AI Tidbits May Round-Up

Google's host of generative AI updates, Anthropic's 100k context window, a new commercially permissible LLM seizing the 1st place on the LLM Dashboard, and much 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 May edition of AI Tidbits, your monthly dose of awe-inspiring developments in the dynamic world of artificial intelligence. This month didn’t fall short of amazing breakthroughs: Google's host of generative AI updates (PaLM 2, Bard Plugins), Anthropic’s new 100k context window, a new commercially permissible LLM seizing the 1st place on Hugging Face’s LLM Dashboard, and an autonomous agent that learns skills on the fly using GPT-4 to dominate Minecraft.

Also, healthcare has seen much progress with the release of Google’s Med-PaLM 2 - the first language model to achieve expert-level performance on U.S. Medical Licensing Examination-style questions, and a paper showing that ChatGPT’s responses are preferred over human physicians.

This progress has also warranted industry leaders and distinguished AI researchers to warn against the potential harms of AI, urging lawmakers to regulate AI development. Let's dive in!

Large Language Models (LLMs)

Commercial

1. Google infuses generative AI across its product lines (summarized Google I/O keynote):

   1. PaLM 2 released with improved multilingual, reasoning, and coding capabilities

   2. Bard Tools - connecting Bard to Google products (Gmail, Docs, Sheets, Maps) and third-party services such as Instacart

   3. Search - AI-powered snapshots of key results, with links to dig deeper

   4. Gmail - generating full emails in one click

   5. Photos - Magic Editor to remove, replace, and recreate parts in pictures

   6. Google Workspace is getting Duet AI - a suite of improvements such as automatic table generation in Sheets and image creation in Slides and Meet

   7. Gemini - a secret collaboration between Google and DeepMind to release a next generation multimodal model

2. Inflection introduces Pi, an Ai-powered chatbot designed to be better at a “natural flowing” conversation

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

4. Google rolls out Med-PaLM 2, Google's variant of the PaLM model that is optimized for medical questions

Open-source

1. TII open sources Falcon - a commercially permissible LLM outperforming open-source models like MPT-7B, StableLM, and RedPajama, seizing the first spot in Hugging Face's Open LLM Dashboard

2. Meta AI unveiled LIMA - a fine-tuned version of its 65B LLaMA achieving Bard-level performance and having strong generalization capabilities on unseen tasks

3. Together releases MosaicML’s MPT-7B, a new family of open-source commercially permissible LLMs, outperforming LLaMa-7B and other 13B models

4. Together open sources 3B and 7B RedPajama-INCITE models, including base, instruction-tuned, and chat models, outperforming the Pythia 7B model

5. Stability AI releases StableVicuna, the first large-scale open-source chatbot trained with human feedback

6. Microsoft Research introduces Gorilla - a fine-tuned LLaMA model explicitly designed for API calls, surpassing the performance of GPT-4 in writing API calls

7. OpenLM open sources OpenLLaMA - a commercially permissible reproduction of Meta AI's LLaMA

8. SambaNova and Together present BLOOMChat - an open-source commercially permissible multilingual chatbot achieving a win rate of 45.25% compared to GPT-4's 54.75% in a human preference study 

9. LLaVA-Adapter V2 was released, achieving ChatGPT-like performance in language tasks and offering advanced visual understanding capabilities

10. MLC releases MLC LLM - a framework allowing any LLM to be deployed natively on a diverse set of platforms, including iPhones and iPads

11. Hugging Face unveils StarCoder - a ChatGPT-like coding model reaching 40.8% on HumanEval benchmark

12. Researchers release Clinical Camel - a 13B fine-tuned LLaMA outperforming GPT-3.5 on the US Medical Licensing Examination