September 2023 - AI Tidbits Monthly Roundup

OpenAI's DALL-E 3 and multimodal GPT, novel prompting techniques paving the way to reliable AI, new LLMs for extracting information from complex documents, and a SOTA model from French OpenAI

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.

This is the monthly curated round-up, so if you're pressed for time and can only catch one AI Tidbits edition, this is the one to read—featuring the absolute must-knows.

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

---

Welcome to the September edition of AI Tidbits, where we unravel the latest and greatest in AI. September marked a particularly innovative month, overflowing with groundbreaking research and product launches that showcase the astounding pace of progress in AI.

Leading the charge were new prompting techniques to substantially boost the accuracy of language models while reducing hallucinations. These advances pave the way for more reliable, trustworthy AI systems. On the open-source front, Mistral and Abacus released powerful new models, Mistral 7B and Giraffe 70B, respectively, outperforming previous open-source models such as Llama 2.

Commercially, heavy hitters like OpenAI, Google, and Meta charged ahead with high-profile releases. OpenAI unveiled DALL-E 3 and the long-anticipated multimodal GPT-4 Vision, while Google released a substantial update to its Bard chatbot, and Meta announced its own swarm of LLM-powered chatbots.

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

• Large Language Models

  • Prompting techniques (5 entries)

  • Open-source (8 entries)

  • Research (8 entries)

  • Commercial (5 entries)

• Autonomous Agents (3 entries)

• Image, Audio, and Video (8 entries)

• Multimodal (4 entries)

• Cool Tools (3 entries)

• Open-source (3 entries)

• Other (3 entries)

Recent AI Tidbits Deep Dives

The era of AI-powered SMBs

The Multiprocessor of Language Models

Open-source Generative AI

Large Language Models (LLMs)

Special feature: Prompting techniques

1. DeepMind proposes a prompt that improves LLMs' accuracy by 9% as part of a new method called OPRO - leveraging LLMs to create prompts, outperforming human-designed prompts

2. Meta presents Chain-of-Verification (CoVe) - a method designed to reduce the occurrence of LLMs' hallucinations by enabling them to self-verify their generated responses

3. Salesforce, Columbia, and MIT propose a new prompting technique called Chain of Density (CoD) to improve GPT-4's content summarization capability

4. Anthropic publishes a prompting guide outlining techniques to maximize the performance of its long context window LLM

5. Microsoft and Tsinghua University develop EvoPrompt - an innovative framework that synergizes evolutionary algorithms and LLMs to optimize prompt creation

   

   DeepMind’s new prompt boosts LLMs’ accuracy by +9%

Open-source

1. Mistral, the French OpenAI competitor, releases Mistral 7B - its first model which outperforms all currently available open-source models up to 13B parameters (including Llama 2) on English and coding benchmarks

2. Adept open-sources Persimmon-8B - a commercially permissively licensed LLM outperforming the 7B Llama 2 and MPT models

3. Microsoft releases phi-1.5 - a 1.3B Transformer-based model that rivals larger LLMs and exhibits improved behavior due to the exclusion of web data in its training set

4. Abacus AI open sources Giraffe 70B - a Llama 2 fine-tuned model featuring a 32k tokens context window, outperforming the recent SOTA model LongChat

5. Cerebras Systems open sources a new commercial permissively licensed 3B parameter model called BTLM, equipped with an 8k context window that outperforms all existing 3B parameter models by up to 5%

6. Deci releases DeciLM 6B - a language model that is 15x faster than Llama 2 7B while maintaining comparable performance 

7. John Hopkins and Microsoft open source ALMA - a fine-tuned LLM for machine translation that outperforms Llama 2 on translation tasks

8. LMSYS open sources LMSYS-Chat-1M - a dataset containing one million real-world conversations with LLMs, including diverse use cases like content moderation, safety benchmarks, and instruction-following models

Mistral 7B’s performance compared to Llama 2

Research

1. DeepMind shows that LLMs are powerful lossless compressors, outperforming domain-specific counterparts (PNG, gzip, FLAC) in compressing data