AI21 Labs: Being in Amazon's Bedrock without any investment from Amazon

AI21 Labs was the first independent startup to surpass OpenAI’s GPT-3. In September 2021, their Jurassic-1, featuring 178 billion parameters, ignited the 2021 race for ever-larger language models. In the same year, Amazon Web Services (AWS) published a case study highlighting how AI21 Labs optimized their language model using Amazon's EC2 P4d Instances and the PyTorch toolkit.

Strangely enough, AI21 Labs, despite its integration into the AWS ecosystem, hasn’t received investment from Amazon, which contrasts with Amazon's investments in Cohere and Anthropic. This is intriguing given AI21 Labs’ solutions are employed by Amazon's Bedrock.

Is there a generous investment from Amazon coming? Let’s learn more about the startup, which recently joined the GenAI Unicorn Family with a $1.4 billion valuation and has, so far, thrived without Amazon's financial backing.

AI21 Labs was co-founded in 2017 by Prof. Yoav Shoham (Professor Emeritus at Stanford), Ori Goshen (Founder, CrowdX), and Prof. Amnon Shashua (Founder, Mobileye). "AI21 Labs was founded to change and fundamentally improve how people read and write. Pushing the frontier of language-based AI requires more than just pattern recognition offered by current deep language models," CEO Shoham told VentureBeat. 

Shoham met Goshen in an Israeli non-profit project that democratizes programming skills. They found each other like-minded and decided to start AI21 Labs ("AI for the 21st century") to build AI systems "with an unprecedented capacity to understand and generate natural language."

Amnon Shashua became the Chairman, while Ori Goshen and Yoav Shoham were appointed as Co-CEOs.

After being in stealth mode for about three years, in October 2020 the team launched its first product, Wordtune, an AI-powered writing tool that helps refine and enhance text for clarity, tone, and style in real-time.

As per Goshen, what inspired them to develop a writing productivity tool as a use case is that "The way we write hasn't drastically changed since Microsoft Word in 1980, and the way we read hasn't changed since the printing press in 1440. We believe both can be made AI-first experiences, with the machine becoming a true writing and reading partner."

The lab was founded to become the next big disruptor in the Natural Language Processing (NLP) space to give machines the human-like ability to understand and generate text. Far from seeing machines as a replacement for humans' skills with written symbols that represent our understanding of the world, AI21 Labs views technology as a revolutionary tool.

Funnily enough, in ancient times (back in June 2022), when OpenAI had not yet announced building a superintelligence, they collaborated with Cohere and AI21 Labs to develop a set of Best Practices for Deploying Language Models. The guidelines were phrased in very grounded language: 'The future of human-machine interaction is full of possibility and promise, but any powerful technology needs careful deployment.’ For example, in the ‘Thoughtfully collaborate with stakeholders’ section, the document suggests building teams with diverse backgrounds, Publicly disclosing lessons learned regarding LLM safety and misuse, and treating all labor in the language model supply chain with respect.

Very good points, right?

Recently, Ori Goshen talked about how people were hyped and scared about generative AI. But he's saying, "There has been a lot of hype – and fear – surrounding generative AI recently, with new exciting tools emerging daily and consumer demand through the roof. The initial fear around AI is starting to wane as the world realizes the benefits. McKinsey recently released a report estimating that AI will boost the world economy by trillions each year."

The notable rounds:

As we’ve mentioned, they first launched Wordtune (2020), a very practical writing assistant for better communication. But that wasn’t exactly what the lab part was interested in. In 2019 they published a paper SenseBERT: Driving Some Sense into BERT – about teaching algorithms to better comprehend human language. In 2021 AI21 Labs was ready to share the real deal, so they dropped two heavyweight NLP models – Jurassic-1 Large and Jurassic-1 Jumbo. At that time, Jumbo was the largest English language model available to the general public so far, with 178 billion parameters.

In June 2022, the AI21 lab spiced things up with an 'Ask Ruth Bader Ginsburg' bot. Ruth Bader Ginsburg served as an associate justice of the Supreme Court of the United States from 1993 until she died in 2020. The bot had been fed a lot of her data: legal opinions, interviews, speeches. But a former law clerk of Ginsburg, Paul Schiff Berman, wasn't too impressed. He told The Washington Post the bot's responses didn't quite cut it.

Today, the company caters to enterprise-level customers. Recently, the team announced Contextual Answers, a plug-and-play engine for businesses to make finding info quicker and more accurate using generative AI. You ask a question about your company's info, and the tool gives you the answers. 

AI21 Labs has been focusing on addressing the challenges of text ambiguity in NLP. Here's a summary of their approach and innovations:

Based on that research and approach, they first developed a language model called HAIM. HAIM means “Halfway Acceptable Interpolating Machine”. The model offered such a name itself.

The demo of HAIM is unavailable now, and the company efforts are concentrated around their most famous Jurassic language models family.

First, it was: J1-Jumbo (178B parameters) and J1-Large (7B parameters). The "Jurassic-1: Technical Details and Evaluation" paper explains that these models are built on the Transformer architecture, with enhancements for efficiency and a larger vocabulary size compared to GPT-3. The paper evaluates Jurassic-1's performance in terms of perplexity, zero-shot, and few-shot learning. It also addresses the challenges of training large-scale models and the inherent biases in language models, emphasizing the need for ongoing research in these areas.

But Jurassic 1 is already outdated as well. In March 2023, Jurassic-2 was introduced along with their Task-Specific APIs, including the Wordtune API set.

Jurassic-2 represents a significant upgrade over its predecessor, Jurassic-1, featuring improvements in quality, zero-shot instruction-following, reduced latency, and multilingual support. The J2 model family includes Large, Grande, and Jumbo sizes, with the Jumbo model ranking high on Stanford’s Holistic Evaluation of Language Models (HELM).

A significant advancement by AI21 Labs is their MRKL system – a modular, neuro-symbolic architecture that combines large language models, external knowledge sources, and discrete reasoning.

Wordtune is based on Jurassic and other proprietary algorithms as well.

How does Wordtune work: It suggests alternative ways to write what one has typed. The interface is similar to translation productivity tools (a.k.a. CAT tools). It's like this: hover your cursor over a word, and a menu of synonyms pops up. Highlight a sentence, and you're given a selection of paraphrased options, categorized into casual (marked by a sneaker) or formal (with a briefcase).

Earlier this year, AI21 Labs conducted a social experiment called "Human or Not". More than 2 million people chatted with each other over 15 million times on their website. Here's the twist – at the end of each chat, they had to guess if the other person was a human or a computer. Nearly a third of them got it wrong.

The creative product lead at AI21 Labs at the time of the experiment, Amos Meron, said they did this experiment to show people, researchers, and policymakers what AI was up to in early 2023. It wasn't just about AI being a handy tool; they wanted everyone to see AI as new members in the online world. You know, at a time when everyone's trying to figure out how AI should fit into our lives. 

The experiment was a sneak peek into the future, and AI21 Labs wanted everyone to get a good look.

Also worth checking the paper ‘Human or Not? A Gamified Approach to the Turing Test’.

AI21 Labs generates revenue through its various AI products and services, with a significant portion of its income coming from Wordtune. The company's revenue was reported to be $20 million in 2022, with a projected revenue of $50 million for the following year, indicating substantial growth in its income.

The pricing structure for these services typically revolves around API usage. The AI21 Studio is a pay-as-you-go developer platform.

They also access to their foundation models:

And access to Task-Specific Models – specialized APIs tailored for reading and writing.

AI21 Labs is often referred to in the media as an OpenAI competitor. However, their story reveals a very different approach. Unlike OpenAI, which has established a closely intertwined relationship with its main investor and key client Microsoft, AI21 Labs opts for a more open-ended strategy. Their collaboration with Amazon, a major client and partner, exemplifies this approach. Despite their integration into Amazon's systems, AI21 Labs operates independently, without Amazon's financial backing. Their pragmatic and grounded approach focuses on leveraging AI as a powerful tool for enterprises, rather than pursuing the elusive goal of Artificial General Intelligence (AGI). It wouldn't be surprising if Amazon were to invest in the company soon, but regardless, AI21 Labs has already positioned itself as an independent entity, emphasizing AI as a tool rather than an existential threat.

In a world with long literature reviews and complex jargon among many scientific hindrances, SciSpace is an AI platform that empowers academicians to conduct their research.

With SciSpace, you can find 282 million+ relevant papers for any topic or question, and analyze and contrast multiple papers on one screen. Further, you can ask an AI Copilot for summaries, explanations, or any question to read papers faster.

Get unlimited access to SciSpace at 40% off. Use coupon code "TP40"

Prof. Yoav Shoham 

Ori Goshen

Prof. Amnon Shashua

Thank you for reading, please feel free to share with your friends and colleagues. In the next couple of weeks, we are announcing our referral program 🤍