August 2023 - AI Tidbits Monthly Roundup

Meta's open source launch spree (Code Llama, DINOv2, Nougat), autonomous agents make their way into commercial applications, and TII unveils Falcon 180B - the largest and SOTA open LLM to date

Welcome to a subscriber-only edition 🔒 of AI Tidbits, where I curate the firehose of AI research papers and tools so you won’t have to. If you find AI Tidbits valuable, share it with a friend, and consider showing your support.

---

Welcome to the August edition of AI Tidbits, where we unravel the latest and greatest in AI. August saw the rise of coding language models with Meta’s release of the commercially permissive Code Llama, followed by a host of other SOTA coding models such as WizardCoder 34B and DeciCoder.

UAE brought us the first Arabic LLM (Jais) as well as the largest open-source LLM to date, Falcon 180B, outperforming Meta’s Llama 2 and on par with Google’s Bard. Moreover, an open-source alternative to OpenAI’s renowned Code Interpreter was released.

Also in August, Stanford open sourced the code for its last March’s groundbreaking autonomous agents paper, with a16z releasing AI Town to help developers jumpstart AI simulation environments.

These and many more exciting updates across multimodal AI, video models, and open-source tools are part of this month’s roundup.

Let's dive in!

---

Overview

• August Deep Dives (3 entries)

• Large Language Models

  • Coding LLMs (8 entries)

  • Open-source (9 entries)

  • Research (2 entries)

  • Commercial (4 entries)

• Autonomous Agents (5 entries)

• Image and Video (6 entries)

• Multimodal (3 entries)

• Cool Tools (6 entries)

• Open-source (6 entries)

August Deep Dives

The Multiprocessor of Language Models

The future of Internet Search in the era of LLMs

Open-source Generative AI

Large Language Models (LLMs)

Special feature: Coding LLMs

1. Meta releases Code Llama - a commercially permissible SOTA model built on top of Llama 2, fine-tuned for generating and discussing code

2. WizardCoder 34B - a fine-tuned language model outperforming ChatGPT, Claude 2, and previous SOTA open-source language models on coding tasks 

3. Stability AI announces StableCode - its first coding LLM supporting multiple programming languages and a 16k context window

4. Refact releases Refact Code LLM - a 1.6B coding LLM outperforming similar-sized coding models across 20 programming languages

5. Deci releases DeciCoder 1B - a code completion model trained on Python, JS, and Java, outperforming SantaCoder

6. Hugging Face introduces SafeCoder - an enterprise coding assistant with a fully compliant and self-hosted pair programmer

7. Researchers present OctoPack, harnessing Git commits to fine-tune LLMs, leading to SOTA coding task performance with OctoCoder and OctoGeeX

8. Defog open sources SQLCoder - a state-of-the-art LLM for SQL generation outperforming GPT-3.5

Meta’s Code Llama

Open-source

1. TII releases Falcon 180B - the largest open source LLM to date with 3.5 trillion tokens outperforming Llama 2 and on par with Google's PaLM 2 

2. A developer releases Open Interpreter - an open-source alternative to OpenAI's Code Interpreter that runs locally

3. Boston University open sources Platypus - a family of finetuned LLMs achieving the top score on the Hugging Face Open LLM Leaderboard 

4. Meta AI introduces Nougat - a visual transformer model for OCR, aiming to convert scientific PDFs, including mathematical expressions, into markup language

5. Meta AI introduces a new method to automatically generate instructions using Llama called Instruction Backtranslation

6. Meta drops SeamlessM4T - a translation multimodal model capable of translating text in ~100 languages and speech in 35 languages 

7. UAE open sources Jais and Jais-chat - the first Arabic language models with 13B parameters

8. Stability AI releases Japanese StableLM - a base and instruction fine-tuned models achieving SOTA compared to other Japanese models

9. Clinical Camel-70B - a new open-source medical LLM for clinical research

Open Interpreter

Research

1. Google shows that RL from AI Feedback (RLAIF) performs on par with Reinforcement Learning from Human Feedback (RLHF), offering a potential solution to the scalability limitations of RLHF

2. DeepMind introduces Reinforced Self-Training (ReST) - a method to efficiently align LLMs with human preferences, significantly enhancing machine translation quality using offline reinforcement learning

Google’s RLAIF vs. RLHF performance

Commercial

1. OpenAI announces GPTBot - a web crawling bot that extracts website data to train and improve OpenAI's models, providing website owners the option to block them from using their data

2. OpenAI announces a GPT 3.5 Turbo fine-tuning API, with GPT-4 fine-tuning to follow in a couple of months