Supported by Nvidia + invested by Turing Award winner, this company on the corporate track has become the third largest unicorn in the field of AI

Source: Silicon Starman (ID: guixingren123)

Author: Juny

edit丨Vicky Xiao

Image credit: Generated by Unbounded AI tools

In recent months, it seems that everyone has been chatting enthusiastically with AI robots such as ChatGPT and Bard in their daily lives, asking them to help write emails, write summaries, and make plans. But a common phenomenon is that once people switch to work mode, these generative AI tools rarely appear in everyone's workflow, and are even banned by some companies.

AI is so powerful, why don't companies use it?

The core reason behind this is actually very simple - the data security and privacy issues of every enterprise are too important. No enterprise dares to put its "lifeblood" completely in the hands of AI that has not yet matured and is controlled by other large companies.

So, is there a way to solve this thorny problem to maximize the potential of generative AI on the enterprise side? In fact, since 2019, an AI startup company called Cohere has been paying close attention to this problem and constantly proposing solutions.

For a long time, enterprise-level generative AI has been a relatively niche market with high barriers, but Cohere has received the support of many bigwigs and giants with its mature technology and keen sense of smell. At present, Cohere’s investors include not only giants such as Nvidia, Oracle, and Salesforce, but also Turing Award winner Geoffrey Hinton, Stanford AI professor Li Feifei, and other big names in the circle. Not long ago, YouTube's former chief financial officer, Martin Kon, also opted to join Cohere as president and chief operating officer.

Nvidia, Oracle and Salesforce are all betting on Cohere

Source: Crunchbase

Riding on the boom of ChatGPT, since this year, more and more people have noticed the potential of Cohere and entered the fast lane of soaring valuation. It has become the No. 1 global AIGC track after OpenAI and Antropic. Three unicorns.

"Born out" from Google, from Canada's top AI circle

Founded in Toronto, Canada, Cohere was co-founded in 2019 by Aidan Gomez, Ivan Zhang and Nick Frosst. All three of them are majoring in computer science at the University of Toronto. Based on the time of admission, all three should be under the age of 30.

Cohere's founding team

Source: Cohere official website

Among them, Aidan Gomez participated in 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 signatories. This paper is the foundation of the famous Transformer machine learning architecture in the future. The beginning is also the cornerstone of future revolutionary architectures such as Google BERT and OpenAI's GPT.

In the same year, Aidan Gomez and fellow student Ivan Zhang founded For.ai, a non-profit artificial intelligence research community, to support and link independent artificial intelligence researchers around the world.

After graduating from his undergraduate degree, Aidan Gomez went to Oxford University to study for a doctorate in computer science. At the same time, he also joined the Google AI team led by the "father of deep learning" and Turing Award winner Geoffrey Hinton to conduct further research based on the Transformer architecture. In the Hinton team of Google Brain, Aidan Gomez met Nick Frosst, who has been engaged in machine learning and cognitive science research.

In the next two years, through in-depth research, everyone learned that Transformer can be expanded into a large neural network with excellent performance, and it performs very well on language-related tasks. Some Transformer paper writers, including Aidan Gomez, have begun to think about the commercialization opportunities behind this. Currently, except for Llion Jones who is still working at Google, the other seven authors have "go to sea" to start 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 to train large models, since Cohere was established in 2019, they have focused on enterprise use cases, trying to create customized large language models based on the proprietary data of different companies.

Do not rely on the cloud, but do enterprise-customized generative AI services

Simply put, Cohere's goal is to become the default NLP toolkit for all types of developers, allowing developers of all types to use large neural networks and state-of-the-art AI to solve any language-related problems, but not relying on any public cloud , allowing the model to run on a private cloud or on-premises.

At present, Cohere's main products mainly revolve around three key areas in the daily operation of enterprises: text generation, text classification and text retrieval, covering almost all text-related areas in enterprise production.

The text generation part mainly has three products: Summarize, Generate, and Command Model. Summarize is a text summary generator driven by a large language model, which can quickly summarize and summarize the key points of a document, and can support input of 100,000 characters and text format options. Generate is a content generator that can generate unique content for various purposes, such as emails and product descriptions.

Next, let's focus on the Command Model. Command is a text generation model launched by Cohere that can accept user personalized commands for training. That is to say, after combining their own data with Command, enterprise users can generate their own unique language model, which can immediately play a role in the actual business of the enterprise.

Command Model

Source: Cohere

It is worth noting that, as a model with only 52 billion parameters, the accuracy of Command has previously surpassed other models trained on a larger scale. The most capable large-scale language model.

Source: Stanford University’s Comprehensive Evaluation of Language Models (HELM) official website

The text retrieval part includes three products: Embed, Semantic Search, and Rerank.

For machine learning teams looking to build their own text analytics applications, Embed helps them spot trends quickly and supports more than 100 languages. Semantic Search is a powerful search tool. Users only need to simply use the API to use the search function. It supports the return of various information based on the meaning of the query rather than just keywords, and is not limited by language. Rerank can analyze and rank search results from existing tools based on semantic relevance, providing richer and more relevant results with minimal requirements for user intervention or programming experience.

The main product of the text classification part is Classify, which enables users to personalize and organize information to help content moderation, user analysis and chatbot experience. For example, it can conduct efficient customer service by quickly marking different categories of customers, and it can also identify positive and negative social media comments to better understand customer feedback.

Source: Cohere

Cohere's business model is to first bear the cost of creating a large Transformer neural network, and then connect companies that need it to these networks, and the company pays according to usage. The main feature of Cohere is that it provides customers with a variety of data hosting options including private cloud, local deployment, Cohere managed cloud and other cloud partners AWS, Google, etc., allowing users to choose according to their own needs, allowing customers to have control over the data.

For developers who wish to learn prototyping and become part of the community, Cohere offers free, limited-use access. However, there will be a fee to go into production, train custom models, access all endpoints, and receive enhanced customer support. Current clients of Cohere include Spotify, Jasper, HyperWrite, etc.

In terms of price, under the embedding function, the default model is 40 cents per 1 million tokens, and the enterprise-defined model is 80 cents. Under the generation function, the default model is 15 dollars per 1 million tokens, and the custom model is 30 dollars. Summary Under the function, $15 per 1 million Tokens, etc.

Cohere prices for different functions

Source: Cohere

However, Cohere's previous pricing was quite advantageous, but after the big price cut of OpenAI yesterday, it is expected to have a big impact on Cohere. For example, the price of OpenAI's embedded model has dropped by 75%, and it only costs US$0.0001 per thousand tokens, which is 10 million tokens for US$1, which is far lower than Cohere.

Supported by big guys and giants, Cohere entered the first camp of AIGC

Cohere, which aims at the pain points of enterprise-level AI data security, stands out in the current AI client competition, including VCs, technology giants, and big names in the field of artificial intelligence. They all voted for it. Since officially entering commercialization in 2021, Cohere's valuation has also been rising steadily, and has now reached about 2.2 billion US dollars, second only to Microsoft-backed OpenAI and Google-backed Anthropic on the AIGC track.

At the beginning of Cohere's establishment, its artificial intelligence academic color seemed to be stronger. In the Cohere A and B rounds of financing in 2021 and 2022, the investment in the AIGC track at that time was still in the cold winter. Who invested in Cohere to support funds? In the investment lists of these two rounds, we have seen the following figures of AI giants.

Source: Crunchbase

In addition to Geoffrey Hinton, the "Father of Deep Learning" and Turing Award winner, several founders directly followed and studied in Toronto, they also include Li Feifei, a professor at Stanford University and head of the Vision Lab, and a professor at the University of California, Berkeley, Berkeley Artificial Intelligence. Pieter Abbeel, the director of the laboratory, and Raquel Urtasun, a professor at the University of Toronto and former director of Uber's driverless car technology research center, are all academic experts in the field of artificial intelligence.

In the latest round of financing announced earlier this month, amid the upsurge of AIGC, Cohere has also attracted the attention of more technology companies in the field. These include Nvidia, the strongest "arms dealer" in artificial intelligence, and cloud giants Salesforce and Oracle. The current total financing has reached 439 million US dollars.

Cohere's rapid development is inseparable from its profound technical background and track selection. From the perspective of large-scale models, Cohere may not be the most leading in the market at present, but they have keenly grasped the pain points of AIGC enterprise applications, and can further provide content generation on the premise of first meeting the security needs of enterprises. , summarization, search and other services.

Their business model enables a large number of companies to customize access to large neural networks without spending a lot of money to build their own models, and by subdividing business modules, companies can pay according to usage, so as to achieve a win-win situation state.

Judging from the increasing popularity of Cohere and OpenAI's recent large-scale price cuts and API upgrades, the AIGC war is spreading from the user side to the enterprise battlefield. At that time, perhaps a real AI productivity revolution will really begin.