Can Google’s investment revive AI21 Labs?

Ori Goshen, co-founder and co-CEO of AI21 Labs / Photo credit: AI21 Labs

Until ChatGPT became mainstream, AI21 Labs’ Jurassic-1 model was considered on par -or even better – than OpenAI’s GPT-3 model. The former had a staggering 3 billion more parameters compared to the latter.

However, with GPT-4 making waves today, the Jurassic-2 – the second-generation large language model (LLM) by AI21 Labs – seems to have lost its sheen.

Jurassic-2 claims to be the world’s most customizable language model. And with five APIs, it gives organizations a platform to build their own applications and services to scale.

Amazon currently includes AI21 Labs’ foundation model in its list of options under Amazon Bedrock.

In its effort to compete against Microsoft’s increasing market share in the sector, Amazon Bedrock launched a fully managed service that offers foundation models from Amazon and leading AI startups through an API.

However, the Jurassic-2 model struggles to provide answers to simple queries, unlike GPT-4 which handles them with ease. For instance, when I asked a mathematical question, the Jurassic-2 model often goes way off track.

The J2-ultra model from AI21 Labs struggles with mathematical questions.

In this instance, the correct answer is 84. But not only does Jurassic-2 get it wrong, it also doesn’t take into account that 150 years is too old for a human. However, GPT-4 gives the correct answer.

While the model does explain what the code does and how it works, it fails to debug it

In another example, I gave Jurassic-2 a python code with a bug, but the model wasn’t able to debug it, something which GPT-4 does easily. I even prompted Jurassic-2 with a code for calculating with zero value in “list_02,” the model overlooks it and provides an answer which isn’t helpful. 

Recently, AI21 Labs raised US$155 million in a series C round, which valued the startup at US$1.4 billion.