Cohere

This Toronto company has come out of stealth with a wealth of funding and a founding team with some serious clout. The team have a close relationship with Google and some of the founders worked on some of the most prominent AI/ML projects within Google prior to starting this company.

Cohere offer a bunch of tools to help with your AI needs and don’t just focus on generating text. You can use their models for sentiment analysis and figuring out patterns. When talking with the team, you get a sense of their deep understanding and passion for this space so it is definitely worth keeping a track of what they are up to.

Cohere offer 3 models for text generation, a small, medium and large. They don’t disclose the size of the models in terms of parameters but they seem well trained and the API is easy to use. The pricing is also very generous compared to others on the market.

One of the areas of Cohere which I think is above anyone else currently is the way that you are able to fine-tune the models and create these yourself in a fully self-serve scenario. This process is simple, fast and get you the API endpoint immediately to implement into production.

As far as prompts go with Cohere, I have found that you need to phrase things slightly differently. Whereas a prompt used in an OpenAI model can be pasted into an AI21 model and work just as well, there is a definite art to using cohere which can be a little frustrating and might put some people off.

Cohere take an approach similar to OpenAI with ethics and morals of the technology wanting all users to go through a launch review prior to putting the technology into production. There are some banned use cases such as not creating full blog posts etc which you should be aware of. Also through testing, there is a filter on the models which is built into the API endpoints. This is a better system than the one from OpenAI but a filter is still frustrating for certain things such as my testing with a female ecommerce fashion brand which kept hitting it for skirts and underwear descriptions.

The price of Cohere really is a strong reason to include it in your AI stack of tools. The team are approachable and willing to help you out and I believe that they will constantly improve the models. The fine-tuning definitely is where the product comes alive. If you have a dataset of over 1MB and want a fine-tune, you will not find an easier way to get the fine-tune done.