July 2023 - AI Tidbits Monthly Roundup

Meta's new SOTA open-source LLM takes the internet by storm, OpenAI's Code Interpreter and GPT-4 released to all, Google's new multimodal medical AI, and step change improvements in image + video AI

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 July edition of AI Tidbits, where we unravel the latest and greatest in AI. This month we saw major progress on LLMs with Meta’s release of Llama 2, which performs competitively with ChatGPT and Bard, allows commercial use, and can run on consumer hardware. OpenAI opened GPT-4 and Code Interpreter to all ChatGPT Plus users, and Anthropic released an improved version of its Claude LLM, supporting up to 200k tokens.

At the same time, Stanford and Berkeley researchers showed ChatGPT has been degrading in performance since March, and Carnegie Mellon researchers managed to jailbreak proprietary models like ChatGPT and Claude.

Lastly, Google DeepMind released a new multimodal version of its cutting-edge medical LLM Med-PaLM, surpassing specialist models by a wide margin and showing examples of zero-shot generalization to novel medical concepts and tasks.

Let's dive in!

Large Language Models (LLMs)

Open-source

1. Meta releases Llama 2 - a new version of its open-source LLM which outperforms previously released LLMs such as Falcon and MPT and is available for commercial use 

2. Stability AI announces two new powerful LLMS - FreeWilly1 based on LLaMA 65B and FreeWilly2 based on Llama 2 70B, achieving ChatGPT-like performance

3. LangChain releases LangSmith - a unified platform for debugging, testing, evaluating, and monitoring of LLM applications 

4. Alignment Lab releases OpenOrca - a dataset that brings GPT-4 reasoning to open models

   

Commercial

1. Anthropic releases Claude 2 - an improved version of its LLM that can support up to 200k tokens and is considered to be the main contender to OpenAI's GPT 

2. OpenAI releases Custom Instructions, equipping ChatGPT with the ability to remember users' preferences across threads

3. OpenAI opens access to GPT-4 and Code Interpreter to all of its paying users 

   

Research

1. Stanford and Berkeley show ChatGPT has been degrading in performance since March

2. Carnegie Mellon researchers manage to jailbreak proprietary models like ChatGPT and Claude by finding and appending a suffix that maximally misaligns the underlying language model

3. Microsoft proposes RetNet - a foundation architecture for LLMs achieving training parallelism and low-cost inference compared to Transformer while achieving strong performance

4. Researchers develop 3D-LLM - equipping LLMs with a 3D world understanding to conduct 3D-related tasks, including captioning, Q&A, and navigation 

5. DeepMind presents WebAgent - an LLM-driven agent for autonomous web navigation substantially outperforming previous methods 

6. Microsoft AI introduces LongNet - a transformer variant scaling token length to over 1 billion tokens

   

Multimodal

1. Google DeepMind releases Med-PaLM Multimodal - a large multimodal AI that interprets biomedical data, including language, imaging, and genomics