OpenAI has curtailed public access to GPT-4's image-recognition abilities as it can potentially recognize specific individuals. The feature could potentially infringe upon privacy laws in regions where the use of biometric information requires explicit consent from users. It could also misrepresent aspects of people's faces, leading to inappropriate or harmful results. There have been instances of the model hallucinating responses, giving users inaccurate information. Microsoft has been testing a limited rollout of the visual analysis tool in its Bing chatbot.
Meta's Llama 2 is a new open-source family of AI language models with a license that allows it to be integrated into commercial products. The models range in size from 7 to 70 billion parameters and they reportedly outperform other open-source chat models on most benchmarks. Llama 2 is available on Microsoft Azure and will be available on AWS, Hugging Face, and other providers. A link to a research paper released by Meta with details on Llama 2's performance, benchmarks, and construction is available in the article.
An anesthetist renowned for their ability to spot dodgy data in medical trials estimates that 44% of trials contain at least some flawed data and that 26% of papers have flaws so large that the studies can't be trusted. These studies resemble real research, but close scrutiny will show that the authors were incompetent or that the data was faked. Without access to raw trial data, even experienced investigators may miss hidden flaws. Journals need to assume that all submitted papers are potentially flawed and review the data before publishing them.
Rocket Lab successfully retrieved the first stage of its Electron rocket during a mission on Monday. The first stage separated two and a half minutes after Electron lifted off and splashed down into the Pacific Ocean. Rocket Lab aims to reuse its boosters for future missions. The company had planned and attempted to catch its boosters midair using helicopters, but it has changed its procedure to simply collecting the stage from the ocean after tests showed that the boosters fared better than expected after being exposed to water.
This article discusses common issues that people don't really talk about when developing deep learning apps. Docker images for AI apps can be enormous and take a long time to configure to work correctly in production environments. They can take forever to build and start and it requires a lot of study to learn how to build them. The fast pace of AI development has led to a special kind of dependency hell where it is difficult to create reproducible environments. Different cloud providers use different architectures, making development requirements much more difficult.
AutoChain aims to enable rapid iteration on generative agents by simplifying agent customization and evaluation. Building and evaluating custom generative AI agents can be difficult and time-consuming with the current generation of tools and frameworks. AutoChain attempts to solve these issues by providing a lightweight and extensible framework, custom tools, and automatic evaluation of agents. It shares similar concepts with LangChain, making it easy for developers familiar with LangChain to navigate.
Daily active users for Meta's Threads dropped from 49 million to 23.6 million between July 7 and July 14. Threads' release caused Twitter's traffic to go down 5% in the first two days, and while it has since bounced back, Twitter's traffic is still down 11% year-over-year. Twitter's user retention has been on the decline. Threads is still missing many basic features and a compelling reason for Twitter users to switch over.
Meta's Llama 2 will be able to run on Qualcomm chips on phones and PCs starting in 2024. Large language models have primarily run on large server farms so far due to the technology's need for vast computational power and data. While Qualcomm's chips are well-suited for the kinds of calculations that AI models require, these chips will still have significantly less power than data centers stocked with cutting-edge GPUs. Qualcomm and Meta have worked together in the past, most notably on chips for Meta's Quest virtual reality headsets.
Code Interpreter allows its users to create, edit, and analyze charts using English and it automatically sets up Python environments so users can focus on their work.