Chapters discuss challenges in training multilingual speech synthesis models, the development of generative AI tools like Lily, the launch of new features in Runway and Adobe Express, Snapchat's Dreams feature and selfie-to-headshot apps, the economics of image generation and Gartner's hype cycle for generative AI, funding trends in the AI industry, the development of AI2 OLMo and Edy Thix, generative assessment and model training, and the impact of China's new AI regulations on global AI development.
Read more
AI Summary
Highlights
AI Chapters
Episode notes
auto_awesome
Podcast summary created with Snipd AI
Quick takeaways
The Allen Institute for AI is developing a large language model called AI-Dalmo, aiming to provide accessibility and education to the research community.
OpenAI has acquired AI design studio Global Illumination to contribute to their core products.
Arthur has introduced Benchmark, a suite of monitoring tools to evaluate and compare large language models, addressing the challenge of model comparison.
China has implemented provisional guidelines for AI activity, including platform registration, security reviews, and content labeling.
Deep dives
OpenAI to release a 70 billion parameter open source language model
The Allen Institute for AI is developing a 70 billion parameter language model called AI-Dalmo, which will be released in early 2024. The model aims to provide access and education to the research community, with all elements of the project being made accessible, including data, code, training curves, and evaluation benchmarks. The model is being developed in collaboration with AMD, CSC, and other partners, and is expected to make a significant impact in the AI research space.
OpenAI acquires AI design studio Global Illumination
OpenAI has made its first public acquisition, acquiring AI design studio Global Illumination. The New York-based startup will join OpenAI to work on core products. The details of the acquisition and the specific focus of the acquired team have not been disclosed.
Arthur releases Benchmark, a suite of monitoring tools for large language models
Arthur, a company focused on large language models, has released Benchmark, a suite of monitoring tools for businesses to integrate into their operations. Benchmark includes an open-source evaluation tool that allows businesses to compare large language models, prompts, and hyperparameters. The tools aim to address the challenge of evaluating and comparing different language models.
Hugging Face introduces Eddythix, a reproduction of DeepMind's Flamingo model
Hugging Face has introduced Eddythix, an open reproduction of DeepMind's Flamingo model. By combining LAMA 1 and OpenCLIP models, Eddythix is able to generate text outputs based on text and image inputs. The open-source nature of Eddythix allows for transparency and accessibility in the field of visual language models.
China's New AI Regulations
China has implemented provisional guidelines for AI activity and management, which require platforms to register and undergo security reviews. The guidelines also mandate labeling for AI-generated content and hold service providers accountable for the content created through their platforms.
AP's AI Guidelines for Journalists
The Associated Press (AP) has set guidelines for the use of AI in journalism. While AP allows journalists to experiment with AI models like ChatGPT, they are not permitted to use it to create publishable content. Additionally, output from AI models should be treated as unvetted source material and subject to AP's existing sourcing standards.
Issues with AI Detection Tools
AI detection tools, such as plagiarism checkers, are falsely accusing international students of cheating. The tools have shown bias against non-native English speakers and label a significant percentage of papers as AI-generated when they are not. These false positives can have severe consequences for students.