Next Week in The Sequence:

• Edge 323: Our series about foundation models explores the different types of memory in LLMs. We discuss the research that outlines how foundation models are computationally universal and the Chroma vector store.

• Edge 324: We deep dive into Code Llama, Meta AI’ new code generation language model.

📝 Editorial: The Other OpenAI Competitor that Just Raised a Lot of Money

When we consider standalone OpenAI competitors, our thoughts turn toward companies such as Anthropic or Cohere, which have been securing remarkable venture funding. However, the realm of OpenAI competitors extends beyond these. Companies like MosaicML were recently acquired for $1.3 billion to enhance DataBricks’ AI offerings. All of these companies share a common DNA, as they are constructing foundational models that serve as alternatives to the OpenAI suite.

Among the lesser-known OpenAI competitors, one has been operating impeccably while remaining surprisingly under the radar. AI21 Labs was founded by AI luminaries Yoav Shoham (Stanford University), Amnon Shashua (Mobileye), and Ori Goshen (CrowdX) with the goal of advancing research in foundational models. The company’s standout product has been the Jurassic models, which enable multi-task LLM capabilities across multiple languages. Jurassic-2 consistently ranks extremely high in state-of-the-art LLM benchmarks, surpassing much larger models.

In addition to its core foundational models, AI21 Labs offers hyper-optimized task-specific models for various domains, such as content management or expert knowledge systems.

Just last week, AI21 Labs announced the completion of a $155 million financing round that valued the company at $1.4 billion. Notable names participated in this round, but one, in particular, stood out: Jensen Huang, founder and CEO of NVIDIA. Although this amount falls short of the funding rounds of OpenAI, Anthropic, or Cohere, it positions AI21 Labs to be quite competitive in the short to medium term. AI21 Labs is undeniably a name to monitor, as they appear to be a formidable presence in the LLM space.

🔎 ML Research

RO-ViT

Google Research published a paper introducing a technique for training vision transformers in a region-aware manner. The idea is to augment the pretraining dataset with position embeddings that improve the spatial recognition capabilities of the model —> Read more.

Algorithm of Thoughts

AI researchers from Microsoft and Virginia Tech University published a paper unveiling Algorith of Thoughts, a new reasoning technique in LLMs. The method uses a new form of in-context learning in which the LLM works through different algorithmic reasoning paths —> Read more.

MoZO

AI researchers from Princeton University published a paper detailing memory-efficientzerothorderoptimizer(MeZO), a high performance fine tuning technique for LLMs. MeZO is able to fine-tune LLMs using the same memory footprint as inference by leveraging a classical ZO-SGD method —> Read more.

TIFA

Researchers from the Allen Institute for AI and the University of Washington published a paper detailing TIFA, a new metric that improves over CLIP by evaluating the faithfullness of generated images —> Read more.

🤖 Cool AI Tech Releases

DSPy

AI researchers from Stanford University open sourced DSPy, an framework for advanced LLM tasks including information retrieval and tool augmentation —> Read more.

FACET

Meta AI open sourced FACET, a new bechmark to evaluate bias and fairness in computer vision models —> Read more.

SynthID

DeepMind released SynthID, a tool for watermarking and identifying AI generated images —> Read more.

Bedrock Foundation Models

Amazon announced support for a new wave of foundation models in its Bedrock platform —> Read more.

🛠 Real World ML

Jupyter Notebooks at Meta

The Meta engineering team discusses the architecture used to schedule Jupyter notebooks at scale —> Read more.

📡AI Radar

• Google announced a new set of generative AI services at its Cloud Next conference.

• OpenAI competitor AI21 Labs raised $155 million at a $1.4 billion valuation.

• OpenAI announced the enterprise edition of ChatGPT.

• DoorDash announced an AI powered voice ordering solution.

• OpenAI unveiled some guidance for educators using ChatGPT in their classrooms.

• Language learning app Speak raised $16 million.

• AI search and analytics platform Rockset raised $44 million.

• Startup studio Betaworks announce a new batch of AI companies for its thematic camp.

• Personalized game startup Martian Lawyers Club raised $2.2 million to expand its platform with generative AI.

• Context.ai raised $3.5 million for its analytics suite for LLM apps.