In recent months, it seems like everyone has been chatting animatedly with AI bots like ChatGPT and Bard in their daily lives, asking them to help write emails, summarize, and make plans. But one common phenomenon is that once people switch to work mode, these generative AI tools rarely appear in everyone’s workflow, and are even explicitly banned by some companies.
AI is so powerful, why don’t companies use it?
The core reason behind this is simple – data security and privacy are too important for every organization. No enterprise dares to put its “lifeblood” completely in the hands of AI, which has not yet matured and is controlled by other large companies. So is there a way to solve this thorny problem and maximize the potential of generative AI on the enterprise side? In fact, an AI startup called Cohere has been closely monitoring this problem and continuously coming up with solutions since the beginning of 2019.
Enterprise-grade generative AI has long been a relatively niche and high-barrier market, but Cohere has been backed by a host of bigwigs and giants with its proven technology and keen sense of smell. At present, Cohere’s investors not only include NVIDIA, Oracle, Salesforce and other giants, but also Turing Award winner Geoffrey Hinton, Stanford AI professor Feifei Li and a number of other big names in the circle. Not long ago, Martin Kon, former CFO of YouTube, also chose to join Cohere as President and COO. Riding on the ChatGPT fire east wind, this year, Cohere’s potential began to be seen by more and more people and entered the fast lane of soaring valuation, has become the third largest unicorn in the global AIGC track, second only to OpenAI and Antropic.
1. “Born out of Google”, from Canada’s top AI circle
Cohere was founded in Toronto, Canada in 2019 by Aidan Gomez, Ivan Zhang and Nick Frosst. All three undergraduates attended the University of Toronto majoring in computer science, and according to the projection of the time of enrollment, all three should not be over 30 years old at present. Among them, Aidan Gomez participated in the research by the Google Brain team during his undergraduate studies in 2017 and published a paper titled “Attention is All You Need” as one of the co-authors, which was the beginning of the later famous Transformer machine learning architecture, and the development of future revolutionary architectures such as Google’s BERT, OpenAI’s GPT, and other revolutionary architectures of the future. 
In the same year, Aidan Gomez and fellow student Ivan Zhang founded For.ai, a non-profit AI research community to support and connect independent AI researchers around the world. After completing his undergraduate degree, Aidan Gomez went on to pursue a PhD in Computer Science at the University of Oxford, where he also joined the Google AI team led by Geoffrey Hinton, the “Father of Deep Learning” and winner of the Turing Award, to further his research based on the Transformer architecture. In the Hinton team at Google Brain, Aidan Gomez met Nick Frosst, who has been engaged in machine learning and cognitive science research.
Over the next two years, through further research, it became clear that the Transformer could be scaled up into a large neural network with excellent performance, and that it could perform very well on language-related tasks. Some of the authors of the Transformer paper, including Aidan Gomez, began to think about the commercialization opportunities behind it, and now, with the exception of Llion Jones, who is still working at Google, the other seven authors have started their own businesses. Among them, Aidan Comez co-founded Cohere with Nick Frosst and Ivan Zhang. unlike Google, Microsoft and other powerful companies that spend a lot of money on training large models, Cohere was founded in 2019, they focus on enterprise use cases, trying to create customized large language models based on the proprietary data of different enterprises.
2.Not relying on cloud services, to make customized generative AI services for enterprises
Simply put, Cohere’s goal is to become the default NLP toolkit for all kinds of developers, allowing all kinds of developers to use large neural networks and state-of-the-art AI to solve any language-related problem, but without relying on any public cloud, allowing the models to run in private cloud services or local deployments. Currently, Cohere’s main products are centered around three key areas of daily business operations: text generation, text categorization, and text retrieval, covering almost all text-related areas of business production.
Cohere’s text generation products include Summarize, Generate, and Command Model, which is a large-scale language model-driven text summary generator that quickly outlines and summarizes the key points of a document, and supports input of 100,000 characters and text formatting options. Generate is a content generator that generates unique content for a variety of purposes, such as emails and product descriptions.
Next, let’s focus on Command Model, a text generation model from Cohere that is trained to accept personalized commands from users. In other words, after combining their own data with Command, business users can generate their own unique language model, which can be immediately useful in the actual business of the enterprise. Notably, as a model with only 52 billion parameters, Command’s accuracy performance has previously outperformed that of other, more massively trained models, and was recently named the world’s most capable large-scale language model by Stanford University’s Holistic Evaluation of Language Models (HELM).
The Text Retrieval section, includes three products, Embed, Semantic Search, and Rerank. For machine learning teams looking to build their own text analytics applications, Embed helps them quickly spot trends and supports more than 100 languages.Semantic Search is a powerful search tool that is available to users with a simple API that can support returning a wide range of information based on the meaning of a query, not just keywords, and is language-independent.Rerank is a powerful search tool that is available to users in a variety of languages, including English, Chinese, French, Spanish, and Spanish. Rerank can analyze and rank search results from existing tools based on semantic relevance to provide richer, more relevant results with minimal user intervention or programming experience.
The main product in the text categorization section is Classify, a feature that enables users to personalize and organize information to help with content audits, user analytics, and for chatbot experiences. For example, it enables efficient customer service by quickly tagging different categories of customers, as well as identifying positive and negative social media comments to better understand customer feedback.
Cohere’s business model is to bear the cost of creating large Transformer neural networks upfront and then connecting in-demand companies to these networks, with the companies paying on a per-use basis.Cohere’s key feature is that it offers its customers a wide range of data-hosting options, including private clouds, local deployments, the Cohere Hosted Cloud and other cloud partners such as AWS, Google, and others, allowing the users to choose according to their needs, giving customers control over their data.
For developers looking to learn about prototyping and become part of the community, Cohere offers free, usage-limited access. However, those wishing to go into production, train custom models, access all endpoints and receive enhanced customer support will need to pay a fee. Current Cohere customers include Spotify, Jasper, HyperWrite and more. In terms of pricing, it’s 40 cents per 1 million Token for the default model and 80 cents per 1 million Token for enterprise-customized models for the embedding feature, $15 per 1 million Token for the default model and $30 for the custom model for the generating feature, and $15 per 1 million Token for the summarizing feature, among others. However, this previous pricing of Cohere was quite favorable, but after yesterday’s big price cut by OpenAI, it is expected to give Cohere a considerable impact as well. For example, the price of OpenAI’s embedding model directly jumped 75%, only $0.0001 per thousand tokens, that is, $1 ten million tokens, far lower than Cohere.
3. Big brother and giant support, Cohere sails into the first camp of AIGC
Aiming at enterprise-level AI data security pain points, Cohere stands out in the current AI user-side killing, including VCs, tech giants and bigwigs in the field of artificial intelligence have given it a vote of support. 2021 formally entered the commercialization of Cohere’s valuation is also climbing, has reached about 2.2 billion U.S. dollars, in the AIGC track is only second to the Microsoft-backed OpenAI and Google-backed OpenAI and Google-backed Anthropic.At the time of Cohere’s inception, it appeared to have a more academic AI bent. In Cohere’s Series A and Series B rounds in 2021 and 2022, when investment in the AIGC track was still in the dead of winter, who threw backing money at Cohere? We saw the following AI bigwigs on the investment list for both rounds.
In addition to the “father of deep learning” and Turing Award winner Geoffrey Hinton, whom several founders studied directly under in Toronto, there are also Stanford University professor and head of the vision lab Feifei Li, UC Berkeley professor and director of the Berkeley Artificial Intelligence Lab Pieter Abbeel, University of Toronto professor and former director of Uber’s Human-Driven Vehicle Technology Research Center, Raquel Urtasun, all of whom are academic gurus in the field of artificial intelligence. And in the latest round of funding announced earlier this month, amidst the AIGC buzz, Cohere is also getting more attention from tech companies in the field. These include NVIDIA, the most powerful “arms dealer” in AI, as well as cloud giants Salesforce and Oracle. Currently, the total amount of funding has reached $439 million.
Cohere’s rapid development can not be separated from the deep technical background and track selection. From a big model perspective, Cohere may not be the market leader at the moment, but they have keenly grasped the pain points of AIGC’s enterprise applications, and are able to provide services in areas such as content generation, summarization, and search, under the premise of first meeting the security needs of the enterprise, and then going further. Their business model enables a large number of companies to customize their access to large neural networks without having to spend a lot of money on building their own models, and through segmentation of business modules, allows companies to pay based on usage, thus achieving a win-win situation. Judging from the growing popularity of Cohere and OpenAI’s recent massive price cuts and API upgrades, the fire of AIGC is spreading all the way to the enterprise battlefield on the user side. And by then, perhaps a real AI productivity revolution will really begin.