell University researchers, revolutionizes language models by efficiently managing lengthy byte sequences without traditional tokenization. It significantly outperforms MegaByte, showcasing superior efficiency and results with fewer computational resources. This breakthrough hints at an exciting future for token-free language modeling in natural language processing.

Physicists have developed a new type of neural network using active colloidal particles instead of electricity. This physical system shows promise for artificial intelligence and time series prediction, offering an alternative to traditional microelectronic chip-based digital calculations.

Millions witnessed nonconsensual deepfake pornography of Taylor Swift on social media platform X, prompting the platform to block searches for her. Generating deepfakes with AI has made it easier to sexually harass people. The fight against nonconsensual deepfakes includes using watermarks, protective shields for images, and stricter regulations to hold perpetrators accountable and protect victims.

In September 2022, former Google AI experts Noam Shazeer and Daniel De Freitas released Character.AI, an advanced chatbot. By May 2023, the app had over 1.7 million downloads and high user engagement. As of 2024, it boasts 20 million global users, with 60% aged 18-24. The app’s success raises concerns about mental health support and…

The late comedian George Carlin’s estate is suing the creators of an AI-generated video impersonating Carlin, claiming copyright infringement and violation of Carlin’s right to publicity. It was initially believed that the show was created by an AI, but the creators have stated that it was actually written by a human. The lawsuit raises questions…

Researchers from The Wharton School explored methods to enhance GPT-4’s creativity in idea generation. Experimenting with various prompting strategies, they found that longer prompts and Chain of Thought (CoT) instructions resulted in more diverse ideas. While GPT-4’s ideas were initially similar, strategic prompting improved diversity, making it a valuable tool in brainstorming.

Research on fitness landscapes in evolutionary biology explores the challenge of mapping and understanding the relationship between genotypes and an organism’s fitness. Conventional methods for assessing this complex relationship are limited, prompting the use of deep learning models to predict and analyze fitness. These innovative approaches offer a more scalable and efficient means of studying…

The web is a vast source of knowledge constantly changing, posing challenges for accurate information retrieval. Language models like chatGPT add complexity, leading to research on Retrieval Augmented Language Models (RALMs). San Jose State University proposed TempRALM, using temporal retrieval to enhance Atlas, outperforming it by 74% with fewer resources. Future applications include fact-checking and…

Alibaba researchers introduce DITTO, a self-alignment method enhancing large language models’ role-play capabilities, addressing the limitations of open-source models compared to proprietary ones. Leveraging extensive character knowledge, DITTO outperforms existing baselines, showcasing proficiency in multi-turn role-play conversations. The method opens new possibilities for LLM applications, marking a significant advancement in the field.