Next Week in The Sequence:

• Edge 317: Our series about generative AI techniques starts exploring in-context learning in LLMs. We review two papers from Stanford Univervisy that explore the emergence of ICL and the GPTCache framework to enable caching in foundation models.

• Edge 318: We deep dive into Microsoft’s LLM-AUGMENTER, a new architecture to augment LLMs with memory, knowledge, and external feedback.

Time to Subscribe:

📝 Editorial: Keeping Up with NVIDIA’s Generative AI Announcements

NVIDIA has made generative AI the center of its value proposition. After all, it was generative AI that pushed NVIDIA briefly to a trillion dollar market cap a few months ago. NVIDIA’s ambitious in generative AI expand way beyond its dominant position in GPUs and onto other hardware and enterprise software domains.

This week at the SIGGRAPH conference, NVIDIA unveiled so many generative AI initiatives that it’s pretty hard to keep up with it. The announcements encompass both new hardware and enterprise software initiatives targeted to streamline the adoption of generative AI. Let’s briefly summarize NVIDIA’s generative AI week:

• GH200: A new version of the version of the Grace Hopper Superchip was unveiled with combines a 72-core Grace CPU with a Hopper GPU and a strong memory architectures.

• AI Workbench: A new toollset to quickly build and fine-tune generative AI models in commody hardware and then scale with NVIDIA DGX Cloud.

• AI Enterprise 4.0: A new release of NVIDIA generative AI software suite for enterprise suite optimized for large scale deployments.

• Hugging Face Partnership: This strategic alliance might extend supercomputing access to million of data scientists in the Hugging Face community.

• Omniverse Gen AI Capabilities: A new version of NVIDIA’s Omniverse platform unveiled new generative AI features and strong support for the Universal Scene Description(OpenUSD) framework.

• OpenUSD: Speaking of OpenUSD, NVIDIA announced a new set of frameworks such as ChatUSD or RunUSD for implementing OpenUSD applications and pipelines.

• New Workstations: NVIDIA unveiled a new set of RTX workstations with embedded support for generative AI applications.

There were additional announcements related to generative AI hardward, software and research. NVIDIA is certainly a force to be reckon with in the generative AI space and its ambitious are not circunvented to the hardware space. Exciting week!"

🔎 ML Research

Adaptive Computation in Foundation Models

Google Research published a paper detailing AdaTape, an adaptive computaiton method for LLMs. By adaptive computation, we refer to a technique that uses a dynamic number of tokens to adapt to changes in the environment —> Read more.

Compressing Token Embeddings

Amazong Science published a paper presenting a compression method for token embedding matrices used in LLMs. The technique combines methods such as low-rank approximation, a novel residual binary autoencoder, and a fresh compression loss function to improve the size of token embedding structures —> Read more.

Advances in Document Understanding

Google Research published a paper outlinign the Visually Rich Document Understanding (VRDU) a new bechmark for document understanding models. The behckmark is based on five key requirements of document understanding tasks —> Read more.

FABRIC

Researchers from ETH Zurich published a paper introducing FeedbackviaAttention-BasedReferenceImageConditioning(FABRIC), a method for text guidancen in Diffusion models. FABRIC uses a combination of positive and negative feedback to adapt results via image-conditioning —> Read more.

🤖 Cool AI Tech Releases

NVIDIA Gen AI Releases

At the SIGGRAPH conference, NVIDIA annuonced a series of major generative AI releases —> Read more.

Claude Instant 1.2

Anthropic released a new version of its lightweight Claude Instant model —> Read more.

StableCode

Staiblity AI released StableCode, a new open source LLM for coding —> Read more.

Project IDX

Google announced a preview of Project IDX, a browser based development environment based on PaLM 2’s Codey model —> Read more.

Gen AI in Jupyter

The Jupyter team released a series of extensions to incorporate generative AI capabilities into Jupyter notebooks —> Read more.

🛠 Real World ML

Inside Instagram Recommendations

Meta discusses the ML architecture powering Instagram’s recommendation system —> Read more.

Generative Image at Salesforce

Salesforce Research details the architecture and training process of their XGen-Image-1 model —> Read more.

📡AI Radar

• Weights and Biases announced a $50 million funding round led by AI-industry luminaries.

• LLM platform deepset, creators of the popular HayStack framework, announced a $30 million funding round.

• Federated learning platform Flower landed $3.6 million in new funding.

• AI-HR platform One Model raised $41 million in new funding.

• Multiplayer, an AI-based distributed software development platform, came out of stealth mode with a $3 million funding raise.

• Tractian, an AI platform for mechanical operations, announced a $45 million series B.

• Salesforce announced the general availability of its Einstein Studio including the ability of bringing your own model.

• AI influencers Chris Messina and Brian McCullough announced a $15 million AI VC fund.

• Defense Advanced Research Projects Agency (DARPA) launched a new challenge to develop AI cyber software.

• Middleware raised $6.5 million to improve cloud observability using AI.

• 3D data exploration platform Virtualitics announced a $37 million series C.

• Rasgo launched a new generative AI agent for data warehouse technologies.

• 11xAI announced a $2 million pre-seed to create autonomous AI productivity agents.