Next Week in The Sequence

• Edge 283: Our series about federated learning(FL) continues exploring the concepts of FL with differential privacy and we review Meta AI’s proposed architecture in this area. Additionally, we review the best differential privacy frameworks for FL models.

• Edge 284: We discuss Dolly, Databricks’ open source foundation model.

📝 Editorial: Big vs. Small, Open Source vs. API Based, the Philosophical Frictions of Foundation Models

Innovation is the foundation of the models space, which is accelerating at a frantic pace, and we are seeing new models popping up everywhere. While the space is still in a very nascent state, we are already seeing conflicting forces that will be highly influential in the evolution of the market. Currently, two major frictions are influencing different philosophical divisions in the foundation model space:

1. The friction between massively large and smaller models.  

2. The friction between open-source and API-based distribution.

Regarding foundation models, the rule that larger is better has proven true for the last few years. Larger models simply exhibited cognitive capabilities that were not possible with smaller architectures. However, in recent months, we have seen the emergence of models like LLaMA and variations with RLHF that have been able to come close to matching the performance of larger alternatives.

The second friction is the generative AI version of the iOS vs. Android debate. Models like GPT-4, LaMDA, and Claude are distributed via commercial APIs, while models like Dolly 2 and Stable Diffusion are distributed via open-source models. The rationale of this debate goes beyond the commercial model and encompasses aspects such as fairness and safety concerns.

The most surprising dynamic of the two frictions at the center of the evolution of foundation models is that they are not defining four camps, but rather two. In a not surprising coincidence, the vendors favoring super large models also rely on API-based distribution, while the open-source models are also relatively smaller. On one camp, we have OpenAI, Anthropic, Microsoft, or Google, while in the other, we can currently see Databricks, Stability AI, and maybe Meta.

Are these two market frictions the same? I personally don’t think so. In the near future, we are likely to see open-source distributions of super large models or smaller models only available via APIs. But also, let’s remember that generative AI is different from any other market.

📣 Re-tooling around LLMs?

Share your POV! Take this one-minute survey and be entered for a chance to win a swag pack or O’Reilly books! Take the flash poll.* 

🔎 ML Research

HuggingGPT

Microsoft Research published a paper detailing HuggingGPT, a framework that uses language models to connect various foundation models for diverse AI tasks. HuggingGPT uses ChatGPT to determine the tasks to execute in ML models for specific tasks —> Read more.

Text-Guided Video Generation

Google Research published a paper proposing UniPi, a model that can learn a diverse, universal policy in text t- video models. UniPi can be seen as a universal interface for inferring actions in videos based on text descriptions —> Read more.

Animated Drawings

Meta AI Research published a paper and open source a dataset to streamline the animation of amateur drawings. The dataset includes 180,000 animated pictures and a demo that could encourage researchers to innovate in this area —> Read more.

🤖 Cool AI Tech Releases

Cache LLM Queries with GPTCache

GPTCache, an MIT-licensed open-source semantic cache, is now available to reduce your ChatGPT bill and improve the performance of your LLM app —> Read more.

Amazon Bedrock

Amazon finally unveiled its play in the generative AI space with the release of Bedrock, a platform that enables interaction with several foundation models —> Read more.

Auto-GPT

A super interesting open source experiment that attempts to make GPT4 more autonomous —> Read more.

StackLLaMA

Hugging Face open sourced StackLLaMA, a model based on Meta AI’s LLaMA and fined tuned on Stack Exchange questions —> Read more.

Dolly 2.0

Databricks open sourced the second version of its Dolly model, a 12B instruction following model which is available for commercial use —> Read more.  

Stable Diffusion SDXL

Stability AI released Stable Diffusion XL Beta as part of its DreamStudio, a new model with enhanced capabilities including the ability to generate legible text in images —> Read more.

DeepSpeed Chat

Microsoft Research open sourced DeepSpeed Chat, a framework for large scale RLHF training of LLMs —> Read more.

🛠 Real World ML 

Responsible AI at LinkedIn

LinkedIn discusses some of the best practices for ensuring responsible AI usage in their applications —> Read more.

Generative AI for Compliance at GitHub

The GutHub engineering team presents some ideas about generative AI can be used for software development compliance tasks —> Read more.  

ML Computing Costs at Lyft

Lyft’s engineering team outlines some of the best practices used to save compute costs in their ML and big data workloads  —> Read more.

📡AI Radar

• Elon Musk is in active talks to build an Open AI competitor à Read more.

• Data as a service provider Cybersyn raised a $69.2M series A led by Snowflake.  

• VideoVerse acquired generative video company Reely.ai.

• InfoGrid raised $90 million for its AI-based building monitoring tech.

• Quora’s conversational app Poe launched the ability to create new chatbots based on prompts.

• Dystil AI announced a $7 million fundraise and a partnership with OpenAI to incorporate generative AI into large scale enterprise scenarios.

• OpenAI announced a bug bounty program.