AI21 Offers Instant Access To Its NLP Models, Jurassic-1 Large and Jurassic-1 Jumbo via AI21 Studio


AI21, An Israeli AI company specializing in Natural Language Processing (NLP), has recently launched two big NLP models, Jurassic-1 Large and Jurassic-1 Jumbo, through an interactive web UI dubbed AI21 Studio. Unlike OpenAI, which has a limited beta, AI21 makes its models available for everyone to try out – there is no waiting list.

As per the researcher, “Using AI21 Studio, businesses can take advantage of text-based AI in the same way that Amazon Web Services makes cloud computing available”

NLP is an area of computer science that aims to design algorithms that can process and generate written natural language. Language Models are systems that can intake text and generate likely continuations. Language models are almost exclusively produced at big tech AI labs, out of reach of a wider developer audience, because training and deploying them into production generally requires expensive computational resources and highly sought-after AI engineers. As a result, language models are almost exclusively produced at big tech AI labs, out of reach of a wider developer audience.

Researchers also said that anyone can now create language-based applications that rival those developed in big AI labs, not just AI researchers at big tech companies. Now publishers, students, artists, business people, and researchers can build applications competing with those in AI labs.

Developers who join the open beta get immediate access to Jurassic-1 Jumbo, which is slightly larger than OpenAI’s GPT-3 with 178 billion computational parameters. AI21 Studio™ also enables developers to quickly customize a private version of Jurassic-1 models, reducing production time and costs.

By limiting the quantity of text generated in the open beta, AI21 Labs hopes to prevent potential misuse of its language models. “We take misuse very seriously, and we’ve taken steps to restrict the potential consequences that have affected others,” said a team member. Developers will require access to bespoke models that aren’t constrained by the open beta to construct apps that will service production-scale traffic. Once clearance is given, developers can raise their usage in a pay-as-you-go cloud services model — AI21 Labs manually reviews every custom model application. AI21 Labs is in charge of the bespoke models’ deployment, serving, and scaling.

Here’s how you can try Jurassic-1

Unlike GPT-3, Jurassic-1 is open to the public. It’s available through the AI21 studio (account creation does require phone-number authentication).

For the Jurassic-1 178B model, the free tier allows 10K tokens each day, three times as much for the smaller Jurassic-1 7.5B model. That’s plenty to try out the web UI, but not enough to perform any tests or benchmarks using the API.

AI21 Studio, which is now in “limited open beta,” will be used to commercialize the company’s models. The company has yet to reveal a price model for commercial use.

Jurassic-1 in action, generating blog post ideas in AI21 Studio Source:

Performance metrics and model sizes

There are larger models available, such as the Chinese Wu Dao 2.0, which is 10 times the size and has 1.75 trillion parameters. However, AI21’s J-1 Jumbo is the most comprehensive English language model currently available to the general public.

J-1 Jumbo’s zero-shot model performance on known benchmarks is comparable to GPT-3 Davinci, OpenAI’s largest GPT-3 model. When the model is “zero-shot,” it is not given any special instructions and is not fine-tuned on any task-specific training data.


Summarizing or simplifying text

Concerns about AI safety, ethics, and biases have plagued neural language models in the past, and they continue to do so with AI21’s models. Putting those concerns aside for a time, AI21’s models appear to be a viable alternative to GPT-3. They do, however, lag on a few fronts:

  1. They cannot employ specialist models such as “GPT-3 Davinci-instruct,” which encourages GPT-3 to follow prompts, or “GPT-3 codex,” specializing in producing code.
  2. The “prompt” ecology is still in its infancy in comparison to GPT-3. Many of the prompts in GPT-3 do not adequately convert to AI21, and there is no comprehensive “official” list of prompts available yet.
  3. The AI21 free token quota is excessively low, and no usage-based pricing has been disclosed yet. It’s tough to perform benchmarks or undertake quick engineering because of this. Still, if you write to them to explain the demand, they will gladly increase the quota.

AI21, on the other hand, is still in its infancy. The AI21 language models are expected to become a viable alternative to the OpenAI language models over time.



Get Access:

AI21 Blog:

Use Cases:

VB Source: