AI Roundup 05/11 -> 05/18/2023

A suite of new techniques to improve LLMs from Microsoft, 100k-tokens prompts, another simulation of humanity using autonomous agents, and OpenAI's generative AI Senate hearing

1. Claude, a GPT competitor from Anthropic, announces its support of a 100k context window, tripling GPT-4’s maximum of 32k (Anthropic Blog)

2. Researchers from South Korea release DarkBERT, a dark web domain-specific LLM aimed to advance research on identifying cyber security threats like ransomware, leak site detection, and more (Paper)

3. Microsoft presents guidance, a framework that enables much greater control over LLMs than traditional prompting or chaining techniques (Github repo)

4. Google open-sources Dreamcast, allowing users to control a PC with their face and head movements by analyzing 468 facial tracking points (The Verge)

5. Microsoft researchers present TinyStories, a synthetic dataset of short stories that is used to train and evaluate smaller language models (<10M parameters) (Paper)

6. Researchers from MIT present Dr. LLaMA, a technique that improves Small Language Models (SLMs) by using LLMs to generate data for medical Q&A tasks (Paper)

7. Google releases SoundStorm, a model capable of synthesizing 30 seconds high quality, natural dialogues with just half a few seconds of audio samples (Paper's Github repo)

8. Stability AI releases Stable Animation, an open-source text-to-animation tool for developers (Stability AI)

9. Researchers from CMU propose a novel approach that enables the generation of high-quality text-to-video content using multiple sketched frames as input (Paper)

10. Researchers introduce InstructBLIP, an open-source model that improves vision-language models through instruction tuning, achieving SOTA performance (Hugging Face)

11. A developer created a customizable group of AI agents with independent personalities and memories, inspired by Stanford's experiment of simulating a virtual town of AI agents (Twitter)

12. Researchers from the University of Michigan introduce a transformer-based model that predicts the location and properties of 3D objects (Paper's website)

13. Google rolls out Med-PaLM 2, Google's variant of the PaLM model that is optimized for medical questions, to select customers for initial testing (Paper)

14. Google researchers introduce a new framework to enable diverse language-based interactions with mobile UI including prompting, screen question-answering, summarization, and more (Google AI)

15. Meta unveils AI Sandbox - a testing ground for advertisers to create different versions of their ad copy, background, and image outpainting (TechCrunch)

16. Microsoft and USD introduce Super In-Context Learning, a novel approach that improves the performance of LLMs by combining them with smaller locally fine-tuned models (Paper)
    

Open source

1. Bard-API: an open-source package to interact with Google's Bard (Github repo)

Announcements

1. Apple introduces new generative AI based accessibility capabilities like Live Speech and Personal Voice (Apple)

2. OpenAI roll out in beta more than 70 third-party Plugins for ChatGPT across multiple industries like travel, food, real estate, and more (Twitter)

Thought-provoking

1. Sam Altman, OpenAI's CEO, urges the US Congress to impose regulation on large-scale AI systems, including OpenAI (Forbes)

Last week’s AI Tidbits

If you find AI Tidbits valuable, share it with a friend, and consider showing your support.