Next Week in The Sequence

• Edge 277 our series about federated learning(FL) continues with an overview of federated transfer learning, the TorchFL paper and the OpenFL framework.

• Edge 278 provides a deep dive of Google’s famous LaMDA model.

📝 Editorial 

In the last three weeks, OpenAI has delivered three major releases. We started with the release of the ChatGPT APIs three weeks ago. That was followed by the anticipated launch of GPT-4. Last week, OpenAI delivered yet another dazzling release by unveiling ChatGPT plugins, an extensbility framework that enables the combination of ChatGPT capabilities with up-to-date information sources. The plugins framework integrates ChatGPT with popular platforms such as Kayak, OpenTable and many others. Furthermore, the framework provides a programming model that allows developers to create their own plugins.

The pace of releases by OpenAI is astonishing but also speaks to an interesting phenomenon we are seeing in the generative AI space. After certain scale, foundation models can become reusable building blocks of new models or platform capabilities in a way that ressembles traditional software developments. This might sound obvious if you are new to machine learning(ML) but will make perfect sense if we remember that the traditional approach to ML assume that every new model needed to be built and trained from the ground up. OpenAI built GPT-3.5 on top of GPT-3, ChatGPT is based on GPT-3.5 and GPT-4 is based on ChatGPT. Foundation models are certainly shorting the release cycles in ML to levels we never seen before.

OpenAI was not the only company with impressive releases in the generative AI space. Google released its anticipated version of its ChatGPT competitor, Bard. Microsoft unveiled the next version of GitHub Copilot. NVIDIA has a mount of generative AI announcements at its GTC conference and Adobe unveiled a new generative AI platform called Firefly. Just another week in the generative AI space.

🔎 ML Research

(this week we added more details about the papers)

BLIP-2

Salesforce Research published a paper and open sourced a version of BLIP-2, a visual-language model with a similar architecture to GPT-4. The model is able to deliver similar results to GPT-4 despite being significatively smaller in size —> Read more.

GNNs for Production Recommendation

Amazon Science published a paper describing a graph neural network technique used for product recommendations. The model uses embedding representations that model product recommendations source and target respectively —> Read more.

ART

Researchers from the University of Washington, Microsoft, Meta AI,  Allen Institute of AI and Univervisty of California, Irvine published a paper detailing ART, a tool that uses frozen LLMs to generate intermediate reasoning steps. ART uses a few-shot technique to decompose a task into multistep micro tasks that simulate reasoning —> Read more.

DeBERTa-V3

Microsoft Research published a paper and open source version of DeBERTaV3, a LLM with improve attention mechanisms and the ability to process large text forms. The new model improves on its previous versions by using a more efficient pretraining technique known as replaced token detection (RTD)  —> Read more.

Vid2Seq

Google Research published a research paper describing Vid2Seq, a visual-language model for describing videos. The core idea of Vid2Seq is to augment the token representation in a language model with special time tokens optimized for predicting event boundaries and textual descriptions in the same sequence —> Read more.

🤖 Cool AI Tech Releases

ChatGPT Plugins

OpenAI released ChatGPT plugin, an extensibility framework that allows ChatGPT access external information sources —> Read more.

Bard

Google announced a limited beta of Bard, its ChatGPT competitor —> Read more.

Copilot X

GitHub unveiled CoPilot X, the new versions of its generative AI powered assistant based on OpenAI’s Codex model —> Read more.

GPT-4 in Azure

Microsoft unveiled the anticipated release of GPT-4 in the Azure cloud —> Read more.

NVIDIA Generative AI Releases

NVIDIA announced a series of software and hardware releases designed to empower generative AI applications —> Read more.

Stable Diffusion Reimagine

Stability AI released Stable Diffusion Reimagine, a tool that allow users to create multiple variations of an image without preset limits —> Read more.

LangFlow

This very cool open source project enables a user interface for the increasingly popular LangChain LLM framework —> Read more.

🛠 Real World ML 

ML Behind LinkedIn’s Skill Graph

LinkedIn discusses the ML architecture used to maintain a skill taxonomy used at its famous Skill Graph —> Read more.

Embeddings at Lyft

Lyft discusses their usage of graph neural networks to generate embedding representations —> Read more.

📡AI Radar

• Google enabled early access to Bard, its ChatGPT alternative powered by the LaMDA models.

• Mozilla launched Mozilla.ai, a new startup focused on building trustworthy open source AI.

• Adobe released Firefly, a new creative generative AI platform.  

• Opera added a new set of features powered by ChatGPT.

• Canva released a new suite of tools powered by generative AI models.  

• Numbers Station raises $17.5 million to enable a generative AI powered data stack.

• AI-powered patent intelligence search platform IPRally announced a $10 million series A.

• Aspecta raised $3.5 million to build its AI-based identity platform for developers.

• Bionic Health raised $3 million to launched a AI-powered health clinic.

• Game studio Ubisoft unveiled Ghostwriter, an AI tool for creating dialogues for non-playing characters in games à Read more.

• AI-based software testing company Codium AI emerged out of stealth model with a an $11m investment.   

• Workera, a platform that uses AI for improving workers skills, raised $23.5 million series B.