News 16 July 2023 - simon-oz/Weekly-AI-news GitHub Wiki

Jul 9 – 14, ACL 2023, Annual Meeting of the Association for Computational Linguistics (ACL), which is one of the top natural language processing conferences in the world, was held in Toronto, Canada.

Jul 11, Anthropic AI released Claude 2, Claude 2 has improved performance, longer responses, and can be accessed via API as well as a new public-facing beta website, claude.ai. The latest model scored 76.5% on the multiple choice section of the Bar exam, up from 73.0% with Claude 1.3.

Jul 11, according to Bloomberg, AI Researcher Who Helped Write Landmark Paper Is Leaving Google, and the last one who is still working for Google, is departing Google, and will start a company after taking time off.

Jul 12, Elon Musk announced xAI, is to understand the true nature of the universe. The new team members have previously worked at DeepMind, OpenAI, Google Research, Microsoft Research, Tesla, and the University of Toronto. Collectively the team contributed some of the most widely used methods in the field, in particular the Adam optimizer, Batch Normalization, Layer Normalization, and the discovery of adversarial examples. xAI aims at implement AGI by the end of 2029, the due date.

Jul 12, businessinsider reported that UC Berkly prof. Stuart Russell warned that AI developers are "running out of text" to train chatbots at a UN summit, and AI's strategy behind training large language models is "starting to hit a brick wall." It also reported that a group of AI researchers, estimated that machine learning datasets will likely deplete all "high-quality language data" before 2026.

Jul 12, Google published a paper on Nature “Large language models encode clinical knowledge”. The paper proposes a human evaluation framework for model answers along multiple axes including factuality, comprehension, reasoning, possible harm and bias. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA3, MedMCQA4, PubMedQA5 and Measuring Massive Multitask Language Understanding (MMLU) clinical topics6), including 67.6% accuracy on MedQA (US Medical Licensing Exam-style questions), surpassing the prior state of the art by more than 17%.

Jul 12, Nature published a paper “Quantum-enhanced Markov chain Monte Carlo”. Researchers have developed a quantum algorithm that can sample from complicated distributions, such as MCMC, arising in several applications. This algorithm is well-suited to current hardware and could ease computational bottlenecks in machine learning, statistical physics, and optimization.

Jul 13, Hashingtonpost reported that the Federal Trade Commission has opened an expansive investigation into OpenAI, probing whether the maker of the popular ChatGPT bot has run afoul of consumer protection laws by putting personal reputations and data at risk. The FTC’s demands of OpenAI are the first indication of how it intends to enforce those warnings. If the FTC finds that a company violates consumer protection laws, it can levy fines or put a business under a consent decree, which can dictate how the company handles data. The FTC in its request also asked the company to provide extensive details about its products and the way it advertises them. It also demanded details about the policies and procedures that OpenAI takes before it releases any new product to the public, including a list of times that OpenAI held back a large language model because of safety risks.

Jul 14, Meta published a paper “Scaling Autoregressive Multi-Modal Models: Pretraining and Instruction Tuning”. The new model, named CM3Leon, is a retrieval-augmented, tokenbased, decoder-only multi-modal language model capable of generating and infilling both text and images. CM3Leon achieves state-of-theart performance in text-to-image generation with 5x less training compute than comparable methods.