Contextual

Contextual is an AI start-up building foundation models for businesses with customizable AI to improve productivity. Contextual aims to provide enterprises with generative AI models to solve the following issues:

• Customizability
• Data privacy
• Hallucination
• Compliance
• Staleness or out-of-date training data
• Latency and cost

Rather than trying to develop a generalist model for AGI (artificial general intelligence), Contextual is developing enterprise AI models using Retrieval Augmented Generation (RAG). General models can require more parameters, increasing the money, latency, and compute required to run. RAG combines generative AI with retrieval-based data stores, such as a vector database, to adapt models for custom data and a specific use case. Contextual believes building LLMs (Large Language Models) that work with an external memory produce systems less prone to hallucination while being more customizable to new data. The business aims to build better performance AI for enterprise by decoupling the memory from the generative capacity of LLMs.

In June 2023, the company announced LENS (Large Language Models ENhanced to See), a modular approach to independent vision modules and LLMs that enables comprehensive multimodal understanding. The release included a paper, its code, and a demo. Existing multimodal models that reason about images require additional data and compute resources for extra training. This slows down the addition of visual capabilities. LENS allows LLMs to leverage these capabilities directly.

Examples of LENS' capabilities.

Contextual states that LENS is competitive with popular multimodal models, such as Flamingo and BLIP-2, without explicitly training the LLM to reason about images. Contextual released the LENS code and plans to support and be supported by the open-source community. This includes leveraging open-source software as much as possible and giving back to the community.

Contextual was founded in February 2023 by Douwe Kiela and Amanpreet Singh, who both have tenured backgrounds in artificial intelligence research. The two founders first met as research leaders at Meta AI in 2016 and continued to build their relationship, working together at Hugging Face. Kiela is also currently an adjunct professor at Stanford University. While at Meta AI, the pair built a multimodal framework for synthesizing information from text, images, and video to help deal with problems, such as detecting hate speech in memes, catching the sale of illicit goods, and fighting misinformation. They also pioneered the first effort around RAG, publishing the first papers and open-sourcing models on it in 2020. The two founders have been involved in a number of AI projects, including the following:

• Foundation models—RAG, FLAVA, UniT, Bloom
• Dense retrieval—Hallucination reduction, MDR (Multi-hop Dense Retrieval), Unsupervised QA
• Evaluation—SentEval, GLUE, SuperGLUE, Dynabench, Eval on the Hub
• Multimodality—Hateful Memes, Winoground, TextVQA, MMF
• Representation learning—InferSent, GroundSent, Poincare embeddings

On June 7, 2023, Contextual emerged from stealth mode with $20M in seed funding led by Bain Capital Ventures (BCV) and with participation from Lightspeed, Greycroft, SV Angel and angel investors including Elad Gil, Lip-Bu Tan, Sarah Guo, Amjad Masad, Harry Stebbings, Fraser Kelton, Sarah Niyogi and Nathan Benaich.