June 2023 - AI Tidbits Monthly Roundup

The rise of open-source LLMs with MPT-30B and Falcon, Meta's AI turning text into music, Google's StyleDrop capable of copying a specific style based on one image, and more!

Welcome to a subscriber-only edition 🔒 of AI Tidbits, helping you stay on top of the recent in AI and leverage the latest breakthroughs to grow personally and professionally. If you find AI Tidbits valuable, share it with a friend, and consider showing your support.

---

Welcome to the June edition of AI Tidbits, where we unravel the latest and greatest in AI. This month we witnessed the rise of further SOTA open-source language models, with the likes of MPT-30B and Salesforce's XGen-7b. In the realm of multimodal AI, Microsoft published Otter, an exceptional model that masters the art of perceiving and reasoning across various modalities.

On the image and video front, Google released StyleDrop, a versatile method that allows the generation of images in specific styles with outstanding quality. Midjourney released v5.2, granting users the power to extend the canvas of images beyond their original boundaries, and Runway unlocked access to Gen-2, its revolutionary text-to-video model.

Meta AI made waves with the release of MusicGen, a beautifully simple and controllable model for music generation based on text prompts and input melodies.

Lastly, MosiacML's staggering $1.3B acquisition, followed by Inflection’s remarkable $1.3B funding round, demonstrates that the flow of investment into AI shows no signs of slowing down. Let's dive in!

Large Language Models (LLMs)

Open-source

1. MosaicML introduces MPT-30B - an open-source commercially licensed LLM with an 8k context window

2. Salesforce introduces XGen-7B - a new 7B 8k context window LLM achieving comparable or better results with SOTA open source LLMs such as MPT and Falcon

3. LMSYS releases a new series of Vicuna models, including a 33B LLM model surpassing MPT and Guanaco 33B

4. Tsinghua University releases ChatGLM 2-6B - a substantially better version of its predecessor open-source bilingual (Chinese-English) chat model, adding support for 32k tokens

5. Berkeley releases OpenLLaMA 13B - a permissively licensed open source reproduction of Meta AI’s 13B trained on the RedPajama dataset

6. Researchers introduce LLM-Blender - a novel ensembling framework that combines multiple open-source LLMs for consistently superior performance

Research

1. Microsoft showcases Orca - a 13B model that imitates the reasoning process of commercial LLMs such as GPT-4, surpassing Vicuna's performance and achieving ChatGPT-like performance

2. Microsoft introduces phi-1 - a smaller and more efficient language model for code, trained on textbook quality data and synthetic exercises 

3. UCSB and Microsoft propose Language Models Augmented with Long-Term Memory (LongMem), enabling LLMs to memorize and utilize long chat sessions 

4. MIT and Microsoft show that only GPT-4 is capable of self-improving upon its generations compared with smaller models such as GPT-3.5 

5. MIT and Stanford introduce a red teaming framework for LLMs that surfaces prompts eliciting toxic and dishonest statements

Autonomous agents