Startup's latest large language model rivals GPT-3
Jurassic-2 is the latest model from AI21 Labs. The model is the successor to Jurassic-1, which was trained on 178 billion parameters, 3 billion parameters more than GPT-3.
Israeli startup AI21 Labs has introduced a new version of its large language model and a set of new APIs to rival other LLMs on the market, most notably OpenAI's GPT and Microsoft's version of ChatGPT in its Bing search engine.
AI21, based in Tel Aviv, launched Jurassic-2 Wednesday along with five new APIs, all generally available now.
Like the previous version, Jurassic-2 is available in AI21 Studio, a natural language processing-as-a-service developer platform that enables developers to create text-based applications such as virtual assistants and chatbots.
The LLM comes in three sizes: large, grande and jumbo. Each size includes a different instruction-based version of the LLM.
New features in Jurassic-2 include support for more languages and faster response for jumbo models compared with Jurassic-1.
Currently, Jurassic LLMs are applied to AI21 Labs' AI editor application Wordtune. However, unlike ChatGPT, the application provides suggestions for writers as they write instead of giving answers in a Q&A format.
Compared with GPT-3, whose knowledge base as of now is current until only the end of 2021, Jurassic-2's training data set is updated up to mid-2022, meaning its data is currently more accurate than GPT-3, AI21 claims.
As the generative AI arms race heats up, it is only a matter of time before other LLMs release an updated version to match this.
Andy ThuraiAnalyst, Constellation Research
However, Jurassic-2 won't be the only updated model for long, according to Andy Thurai, an analyst at Constellation Research.
"As the generative AI arms race heats up, it is only a matter of time before other LLMs release an updated version to match this," Thurai said.
AI21's pricing and size of the large language model -- based on the Starbucks coffee chain's pricing convention -- provides more flexibility than some other LLMs and a usage model depending on the need of users. By comparison, other LLMs, such as GPT, have one size fits all jumbo models and so have fixed latency (how fast the chatbot responds to a query) and per-usage cost, Thurai noted.
Possible drawbacks
Despite its apparent differentiators from other LLMs, Jurassic-2 is considerably less known, Thurai continued.
"ChatGPT is so popular, so the traction and commercialization aspect could be much faster than AI21 Labs," Thurai said.
Moreover, AI21 may worry about being driven out of the market by tech giants like Microsoft and its partner, OpenAI.
The bigger vendors "may be able to train the models more frequently going forward if needed without worrying about the costs, and might reduce the usage costs to drive the competition out," Thurai said.
AI21's recent fundraising success shows investors are interested in the vendor in much the same way Microsoft is interested in OpenAI as an R&D engine.
However, the preference for which LLMs developers use might depend on their application, Gartner analyst Bern Elliot said. Moreover, as language models improve, large models may not be needed as much.
"Different models will become more specialized so that you can get models that are more specific for your purpose and don't have to work with such massive models if you have a narrow purpose," Elliot said.
Along with Jurassic-2, AI21 released five new APIs that enterprises can use to integrate generative AI features into their websites. The APIs provide paraphrasing, summarization, text recommendation, text segmentation and grammatical error correction.
