AI21 Labs - Should I add it to my AI Stack?

Introduction to AI 21

Stuart introduces AI 21, a large language model provider and competitor to OpenAI. He discusses the new pricing structure and upgrades made to the models.

Pricing Structure

• AI 21 has a minimum billing of $29 per month, but offers a 90 credit trial for three months without requiring a card on file.

• The foundation models are Babbage, Curie, DaVinci style models that can be built upon. The jumbo model is the most expensive at $0.015 per thousand tokens, followed by the grand model at $0.01 per thousand tokens and the large model at $0.003 per thousand tokens.

• Once you fine-tune a model in AI 21, it costs the same as the base model you fine-tuned it on.

Upgrades Made to Models

• All of AI 21's models have been upgraded recently.

• The tokenizer in AI 21 is more lenient than OpenAI's tokenizer, resulting in more output for your buck.

• Task-specific APIs are available for paraphrasing, summarizing, grammatical error corrections and text segmentation.

Differences Between J1 and J2 Models

• Stuart discusses differences between J1 and J2 versions of models in the playground.

Token Limits and Fine Tuning

This section discusses the token limits of various models, including DaVinci and J2. It also covers the benefits of fine tuning with AO21 over OpenAI.

Token Limits

• Most models have a token limit of 2048.

• J1 has an 8K token limit, while J2 has a limit of 8191 tokens.

• The memory upgrade in J2 allows for more context and better outputs.

Fine Tuning

• Fine tuning with OpenAI is limited to base models with a 2048 token limit.

• AO21 allows for larger text within each example, making it ideal for fine tuning with J2 models.

• AO21 terms are less restrictive than OpenAI's, allowing for more liberal use cases.

Building Data Sets and Fine Tuning Models

This section covers how to build data sets using Riku's data set studio and CSV to JSON converter. It also explains how to fine tune models using AO21 or OpenAI.

Building Data Sets

• Use Riku's data set studio to build out your data sets line by line.

• Use the CSV to JSON converter if you have all your data in a CSV file.

Fine Tuning Models

• Upload your JSON file and choose whether you want to use AO21 or OpenAI for fine tuning.

• Fine tuning with AO21 is ideal for J2 models due to its larger text allowance per example.

Update on AI 21

The speaker recommends exploring AI 21's updated models and suggests reloading or registering a new account to try them out. The models are free to use until May 1st.

Upgraded Models

• AI 21 has upgraded from Jurassic 1 to Jurassic 2 versions of their models.

• The token length has increased significantly.

• All of the models are currently free to use until May 1st.

Experimentation and Use Cases

• It is recommended to experiment with different large language models to find the best fit for your needs.

• Some use cases may be disallowed in the terms and conditions of different large language model providers, so it's important to choose one that fits your specific use case.

• AI 21 is a good alternative for those who have looked into OpenAI but cannot meet their terms or are looking for something different.

Conclusion

• The J2 models are now live on Riku, where users can experiment with all the best large language models in one place without any code.

• Those interested in learning more about AI can join the speaker's Facebook group.