Israeli artificial intelligence company AI21 Labs has launched a new version of its popular large language model Jurassic and the focus has shifted to performance over size. The company says it gets better performance from a small footprint thanks to optimisation efforts. It says the smallest version of Jurassic-2 outperforms the largest version of Jurassic-1.
The Wordtune maker says it is working to create the best text engine possible, create a product that can be packaged and sold and have it work across multiple languages. It has added support for Spanish, French, German, Portuguese, Italian and Dutch in the latest update.
The Jurassic-2 foundation models are currently available to use for free until 1 May, as a way to encourage people to use the technology, after that point pricing is based on usage with a minimum subscription fee of $29 per month. This also allows for unlimited seats, custom model training and access to the task-specific APIs.
To emphasise the product-like nature of the model, the company is opening five distinct APIs and reducing the overall cost of querying the model. The new APIs include the ability to provide a full summary of a long-form article, offer up improvements to a given text as seen in Wordtune Spices and the ability to paraphrase a paragraph of text that matches the style of the wider work.
In addition to these writing tools, Jurassic-2 will include a grammatical error correction API and text segmentation which can split long pieces of text into appropriate segments based on topics. The company hasn’t given any information on the specific size of the LLM or how many parameters but claims it ranks second behind OpenAI’s Davinci models in the Stanford HELM (Holistic Evaluation of Language Models) assessment.
The aim of the HELM tool was to create a standard metric to evaluate models produced by a range of companies and has been used by Anthropic, Big Science, Google, Meta, Microsoft, OpenAI and others in addition to AI21 Labs. It measures each model against a range of criteria including efficiency, fairness and even whether the text contains toxic content.
AI21 Labs says its goal is to “fundamentally change the way people read and write using generative AI” and rather than have it produce an entire body of text, it works to improve or change text, describing Jurassic-2 as “the world’s most customizable language model”.
Focus on building a product
It builds on the work of Jurassic-1, a large language model that can in a range of versions from a few billion to 175 billion parameters. That version had been used in the production of gaming worlds, the creation of chatbots used by Harambee to increase sign-ups for youth employment programs and Verb which used it to help writers across a range of industries.
Jurassic-2 is more efficient than its predecessor, with Udi Karpas, Squad Director at AI21 Labs telling Tech Monitor the smallest version of Jurassic-2 outperforms the largest version of Jurassic-1. “Grande which is our medium-sized model is punching so much higher than its weight,” he said. “The only models that it’s trailing behind in HELM are three times to 30 times larger. If you care about quality only, we have jumbo.”
In the HELM rankings, the Grande model is ranked fifth. Places one through four are 30 times larger, which Karpas says makes it the perfect trade-off between efficiency and accuracy.
The new model was put through comprehensive instruction tuning using proprietary data, rather than more public open data sets more commonly used in this sort of approach. It also has up to 30% faster response times compared to the previous generation.
More up-to-date data
Most importantly its data is more up to date, a problem that plagues other models that rely on datasets up to two years old. Karpas said the model was trained on data as recently as the middle of last year and they are working on ways to close the training gap.
It is available as part of the AI21 Labs NLP-as-a-service developer platform in Large, Grande and Jump with a specific instruction-tuned version inside the platform and the ability to build a range of tools including chatbots like ChatGPT, virtual assistants and content moderation tools.
“When we released Jurassic-1, we were the first company to offer open access to LLMs without any waiting list, allowing anyone to build textual generative AI apps,” said Ori Goshen, Co-CEO and Co-Founder of AI21 Labs.
“The Jurassic-2 family of models represents the next leap forward and will enable developers and organisations to build text-based applications, faster, with state-of-the-art performance at a fraction of the cost. For those who don’t need all the capabilities for Jurassic-2 but still want to integrate generative AI into their websites, we also have new task-specific APIs to make our language models easily available to businesses and their needs.”