Cohere AI Chronicle — Challenging Tech Giants With AI For All

This is our fifth edition of Unicorn Chronicles (#1 OpenAI, #2 Anthropic, #3 Inflection, #4 Hugging Face), focusing on generative AI unicorns. Today, we’re exploring Cohere, which lost two positions in the ranking of AI unicorns after the sudden rise of Hugging Face and Inflection.

Now, Cohere is #5, but does it mean the company loses its position in the market? Let's dive deeper and discover more about the birth of Cohere, its mission, financial situation, main products, and other crucial details! Things are not as simple as it seems.

1. The starting point - Attention and a Dropout is All You Need

2. Mission of the company and its founders

3. Financial situation

4. What Cohere offers and their tech specs

5. How does Cohere make money?

6. Bonus: All important links about the founders

The starting point - Attention and a Dropout is All You Need

Co-founder of Cohere Aidan Gomez started his career as a Machine Learning Intern at Venture Media, where his work revolved around tackling real-time audio and sheet-music alignment using recurrent neural networks on mobile devices. Following this, Gomez made several career moves, first at Microsoft and then at Google Brain, where he crossed paths with Lukasz Kaiser. Together, they collaborated on the development of multi-modal networks, capable of addressing eight distinct problems spanning the realms of vision, audio, and language.

The camaraderie between Gomez and Kaiser endured even after Gomez transitioned to the University of Toronto in early 2017. This enduring partnership culminated in the publication of the renowned "Attention Is All You Need" paper, co-authored by Gomez, Kaiser, and six other researchers from Google Brain. In the comments accompanying the paper, it is noted that “Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.” Interestingly, Kaiser made the move from the Google Brain team to OpenAI, which happened to be one of Cohere's main competitors, in the same year that Cohere secured its inaugural funding round.

Meanwhile, after his undergraduate studies at the University of Toronto, the second co-founder Nick Frosst became the first employee at Google Brain’s Toronto lab, led by Geoffrey Hinton, sometimes called 'the godfather of AI’. Frosst's tenure at the lab spanned from 2016 to 2020, and it was during this time that he met Aidan Gomez.

The third co-founder, Ivan Zhang, provided insight into his journey and the genesis of the company in a podcast hosted by Madrona. Zhang decided to leave the university as he said “to get my hands on the technology to learn.” He dropped out of school to work at his friend’s startup and then he met Aiden Gomez who wanted to do an indie research group. This is when the match happened between Zhang's personality who thought that “it would be pretty badass to publish papers as a dropout” and Gomez’s idea. They both wanted to be independent.

This is how For.ai started in 2017, as described on their website: "a team of friends, classmates, and engineers started a distributed research collaboration, with a focus on creating a medium for early-career AI enthusiasts to engage with experienced researchers." At that time, For AI stood as one of the pioneering community-driven research groups supporting independent researchers across the globe.

In 2019, Zhang suggested to Aidan the idea of starting a new venture together as they both learned so much and got experience at For.ai. Even in 2019, there weren't many firms that used deep learning while Gomez experienced the proliferation of the new, powerful thing called transformers at Google. Zhang shared their experience at that time:

By the time GPT-2 went out, Zhang and Aiden noticed how powerful these transformers could be and also witnessed their architectural change when they became decoder-only meaning these models could write.

That’s how Cohere was born.

Mission of the company and its founders

In 2019 tech giants who actively developed new ideas and implemented them in their products remained the leaders of the market. But Cohere became an exception. What started as pure enthusiasm has become a product that competes with other NLP model providers, the tech giants.

At the same time, Cohere reaffirms that they also work on making this technology secure for the users mitigating the biases that could be presented in the data LLMs are trained on. The topic gained substantial traction after Google showed the exit door to AI researcher Timnit Gebru and a few more researchers when they started to point out such pitfalls of LLMs.

Financial situation

In May 2019, Cohere emerged from stealth mode, not empty-handed but armed with a handful of test customers and support from some of the brightest minds in the AI industry.

PitchBook reports that Cohere conducted its first early-stage VC round of $5 million in 2020.

Two years later, in 2021, the three co-founders successfully secured their first funding round, raising an impressive $40 million. This backing came from notable figures in the AI sphere, including Turing Award winner Geoffrey Hinton, Fei-Fei Li, Pieter Abbeel, and Raquel Urtasun, in addition to venture capital firms. The co-founder of Index Ventures, Mike Volpi, said

In February 2022, Cohere secured another round of funding, amassing $125 million from the same group of investors as in 2021, but this time, the round was led by Tiger Global Management. The press release stated that the most recent fundraise has brought Cohere's total funding to date to over $170 million, thus confirming the initial $5 million round.

As part of its Series C round, in June 2023, Cohere raised another $270 million at a $2.2 billion valuation. Notably, these investments came from prominent companies rather than individual investors, with significant contributions from industry giants such as NVIDIA, Oracle, and Salesforce.

Fast forward to the present, according to our research, Cohere lost two positions in the ranking of AI unicorns after the sudden rise of Hugging Face and Inflection. Now, Cohere is #5, but does it mean the company loses its position in the market? In February, Reuters speculated that Cohere aimed to raise "hundreds of millions" at a valuation of over $6 billion. If they bring more enterprise clients on board as planned, perhaps the $270 million raise is just the beginning.

What Cohere offers and their tech specs

The company has positioned itself as a neutral provider for enterprises to use models that are not tied to clouds like Microsoft’s Azure and Amazon Web Services (AWS), it can also work on a customer's existing cloud setup or even on their computers. As of now, Cohere’s NLP capability’s customers vary from recent unicorn company Ada to the music streaming platform Spotify.

Cohere models are divided into two primary categories: Generative and Embeddings.

• Generative Models: The initial offering was what Cohere termed as 'base' or 'command' models. The Command is trained on a vast corpus of internet data. What sets Cohere apart is its frequent update schedule, with models showing week-to-week improvements. The latest iteration of Cohere’s Command model ranks competitively in Stanford HELM rankings Another unique feature is a specialized model set called 'Command Nightly,' targeted for more frequent updates, aimed at eventually reaching a nightly cadence.

• Embeddings Models: These are multilingual models that support over 109 languages and are designed for large enterprises. The model enables search capabilities across multiple languages and is particularly useful for indexing a diverse data corpus.

Recently, Cohere has also launched Coral, an enterprise-grade knowledge assistant designed to boost productivity for strategic teams. Powered by Cohere's Command model, Coral engages in intelligent chat and task automation. It offers customization via 100+ data integrations, citation-backed responses, and private deployment options for data security. The service is currently in private access and has garnered a partnership from Oracle.

Furthermore, the company has an LLM University (LLMU)! Through this service, the company offers a comprehensive set of learning resources for anyone interested in NLP, from beginners to advanced learners.

How does Cohere make money?

Cohere's revenue generation is designed to cater to enterprise-level clientele. These enterprises can use Cohere’s platform by training their AI models using proprietary data, all while keeping that data confidential. Before integrating the platform into their products, customers have the advantage of testing its capabilities at no initial cost.

The cost structure for the paid plans is determined by the specific model in use and the volume of tokens purchased. Moreover, the company extends personalized pricing options through its support team. For the Default model, the pricing commences at $0.04 for every 1 million tokens, while the Custom model starts at $0.08 for the same token volume.

Among Cohere's clientele are prominent names like BambooHR, Florite, Glean, Oracle, and Deepjudge, although this list is by no means exhaustive.

Bonus: All important links about the founders

Aidan Gomez, Co-founder & CEO

• His Twitter / LinkedIn Profile / Personal website / Publications / GitHub

Nick Frosst, Co-founder

• His Twitter / LinkedIn Profile / Personal Website / Publications / GitHub

• Nick is also a Singer at Good Kid

Ivan Zhang, Co-founder, CTO

• His Twitter / LinkedIn Profile / Personal Website / Publications / GitHub

To be continued…