AI21 Labs claims it’s building the next phase of natural language processing

AI21 Labs, a Tel Aviv, Israel-based AI lab and product company, says it’s developed a new system called the modular reasoning knowledge and language system, or MRKL system (pronounced “miracle”), to augment the power of large language models, enhancing them with discrete reasoning experts like online calculators and currency converters. With its first implementation, Jurassic-X, the MRKL system, includes different language models that it augments with external knowledge sources such as weather apps and Wikidata.

AI21 Labs is both a research lab and a commercial business that offers natural language processing (NLP) as a service. Advised by Udacity founder Sebastian Thrun, AI21 Labs was cofounded in 2017 by Crowdx founder Ori Goshen, Stanford University professor Yoav Shoham and Mobileye CEO Amnon Shashua. The company says its mission is to shift and shape the way people read and write by making AI a thought partner to humans. AI21 Labs has developed AI language applications like Wordtune and Wordtune Read, with the capacity to generate natural language and understand meaning.

Last year, the company released Jurassic-1, a 178B-parameter autoregressive language model, attracting over 10,000 developers. While other language models like Jurassic-1, GPT-3, and others unfold numerous opportunities, they are also limited — lacking the ability to reason, perform simple arithmetic and connect to live data sources. With a MRKL system like Jurassic-X, users benefit from the existing mega language models without experiencing any of the disadvantages.

Collaborating with AI21 Labs as design partners for Jurassic-X is Publicis Groupe, an advertising and marketing agency. According to Shoham, both companies will work together to create a fantastic user experience for the new architecture. Publicis Groupe will provide creative direction for AI21 Labs’ design team, which includes several highly qualified designers. This collaboration is part of an ongoing effort by both companies to make it easier for marketers and brands alike to harness the power of artificial intelligence (AI). By partnering, they hope to provide users with applications that are better personalized. 

Augmenting language models

The first implementation of the MRKL framework was unveiled as Jurassic-X and is available on AI21 Studio — a developer platform offering immediate access to Jurassic-1. The MRKL system includes different language models augmenting with external knowledge sources and symbolic reasoning experts. Shoham said this new system is the next evolution in NLP that will revolutionize how AI processes the way humans speak.

“It is impossible for pre-trained language models to keep up with dynamic information such as fluctuating exchange rates, weather forecasts, and constantly updating sources like Wikidata. In this way, current language models fall short of their true potential: a general-purpose natural language interface, which receives a wide variety of inputs in natural language that are relevant for real-time purposes, and which replies similarly, in natural language,” said Shoham.

To augment language models and enhance them with broad skills, Jurassic-X was developed to address the weaknesses in mega models like Open AI’s GPT-3 and Jurassic-1. The company claims Jurassic-X enables companies to plug in their database and extract information using free language. Jurassic-X also helps in text generation by combining a powerful language model with access to Wikidata, according to Shoham.

He said Jurassic-X can answer non-trivial math operations phrased in natural language with the combination of a language model and a calculator, adding that it also simplifies complex questions, answers sub-questions, and gathers the answer.

What to expect

As the enterprise continues to move toward a more digitally reliant economy, NLP is becoming increasingly crucial. Shoham said AI is great, but the world is still far from capturing real intelligence. AI21 Labs says it aims to get a little bit closer in the quest to build more accurate AI models.